Machine Translation Divergences: 
A Formal Description and Proposed 
Solution 
Bonnie J. Dorr* 
University of Maryland 
There are many cases in which the natural translation of one language into another results in a 
very different form than that of the original. The existence of translation divergences (i.e., cross- 
linguistic distinctions) makes the straightforward transfer from source structures into target 
structures impractical. Many existing translation systems have mechanisms for handling diver- 
gent structures but do not provide a general procedure that takes advantage of the systematic 
relation between lexical-semantic structure and syntactic structure. This paper demonstrates 
that a systematic solution to the divergence problem can be derived from the formalization of 
two types of information: (1) the linguistically grounded classes upon which lexical-semantic 
divergences are based; and (2) the techniques by which lexical-semantic divergences are resolved. 
This formalization is advantageous in that it facilitates the design and implementation of the 
system, allows one to make an evaluation of the status of the system, and provides a basis for 
proving certain important properties about the system. 
1. Introduction 
There are many cases in which the natural translation of one language into another 
results in a very different form than that of the original. The existence of translation 
divergences (i.e., cross-linguistic distinctions) makes the straightforward transfer from 
source structures into target structures impractical. This paper demonstrates that a sys- 
tematic solution to the divergence problem can be derived from the formalization of 
two types of information: (1) the linguistically grounded classes upon which lexical- 
semantic divergences are based; and (2) the techniques by which lexical-semantic 
divergences are resolved. An important result of this formalization is the provision of 
a framework for proving that the lexical-semantic divergence classification proposed 
in the current approach covers all source-language/target-language distinctions based 
on lexical-semantic properties. Other types of divergences and mismatches are outside 
of the scope of this paper; these include distinctions based on purely syntactic informa- 
tion, idiomatic usage, aspectual knowledge, discourse knowledge, domain knowledge, 
or world knowledge/ 
Although other translation approaches have attempted to account for divergences, 
the main innovation of the current approach is that it provides a formalization of these 
divergences and the techniques by which they are resolved. This is advantageous from 
a computational point of view in that it facilitates the design and implementation of 
• Department of Computer Science, University of Maryland, A. V. Williams Building, College Park, MD 
20742, USA. 1 The reader is referred to Dorr (1993a) for a discussion of how syntactic divergences are handled. 
Aspectual divergences are treated by Dorr (1992a). The relatio.n of the current framework to other types 
of knowledge outside of lexical semantics is discussed by Dorr and Voss (1993b). 
© 1994 Association for Computational Linguistics 
Computational Linguistics Volume 20, Number 4 
(1) Thematic divergence: 
E: I like Mary ~ S: Maria me gusta a mi 
'Mary pleases me' 
(2) Promotional divergence: 
E: John usually goes home 4=~ S: Juan suele ira casa 
'John tends to go home' 
(3) Demotional divergence: 
E: I like eating ~ G: Ich esse gem 
'I eat likingly' 
(4) Structural divergence: 
E: John entered the house 4=~ S: Juan entr6 en la casa 
'John entered in the house' 
(5) Conflational divergence: 
E: I stabbed John ~ S: Yo le di pu~aladas a Juan 
'I gave knife-wounds to John' 
(6) Categorial divergence: 
E: I am hungry ~ G: Ich habe Hunger 
'I have hunger' 
(7) Lexical divergence: 
E: John broke into the room ~ S: Juan forz6 la entrada al cuarto 
Figure 1 'John forced (the) entry to the room' 
Examples of translation divergences with respect to English, Spanish, and German. 
the system: the problem is clearly defined in terms of a small number of divergence 
categories, and the solution is systematically stated in terms of a uniform translation 
mapping and a handful of simple lexical-semantic parameters. In addition, the for- 
malization allows one to make an evaluation of the status of the system. For example, 
given the formal description of the interlingua and target-language root words, one is 
able to judge whether a particular target-language sentence fully covers the concept 
that underlies the corresponding source-language sentence. Finally, the formalization 
of the divergence types and the associated solution allows one to prove certain proper- 
ties about the system. For example, one might want to determine whether the system 
is able to handle two or more simultaneous divergences that interact in some way. 
With the mechanism of the current approach, one is able to prove formally that such 
cases are handled in a uniform fashion. 
This paper will focus on the problem of lexical-semantic divergences and will 
provide support for the view that it is possible to construct a finite cross-linguistic 
classification of divergences and to implement a systematic mapping between the 
interlingual representation and the surface syntactic structure that accommodates all 
of the divergences in this classification. The types of divergences under consideration 
are those shown in Figure 1. The first divergence type is thematic: in (1), the theme 
is realized as the verbal object (Mary) in English but as the subject (Maria) of the 
main verb in Spanish. The second divergence type, promotional, is one of two head 
switching divergence types: in (2), the modifier (usually) is realized as an adverbial 
phrase in English but as the main verb soler in Spanish. The third divergence type, 
demotional, is another type of head switching divergence: in (3), the word like is realized 
as a main verb in English but as an adverbial modifier (gern) in German. 2 The fourth 
2 The distinction between promotional and demotional divergences is not intuitively obvious at first 
glance. In both (2) and (3), the translation mapping associates a main verb with an adverbial satellite, 
or vice versa (i.e., in (2), the main verb soler is associated with the adverbial satellite usually, and in (3) 
the main verb like is associated with the adverbial satellite gern). The distinction between these two 
598 
Bonnie J. Dorr Machine Translation Divergences 
divergence type is structural: in (4), the verbal object is realized as a noun phrase 
(the house) in English and as a prepositional phrase (en la casa) in Spanish. The fifth 
divergence type is conflational. Conflation is the incorporation of necessary participants 
(or arguments) of a given action. In (5), English uses the single word stab for the two 
Spanish words dar (give) and pu~aladas (knife-wounds); this is because the effect of the 
action (i.e., the knife-wounds portion of the lexical token) is conflated into the main verb 
in English. The sixth divergence type is categoriah in (6), the predicate is adjectival 
(hungry) in English but nominal (Hunger) in German. Finally, the seventh divergence 
type is a lexical divergence: in (7), the event is lexically realized as the main verb break 
in English but as a different verb forzar (literally force) in Spanish. 
The next section discusses the divergence classification given above, comparing 
the current divergence categories with those of other researchers. Section 3 formally 
defines the terms used to classify divergences. Section 4 uses this terminology to 
formalize the divergence classification and to define the solution to the divergence 
problem in the context of detailed examples. Finally, Section 5 discusses certain issues 
of relevance to the divergence problem including the resolution of several (recursively) 
interacting divergence types. 
2. Classification of Machine Translation Divergences 
The divergence problem in machine translation has received increasingly greater at- 
tention in recent literature (see, for example, Barnett et al. 1991a, 1991b; Beaven 1992a, 
1992b; Dorr 1990a, 1990b; Kameyama et al. 1991; Kinoshita, Phillips, and Tsujii 1992; 
Lindop and Tsujii 1991; Tsujii and Fujita 1991; Whitelock 1992; related discussion can 
also be found in work by Melby \[1986\] and Nirenburg and Nirenburg \[1988\]). In par- 
ticular, Barnett et al. (1991a) divide distinctions between the source language and the 
target language into two categories: translation divergences, in which the same infor- 
mation is conveyed in the source and target texts, but the structures of the sentences 
are different (as in previous work by Dorr \[1990a, 1990b\]); and translation mismatches, 
in which the information that is conveyed is different in the source and target lan- 
guages (as described by Kameyama et al. \[1991\]). 3 Although translation mismatches 
are a major problem for translation systems that must be addressed, they are outside 
the scope of the model presented here. (See Barnett et al. 1991a, 1991b; Carbonell and 
Tomita 1987; Meyer, Onyshkevych, and Carlson 1990; Nirenburg, Raskin, and Tucker 
1987; Nirenburg and Goodman 1990; Nirenburg and Levin 1989; Wilks 1973; among 
others, for descriptions of interlingual machine translation approaches that take into 
account knowledge outside of the domain of lexical semantics.) 
Although researchers have only recently begun to classify divergence types sys- 
tematically, the notion of translation divergences is not a new one in the machine 
translation community. For example, a number of researchers working on the Euro- 
tra project have sought to solve divergent source-to-target translations, although the 
divergences were named differently and were resolved by construction-specific trans- 
fer rules. (For cogent descriptions of the Eurotra project, see, for example, Arnold 
and des Tombe 1987; Copeland et al. 1991; and Johnson, King, and des Tombe 1985). 
head switching cases will be made clearer in Section 4.3. 
3 An example of the latter situation is the translation of the English word fish into Spanish: the 
translation is pez if the fish is still in its natural state, but it is pescado if the fish has been caught and is 
suitable for food. It is now widely accepted that, in such a situation, the machine translation system 
must be able to derive the required information from discourse context and a model of the domain that 
is being discussed. 
599 
Computational Linguistics Volume 20, Number 4 
Figure 2 
Formal definition of lexical conceptual structure. 
A comprehensive survey of divergence examples is presented by Lindop and Tsujii 
(1991). The term used in this work is "complex transfer," but it describes a class of 
problems inherent in machine translation itself, not just in the transfer (or interlingual) 
approaches. 
One of the claims made by Lindop and Tsujii (1991) is that the non-Eurotra liter- 
ature rarely goes into great detail when discussing how divergences are handled. An 
additional claim is that combinations of divergences and interaction effects between 
divergent and nondivergent translations are not described in the literature. This paper 
seeks to change this perceived state of affairs by providing a detailed description of a 
solution to all of the (potentially interacting) divergences shown in Figure 1, not just 
a subset of them as would typically be found in the description of most translation 
systems. The framework assumed for the current approach makes use of a linguis- 
tically grounded classification of divergence types that can be formally defined and 
systematically resolved. 
We now turn to a formal description of the terminology used to define the diver- 
gence problem. 
3. Definitions 
This section formally defines the lexical-semantic representation that serves as the 
interlingua of the system (Definitions 1-3). This representation, which is influenced 
primarily by Jackendoff (1983, 1990), has been described in detail elsewhere (see, for 
example, Dorr 1992b, 1993a) and thus will not be the focus of this paper. In addition 
to a formal description of the lexical-semantic representation, definitions are provided 
for syntactic phrases (Definition 4) and two translation mappings (Definitions 5 and 6). 
Definition 1 
A lexical conceptual structure (LCS) is a modified version of the representation proposed 
by Jackendoff (1983, 1990) that conforms to the following structural form: 
\[T(X') X' (\[T(W') Wt\], \[T(Zq) Ztl\] "'" \[T(Z',,) Ztn\] \[T(Q',) Q'I\] -" \[T(Q',,,) Q'm\])\] 
This corresponds to the tree-like representation shown in Figure 2, in which (1) X' is 
the logical head; (2) W' is the logical subject; (3) Z~... Z~ are the logical arguments; and 
(4) Q~ ... Q~m are the logical modifiers. These four positions are relevant to the mapping 
600 
Bonnie J. Dorr Machine Translation Divergences 
( oo . 1 
I JOHN TOLo c HAPPILY 
• Thing path Manner 
! I 1 
Thing Location 
Figure 3 
CLCS representation for John went happily to school. 
between the interlingual representation and the surface syntactic representation. In 
addition, T(~) is the logical type (Event, State, Path, Position, etc.) corresponding to 
the primitive ~ (CAUSE, LET, GO, STAY, BE, etc.); Primitives are further categorized 
into fields (e.g., Possessional, Identificational, Temporal, Locational, etc.). 4 
Example 1 
The LCS representation of John went happily to school is 
\[Event GOLoc 
(\[Thing JOHN\], 
\[Path TOLoc (\[Position ATLoc (\[Thing JOHN\], \[Location SCHOOL\])\])\] 
\[M ..... HAPPILY\])\] 
This corresponds to the tree-like representation shown in Figure 3, in which (1) the 
logical head is GOLoc (of type Event); (2) the logical subject is JOHN (of type Thing); 
(3) the logical argument is TOcoc (of type Path); and (4) the logical modifier is HAPPILY 
(of type Manner). Note that the logical argument is itself a LCS that contains a logical 
argument, SCHOOL (of type Location), i.e., LCSs are recursively defined. 
The LCS representation is used both in the lexicon and in the interlingual repre- 
sentation. The former is identified as a root LCS (RLCS) and the latter is identified as 
a composed LCS (CLCS): 
Definition 2 
A RLCS (i.e., a root LCS) is an uninstantiated LCS that is associated with a word 
definition in the lexicon (i.e., a LCS with unfilled variable positions). 
Example 2 
The RLCS associated with the word go (from Example 1) is 
\[Event GOLoc (\[Thing X\], \[Path TOLoc (\[Position ATLoc (\[Thing X\], \[Location Z\])\])\])\] 
4 The validity of the primitives and their compositional properties is not discussed here. The LCS has 
been studied as the basis of a representation for multiple languages (see, for example, Hale and Keyser 
1986a, 1986b, 1989; Hale and Laughren 1983; Levin and Rappaport 1986; Zubizarreta 1982, 1987) and is 
discussed in the context of machine translation by Dorr (1992b). 
601 
Computational Linguistics Volume 20, Number 4 
( 
{ ) 
x / \ ... 
, l 
Position ) 
I(' \[ x 7. Thing Location 
Figure 4 
RLCS representation for go. 
which corresponds to the tree-like representation shown in Figure 4. 
Definition 3 
A CLCS (i.e., a composed LCS) is an instantiated LCS that is the result of combining two 
or more RLCSs by means of unification (roughly). This is the interlingua, or language- 
independent, form that serves as the pivot between the source and target languages. 
Example 3 
If we compose the RLCS for go (in Figure 4) with the RLCSs for John (\[ThingJOHN\]), 
school (\[Location SCHOOL\]), and happily (\[Manner HAPPILY\]), we get the CLCS corre- 
sponding to John went happily to school (shown in Figure 3). 
Each (content) word in the lexicon is associated with a RLCS, whose variable 
positions may have certain restrictions. The CLCS is a structure that results from com- 
bining the lexical items of a source-language sentence into a single underlying pivot 
form by means of LCS composition. 5 The notion of unification (as used in Definition 3) 
differs from that of the standard unification frameworks (see, for example, Shieber 
et al. 1989, 1990; Kaplan and Bresnan 1982; Kaplan et al. 1989; Kay 1984; etc.) in that 
it is not directly invertible. That is, the generation process operates on the CLCS in a 
unification-like fashion that roughly mirrors the LCS composition process, but it is not 
a direct inverse of this process. The notion of unification used here also differs from 
others in that it is a more "relaxed" notion: those words that are mapped in a relaxed 
way are associated with special lexical information (i.e., the :INT, :EXT, :PROMOTE, 
:DEMOTE, ,, :CAT, and :CONFLATED parameters, each of which will be formalized 
shortly). 
A fundamental component of the mapping between the interlingual representation 
and the surface syntactic representation is the syntactic phrase. 
Definition 4 
A syntactic phrase is a maximal projection that conforms to the following structural 
form: 
5 This process is described in detail in Dorr (1992b). 
602 
Bonnie J. Dorr Machine Translation Divergences 
Y-MAX 
Q-MAXj+I ... Q-MAXk Y-MAX Q-MAXk+ I ... Q.MAX m 
W-MAX X-MAX 
X Z-MAX1 ... Z-MAX a 
Q-MAX 1 ... Q-MAXi X Q-MAXi+ 1 ... Q-MAXj 
Figure 5 
Formal definition of syntactic phrase. 
\[Y-MAX 
Q-MAXj+~ ... Q-MAXk 
\[Y-MAX 
W-MAX 
\[X-MAX \[X Q-MAX1 ... Q-MAXi X Q-MAXi+I ... Q-MAX i\] 
Z-MAX1 ... Z-MAXn\]\] 
Q-MAXk+1 ... Q-MAXm\] 6 
This corresponds to the tree-like representation shown in Figure 5, in which (1) X is 
the syntactic head (of category V, N, A, P, I, or C); (2) W-MAX is the external argument; 
(3) Z-MAX1, ..., Z-MAXn are the internal arguments; and (4) Q-MAX1 ..... Q-MAXm 
are the syntactic adjuncts. 
Example 4 
The syntactic phrase corresponding to John went happily to school is 
\[C-MAX \[I-MAX \[N-MAX John\] 
\[V-MAX \[v went\] \[ADV happily\] \[P-MAX to \[N-MAX school\]\]\]\]\] 
This corresponds to the tree-like representation shown in Figure 6, in which (1) the 
syntactic head is \[v went\]; (2) the external argument is \[N-MAX John\]; (3) the internal 
argument is \[P-MAX a ...\]; and (4) the syntactic adjunct is \[ADV happily\]. Note that 
the internal argument constituent is itself a syntactic phrase that contains an internal 
argument, \[N-MAX school\], i.e., syntactic phrases are recursively defined. 
In addition to the representations involved in the translation mapping, it is also 
possible to formalize the mapping itself. The current approach is to map between 
6 These syntactic structures are based on the X framework of government-binding theory (see Chomsky 
1981, 1982, 1986a, 1986b). For ease of illustration, the word order used in all formal definitions is 
head-initial/spec-initial (i.e., the setting for English). The syntactic operations that determine word 
order are completely independent from the lexical-semantic operations that use these definitions. Thus, 
the formal definitions can be stated in terms of an arbitrary ordering of constituents, without loss of 
generality, as long as it is understood that the constituent order is independently determined. 
603 
Computational Linguistics Volume 20, Number 4 
/ 
N-MAX 
A 
John 
C-MAX 
I 
I-MAX 
\ 
V-MAX 
V ADV P-MAX 
I I /\ 
went happily to N-MAX 
school 
Figure 6 
Syntactic phrase representation for John went happily to school. 
the LCS representation and the surface syntactic form by means of two routines that 
are grounded in linguistic theory: a generalized linking routine (G£T4) and a canonical 
syntactic realization (CST4). These routines are defined formally here: 
Definition 5 
The ~£T4 systematically relates syntactic positions from Definition 1 and lexical- 
semantic positions from Definition 4 as follows: 
1. X I 4=~X 
2. W'~W 
3. Z I\] • • • Kin ~ Zl • .. Zn 
4. Q'I--. Q'm 4=~ Q1,.. Qm 
Example 5 
The correspondence between the LCS of Example 1 and the syntactic structure of 
Example 4 (i.e., for the sentence John went happily to school) is (1) X ~ = GOLoc 4=~ X = Iv 
went\]; (2) W' = JOHN 4=~ W = \[N-MAX John\]; (3) Z' = TOcoc 4~ Z = \[pp to., .\]; and (4) 
Q' = HAPPILY ~ Q ~--~\[ADV happily\]. 
Definition 6 
The CST4 systematically relates a lexical-semantic type T(~') to a syntactic category 
CAT(h), where ~t is a CLCS constituent related to the syntactic constituent ¢> by the 
~£~. 
Example 6 
The LCS type Thing corresponds to the syntactic category N, which is ultimately 
projected up to a maximal level (i.e., N-MAX). The full range of realization possibilities 
is given in Figure 7. 
604 
Bonnie J. Dorr Machine Translation Divergences 
LCS Type 
EVENT 
STATE 
THING 
PROPERTY 
PATH 
POSITION 
LOCATION 
TIME 
MANNER 
INTENSIFIER 
PURPOSE 
Syntactic Category 
V 
V 
N 
A 
P 
P 
ADV 
ADV 
ADV 
ADV 
~ DILl, 
Figure 7 
CST4 mapping between LCS types and syntactic categories. 
Now that we have formally defined the representations and mappings used dur- 
ing translation, we will turn to a classification of divergences that is based on these 
definitions. 
4. The Divergence Problem: Formal Classification and Solution 
In general, translation divergences occur when there is an exception either to the 
~£T4 or to the CST4 (or to both) in one language, but not in the other. 7 This premise 
allows one to define formally a classification of all lexical-semantic divergences that 
arise during translation (i.e., divergences based on properties associated with lexical 
7 Most of the examples in this paper seem to suggest that a divergence is defined in terms of a 
language-to-language phenomenon: a divergence occurs when a sentence in language L1 translates into a 
sentence in L2 in a very different form (i.e., differently shaped parse trees or similarly shaped trees with 
different basic categories). This definition implies that a divergence may arise between two languages 
L1 and L2, independent of the way the translation is done (i.e., direct, transfer, or interlingual). 
However, it is also possible to define a divergence from an interlingual point of view, i.e., with respect 
to an underlying representation (lexical conceptual structure) that has been chosen to describe the 
source and target language sentences. From this point of view, a divergent mapping may apply even in 
cases in which the source- and target-language pairs do not exhibit any distinctions on the surface (e.g., 
the translation of the German sentence Hans kuflt Marie gern as the equivalent Dutch sentence Hans kust 
Marie graag, both of which literally translate to Hans kisses Mary likingly). In such cases, there are 
generally two occurrences of a language-to-interlingua divergence: one from the surface structure and one 
to the surface structure. (The terms language-to-language and language-to-interlingua are taken from Dorr 
and Voss 1993a.) At first glance, it might seem odd to introduce the notion of a language-to-interlingua 
divergence for cases that do not exhibit a language-to-language divergence. However, it is clearly the 
case that language-to-language divergences--a special case of language-to-interlingua divergences--do 
exist regardless of the translation approach adopted. Thus, we can view divergences more generally as 
a consequence of the internal mapping between the surface structure and the interlingual 
representation rather than as an external distinction that shows up on the surface. The result is that the 
interlingua appears to have been simplified to the extent that it accommodates constructions in one 
language (without any special information) more readily than it accommodates the corresponding 
construction in another language. However, as one reviewer points out, this is not an undesirable 
consequence, since the development of a suitable representation is where the interlingua builder has a 
choice and should choose the simplest representation format. The appropriate question to ask is 
whether an approach that addresses the divergence problem from a language-to-interlingua perspective 
is an improvement over an approach that addresses the problem strictly from a language-to-language 
point of view. This paper argues that the language-to-interlingua approach is the correct one given that 
the alternative would be to handle language-to-language divergences by constructing detailed 
source-to-target transfer rules for each lexical entry in the source and target language. Introducing the 
notion of language-to-interlingua divergence allows the translation mapping to be defined in terms of a 
representation that is general enough to carry over to several different language pairs. 
605 
Computational Linguistics Volume 20, Number 4 
CLCS: Syntax: 
Y-MAX 
Y-~x Q 
W X-MAX 
IN 
X Z 
Figure 8 
G£T¢ mapping between the CLCS and the syntactic structure. 
entries that are not based on purely syntactic information, idiomatic usage, aspectual 
knowledge, discourse knowledge, domain knowledge, or world knowledge). 
Before we define and resolve each divergence type, we will first make some revi- 
sions to the representations used in Definitions 1 and 4 to simplify the presentation. 
The representation given in Definition 1 is revised so that Z' is used to denote a logical 
argument from the set {Z~l ... Z~n} and Q' is used to denote a logical modifier from 
the set {Q'I ... Q'm}. The resulting representation is considerably simplified: 
(8) \[T(X') X' (\[T(W') Wt\], \[T(Z') Z'\], \[T(Q') Q'\])\] 
Similarly, the representation given in Definition 4 is revised so that W is used to denote 
the external argument, Z is used to denote an internal argument from the set {Z-MAX1 
... Z-MAXn}, and Q is used to denote a syntactic adjunct from the set {Q-MAXI ... 
Q-MAXm}. The resulting representation has the following simplified form: 
(9) \[Y-MAX \[Y-MAX W IX-MAX X Z\]\] Q\]8 
With these simplifications, the G£T4 can be conceptualized as the following set of 
relations: 
(10) Simplified G£T4: 
1. X' ~X 
2. W~ ~W 
3. Z' ~Z 
4. Q'~Q 
Figure 8 shows the simplified ~£T4 in terms of tree-like representations. 9 
We are now prepared to define and resolve the translation divergences of Fig- 
ure 1 on the basis of the simplified formalization presented in (8)-(10) above. The 
8 For the purposes of this discussion, we will retain the convention that syntactic adjuncts occur on the 
right at the maximal level. Note that this is not always the case: the setting of an adjunction parameter 
(described by Dorr \[1993b\]) determines the side and level at which a particular adjunct will occur. 
9 For ease of illustration, this diagram omits the type specification. (There is no loss of generality, since 
the G£'R mapping does not make use of this specification.) We will retain this convention throughout 
the rest of this paper. 
606 
Bonnie J. Dorr Machine Translation Divergences 
(a) Thematic Divergence 
CLCS: Syntax: 
Y-MAX 
I\ 
Y-MAX Q 
/ I 
X-MAX 
I\ 
X W 
(b) Promotional Divergence 
CLCS: Syntax: 
Y-MAX 
/i 
X-MAX 
! \ 
Q X 
(c) Demofional Divergence 
CLCS: Syntax: 
I\ 
Y-MAX X 
X-MAX 
I 
Z 
(d) Structural Divergence 
CLCS: 
~ Y-M~0C 
Y-MAX O 
X-MAX 
IN 
X R 
I 
Z 
(e) Conflational Divergence 
CLCS: Syntax: 
/I 
X-MAX 
I 
x 
Q 
Figure 9 
Translation mappings for cases in which G£~ default positions are overridden. 
solution to the divergence problem relies solely on three types of information: the 
G£T4; the CST4; and a small set of parametric mechanisms. The G£T4 and C$T4 are 
intended to be language independent, whereas the parameters are intended to encode 
language-specific information about lexical items. Because the interlingual representa- 
tion preserves relevant lexical-semantic relations, these three types of information are 
all that are required for providing a systematic solution to the divergence types shown 
in Figure 1. In particular, the solution given here eliminates the need for transfer rules 
and relies instead on parameterized mappings that are defined and applied uniformly 
across all languages. Seven parameters are used to invoke exceptions to the ~£T4 and 
C$T4 functions in the context of translation divergences: :INT, :EXT, :PROMOTE, :DE- 
MOTE, ,, :CAT, and :CONFLATED. We will now present a formal description of each 
divergence type and its associated parameter. 
4.1 Thematic Divergence 
The first divergence type to be formalized is the one for thematic divergence, i.e., the 
repositioning of arguments with respect to a given head. This type of divergence arises 
607 
Computational Linguistics Volume 20, Number 4 
in cases in which the ~£T4 invokes the following sets of relations in place of steps 2 
and 3 of (10): 
(11) 2.' W' ~ Z 
3/Z l ~ W 
Figure 9a shows the revised mapping. 
Thematic divergence arises only in cases in which there is a logical subject. An 
example of thematic divergence is the reversal of the subject with an object, as in 
the thematic divergence example given earlier in (1). The syntactic structures and 
corresponding CLCS are shown here: 
(12) \[C-MAX \[I--MAX IN-MAX I\] \[V-MAX \[V like\] \[N-MAX Mary\]\]\]\] 
\[State BEIdent (\[Thing II, 
\[Position aTIdent (\[Thing I\], \[Thing MARY\])\], 
\[Manner LIKINGLY\])\] 
\[C-MAX \[I-MAX IN-MAX Maria\] \[V-MAX IV me gusta\]l\]\] 1° 
Here the object Mary has reversed places with the subject I in the Spanish translation. 
The result is that the object Mary turns into the subject Maria, and the subject I turns 
into the object me. 
This argument reversal is resolved by means of the :INT and :EXT parameters, 
which force the ~£T4 mapping to be overridden with respect to the positioning of the 
logical subject and logical argument in Spanish. The lexical entries for like and gustar 
illustrate the difference in the use of these parameters: 
(13) (i) 
(ii) 
Lexical entry for like: 
\[State BEIdent (\[Thing W\], 
\[Position aTIdent (\[Thing Wl, \[Thing Z\])\], 
\[Manner LIKINGLYI)\] 
Lexical entry for gustar: 
\[State BEldent (\[Thing :INT W\], 
\[Position aTIdent (\[Thing W\], \[Thing :EXT Z\])\], 
\[Manner LIKINGLY\])\] 
Because the English entry does not include these parameters, the translation relies on 
the default argument positionings imposed by the ~£T4. By contrast, the :INT/:EXT 
markers specified in the Spanish entry force the internal and external arguments to 
swap places in the syntactic structure. 
10 For the purposes of this discussion, the Spanish sentence is given in its uninverted form. There are 
other ways of realizing this sentence. In particular, a native speaker of Spanish will frequently invert 
the subject to post-verbal position: 
\[C-MAX \[I--MAX ei \[V-MAX IV--MAX \[V me gusta\]\] IN--MAX Mafia\]i\]\]\]. 
However, this does not affect the internal/external reversal scheme described here, since inversion is a syntactic operation that takes place independently of the process that handles thematic divergences. 
608 
Bonnie J. Dorr Machine Translation Divergences 
The general solution to thematic divergence is diagrammed as follows: 
(14) RLCS 1: \[T(X,) X' (\[T(W') W'\], \[r(z,) Z'\] \[T(Q') Q'\])\] 
RLCS 2: \[r(x,) X' (\[r(w') :INT W'\], \[T(Z') :EXT Z'\] \[T(Q') Q'I)\] 
Translation: \[Y-MAX \[Y-MAX W \[X-MAX X Z\]\] Q\] 
\[T(X') X' (\[T(W') W'\], \[T(Z') Z'\] \[T(Q') Q'\])\] 
\[Y-MAX \[Y-MAX Z \[X-MAX X Wll QI 
This assumes that there is only one external argument and zero or more internal 
arguments. If the situation arises in which more than one variable is associated with 
the :EXT markers, it is assumed that there is an error in the word definition. N Note 
that the :INT and :EXT markers show up only in the RLCS. The CLCS does not include 
any such markers, since it is intended to be a language-independent representation 
for the source- and target-language sentences. 
Thematic divergence is one of three types of possible positioning variations that 
force the G£T4 to be overridden. Two additional positioning variations are promo- 
tional and demotional divergences, which will be defined in the next two sections. 
Whereas thematic divergence involves a repositioning of two satellites relative to a 
head, promotional and demotional divergences involve a repositioning of the head 
itself\] 2 We will see in Section 5.1 that these three divergences account for the entire 
range of repositioning possibilities. 
4.2 Promotional Divergence 
Promotional divergence is characterized by the promotion (placement "higher up") of 
a logical modifier into a main verb position (or vice versa), as shown in Figure 9b. 
In such a situation, the logical modifier is associated with the syntactic head position, 
and the logical head is then associated with an internal argument position. Thus, 
promotional divergence overrides the G£T4, invoking the following sets of relations in 
place of steps 1 and 4 of (10): 
(15) 1.' X ! ~ Z 13 
4.'Q' ¢~X 
Figure 9b shows the revised mapping. 
11 The parameters associated with the RLCS are assumed to be correctly specified for the purposes of this 
formal description. However, in practice, there might be errors in the lexical entries, since they are 
constructed by hand in the current implementation. Eventually, the intent is to automate the process of 
lexical entry construction so that these errors can be avoided. 
12 The notions of demotion and promotion are not the same as the notions of demotion and advancement 
in the theory of relational grammar (see Perlmutter 1983). Dorr (1993b, pp. 269-274) argues that, 
although the relational representation might be a convenient tool for illustrating the promotion and 
demotion operations as used in the current approach, this representation is not an appropriate vehicle 
for interlingual translation for a number of reasons. 
13 This relation does not mean that X replaces Z (if there is a Z), but that X retains the same structural 
relation with Z (i.e., Z remains an internal argument of X). To simplify the current description, Z is not 
shown in the syntactic structure of Figure 9b. 
609 
Computational Linguistics Volume 20, Number 4 
An example of promotional divergence is the case given earlier in (2). The syntactic 
structures and corresponding CLCS are shown here: 
(16) \[C-MAX \[I-MAX \[N--MAX John\] 
\[V-MAX IV usually Iv goes\]\] IN-MAX home\]\]\]l 
\[Event nOLoc (\[Thing JOHN\], 
\[Path TOcoc (\[Position ATcoc (\[Thing JOHN\], \[Location HOUSE\])\])\], 
\[Manner HABITUALLY\])\] 
\[C-MAX \[I--MAX IN-MAX Juan\] 
\[V-MAX Iv suele\] \[V-MAX \[V ir\] \[P-MAX a casa\]\]\]\]\] 
Here the main verb go is modified by an adverbial adjunct usually, but in Spanish, 
usually has been placed into a higher position as the main verb soler, and the "going 
home" event has been realized as the internal argument of this verb. 
Promotional divergence is resolved by the :PROMOTE parameter, which forces 
the ~£T4 mapping to be overridden with respect to the positioning of the logical head 
and the logical modifier. The lexical entries for usually and soler illustrate the difference 
in the use of this parameter: 
(17) (i) 
(ii) 
Lexical entry for usually: 
\[Manner HABITUALLY\] 
Lexical entry for soler: 
\[Manner :PROMOTE HABITUALLY\] 
Because the English entry does not use this parameter, the translation relies on the de- 
fault argument positionings imposed by the G£T4. By contrast, the :PROMOTE marker 
specified in the Spanish entry forces the head and adjunct to swap places in the syn- 
tactic structure. 
The general solution to promotional divergence is diagrammed as follows: 
(18) RLCS 1: \[T(Q') Q'\] 
RLCS 2: \[T(Q') :PROMOTE Q'\] 
Translation: \[Y-MAX \[Y-MAX W \[X--MAX X Z\]\] Q\] 
\[T(X') X' (\[T(W') W'\], \[T(Z') Z'\] \[T(Q') Q'\])\] 
\[Y-MAX \[Y-MAX W \[X-MAX Q \[ ... X Z\]\]\]\] 
4.3 Demotional Divergence 
Demotional divergence is characterized by the demotion (placement "lower down") of 
a logical head into an internal argument position (or vice versa), as shown in Figure 9c. 
In such a situation, the logical head is associated with the syntactic adjunct position, 
and the logical argument is then associated with a syntactic head position. Thus, 
610 
Bonnie J. Dorr Machine Translation Divergences 
demotional divergence overrides the g£T4, invoking the following sets of relations in 
place of steps 1 and 3 of (10): 
(19) 1.' X' 4:~ Q14 
3.' Z' 4:> X 
Figure 9(c) shows the revised mapping. 
An example of demotional divergence is the case given earlier in (3). The syntactic 
structures and corresponding CLCS are shown here: is 
(20) \[C-MAX \[I-MAX IN-MAX I\]i IV-MAX IV like\] \[C-MAX PROi to eat\]\]\]\] 
\[State BEcirc (\[Thing I\], 
\[Position ATcirc (\[Thing I\], \[Event EAT (\[Thing I\], \[Thing FOOD\])\])\] 
\[Manner LIKINGLY\])\] 
\[C-MAX \[I-MAX IN-MAX Ich\] IV-MAX IV \[ADV gern\] Iv esse\]\]\]\]\] 16 
Here the main verb like takes the "to eat" event as an internal argument; but in German, 
like has been placed into a lower position as the adjunct gern, and the "eat" event has 
been realized as the main verb. 
The distinction between promotional and demotional divergences may not be in- 
tuitively obvious at first glance. In both cases, the translation mapping appears to 
associate a main verb with an adverbial satellite, or vice versa. However, the dis- 
tinction between these two head switching cases becomes more apparent when we 
consider the status of the participating lexical tokens more carefully. In the case of 
soler-usually, the main verb soler is, in some sense, the token that "triggers" the head 
switching operation: its presence forces the adverbial satellite usually to appear in En- 
glish, even if we were to substitute some other event for ir in Spanish (e.g., correr a la 
tienda, leer un libro, etc.). By contrast, in the case of like-gern, the triggering element is 
not the main verb like, since we are able to use like in other contexts that do not require 
gern (e.g., I like the car ~ Mir gefdllt der Wagen); instead, the triggering element is the 
adverbial satellite gern: its presence forces the verb like to appear in English even if 
we were to substitute some other event in place of essen in German (e.g., zum Geschdft 
laufen, das Buch lesen, etc.). We will return to this point in Section 5.2. 
Demotional divergence is resolved by the :DEMOTE parameter, which forces the 
~£T4 mapping to be overridden with respect to the positioning of the logical head 
and the logical argument. The lexical entries for like and gern illustrate the difference 
in the use of this parameter: 17 
14 This relation does not mean that X replaces Q (if there is a Q), but that X retains the same structural 
relation with Q (i.e., Q remains a syntactic adjunct of X). To simplify the current description, Q is not 
shown in the syntactic structure of Figure 9c. 
15 The default object being eaten is FOOD, although this argument does not appear on the surface for the 
current example. 
16 The German syntactic structure is shown here in the uninverted base form. In the German surface 
structure, the verb is moved up into verb-second position and the subject is topicalized: 
\[C--MAX IN--MAX IchJ/ Iv esse\]j \[I--MAX IN--MAX t\]l IV--MAX Iv \[ADV gem\] Iv t\]/\]\]/\]. 
17 Both definitions of like in (21) use the circumstantial field, which means that the Y argument must be 
an Event (e.g., like ta eat) rather than a Thing (e.g., like Mary). Thus, the definitions for like and gern are 
slightly different from the definitions of like given earlier in (13) (i.e., these are additional lexical entries 
for like). 
611 
Computational Linguistics Volume 20, Number 4 
(21) (i) 
(ii) 
Lexical entry for like: 
\[State BEcirc (\[Thing W\], 
\[position ATcirc (\[Thing W\], \[Event Z\])\], 
\[Manner LIKINGLY\])\] 
Lexical entry for gem: 
\[State BECirc (\[Thing W\], 
\[Position ATCirc (\[Thing W\], \[Event :DEMOTE Z\])\], 
\[Manner LIKINGLY\])\] 
Because the English entry does not use this parameter, the translation relies on the 
default argument positionings imposed by the ~£T4. By contrast, the :DEMOTE marker 
specified in the German entry forces the head and internal argument to swap places 
in the syntactic structure. 
The general solution to demotional divergence is diagrammed as follows: 
(22) RLCS 1: \[v(x') X' (\[T(W') W'\], \[T(Z') Z'\] \[T(Q') Q'\])\] 
RLCS 2: \[T(X') X' (\[T(W') W'\], \[T(Z') :DEMOTE Z'\] \[T(Q') Q'\])\] 
Translation: \[Y-MAX \[Y-MAX W IX-MAX X Z\]\] Q\] 
\[v(x') X' (\[T(W'/ W'\], \[T(Z') Z'\] \[T(Q'/ Q'\])l 
\[Y--MAX \[Y-MAX W IX-MAX Z\] \[ ... X QI\]\] 
4.4 Structural Divergence 
Structural divergence differs from the last three divergence types in that it does not 
alter the positions used in the ~£T4 mapping, but it changes the nature of the relation 
between the different positions (i.e., the "4=~" correspondence). Figure 9d characterizes 
the alteration that takes place. Note that the mapping of Z' to the corresponding 
internal argument position is altered so that it is positioned under the constituent that 
corresponds to W. 
An example of structural divergence is the case given earlier in (4). The syntactic 
structures and corresponding CLCS are shown here: 
(23) \[C-MAX \[I-MAX IN-MAX John\] 
IV-MAX IV entered\] \[N-MAX the house\]\]\]\] 
\[Event GOLoc (\[Thing JOHN\], 
\[Path TOcoc (\[Position INcoc (\[Thing JOHN\], \[Location HOUSE\])\])\])\] 
\[C-MAX \[I-MAX IN-MAX Juan\] 
\[V-MAX IV entr6\] \[P-MAX en \[N-MAX la casa\]\]\]\]\] 
Here the verbal object is realized as a noun phrase (the house) in English and as a 
prepositional phrase (en la casa) in Spanish. 
Structural divergence is resolved by means of the * marker, which forces logical 
constituents to be realized compositionally at different levels. In particular, the * serves 
as a pointer to a RLCS position that must be combined with another RLCS in order 
to arrive at a (portion of a) CLCS. The lexical entries for enter and entrar illustrate the 
difference in the use of this parameter: 
612 
Bonnie J. Dorr Machine Translation Divergences 
(24) (i) 
(ii) 
Lexical entry for enter: 
\[Event GOLoc (\[Thing W\], 
\[Path TOLoc (\[Position INLoc (\[Thing W\], \[Location * Z\])\])\])\] 
Lexical entry for entrar: 
\[Event GOLoc (\[Thing W\], 
\[Path * TOLoc (\[Position INcoc (\[Thing W\], \[Location Z\])\])\])\] 
Because the English entry contains a * marker in the \[Location Z\[ position, this constituent 
is realized on the surface as the object (i.e., the house) of the main verb. By contrast, the 
• marker is associated with a "higher" position \[Path TOcoc ...\] in the Spanish entry, 
thus forcing this constituent to have a more complex realization (i.e., en la casa) in the 
syntactic structure. 
The general solution to structural divergence is diagrammed as follows: 
(25) RLCS 1: \[T(X,) X' (\[T(W') W'\], \[T(R') R' (\[T(Z') * Z'\])\] \[r(Q,) Q'\])\] 
RLCS 2: \[r(x,) X' (\[r(w,) W'\], \[r(R') * R' (\[r(z') Z'\])\] \[T(Q') Q'\])\] 
Translation: \[Y-MAX \[Y-MAX 
\[r(x,) X' (\[T(W') 
\[Y-MAX \[Y-MAX 
W \[X-MAX X Z\]\] Q\] 
W'\], \[T(Z') Z'\] \[T(Q') Q'\])\] 
W \[X-MAX X \[ ... R Z\]\]\] Q\] 
Note that the logical argument R' is associated with a * marker in the RLCS of the 
target language, but not in the RLCS of the source language. This forces the target 
language syntactic structure to realize a phrase R that dominates Z; in contrast, no 
such dominating phrase occurs in the source-language structure. 
4.5 Conflational Divergence 
Conflational divergence is another case in which the "~" correspondence is changed. 
In particular, conflational divergence is characterized by the suppression of a CLCS 
constituent (or the inverse of this process). The constituent generally occurs in logical 
argument or logical modifier position; thus, the "4=>" correspondence of either step 3 
or step 4 of the ~£T4 is changed, depending on which position is conflated. Figure 9e 
characterizes the alteration that takes place. Note that the Z' position in the CLCS does 
not have a corresponding realization in the syntax. 
An example of conflational divergence is the case given earlier in (5). The syntactic 
structures and corresponding LCS are shown here: 
(26) \[C-MAX \[I-MAX IN-MAX I\] 
\[V-MAX \[V stabbed\[ \[N-MAX John\]\]\[\[ 
\[Event CAUSE 
(\[Thing I\], 
\[Event GOposs 
(\[Thing KNIFE-WOUND\], 
\[Path TOWARDposs 
(\[Position ATposs (/Thing KNIFE-WOUND\], \[Thing JOHN\])\])\[)\])\] 
\[C-MAX \[I-MAX IN-MAX Yo\] 
\[V-MAX \[v le di\] \[N-MAX pufialadas\] \[P-MAX a Juan\]\[\]\] 
613 
Computational Linguistics Volume 20, Number 4 
Here, English uses the single word stab for the two Spanish words dar (give) and 
pu~aladas (knife-wounds); this is because the effect of the action (i.e., the knife-wound 
portion of the lexical token) is incorporated into the main verb in English. 
Conflational divergence is resolved by means of the :CONFLATED marker, which 
suppresses the realization of the filler of a particular position. The lexical entries for 
stab and dar illustrate the difference in the use of this parameter: is 
(27) (i) 
(ii) 
Lexical entry for stab: 
\[Event CAUSE 
(\[Thing W\], 
\[Event GOposs 
(\[Thing :CONFLATED KNIFE-WOUND\], 
\[Path TOWARDposs 
(\[Position ATposs (\[Thing KNIFE-WOUND\], \[Thing Z\])\])\])\])l 
Lexical entry for dar: 
\[Event CAUSE 
(\[Thing W\], 
\[Event GOposs 
(\[Thing ~ Y\], 
\[Path TOWARDposs 
(\[Position ATposs (\[Thing Y\], \[Thing ZI)\])\])\])\] 
Because the English entry contains a :CONFLATED marker in the logical position cor- 
responding to KNIFE-WOUND, this constituent is suppressed in the syntactic struc- 
ture. By contrast, this marker does not appear in the corresponding position in Span- 
ish, thus forcing this constituent to be overtly realized (i.e., puflaladas) in the svntactic 
structure. 
The general solution to conflational divergence is diagrammed as follows: 
(28) RLCS 1: \[~(x,) X' (\[T(W') W'\], \[T(Z') * Z'\] \[T(Q') Q'\])\] 
RLCS 2: \[T(X') X' (\[z(w') W'\], \[T(Z') :CONFLATED Z'\] \[T(Q') Q'\])\] 
Translation: \[Y-MAX \[Y-MAX W \[X-MAX X Z\]\] Q\] 
\[T(X'~ X' (\[T(W'/ W'\], \[T(Z'} Z'\] \[r;Qq Q'\])\] 
\[Y-MAX \[Y-MAX W \[X-MAX XI\[ QI 
Note that the logical argument Z' is associated with a :CONFLATED marker in the 
RLCS of the target language, but not in the RLCS of the source language. This forces 
the target language syntactic structure to suppress the realization of this constituent. 
18 Note that the :CONFLATED marker appears to be in complementary distribution with the • marker. In 
fact, one might consider the use of the :CONFLATED marker to be unnecessary, since its presence 
could be implied by the absence of the * marker. However, the :CONFLATED marker plays an 
important role in the lexical-semantic representation: it specifies that the "constant" term (e.g., the 
KNIFE-WOUND of the stab RLCS) must obligatorily fill the position and, moreover, that this constant 
must be a legal LCS primitive of the system. In addition, there is an inherent asymmetry between the 
:CONFLATED marker and the * marker: whereas the former always occurs in a leaf node position, the 
latter may occur in any position in the RLCS. Because the notion of conflation is not meaningful in 
non-leaf node positions, it would be unreasonable to make the assumption that every non-leaf position 
without the * marker is conflated. The :CONFLATED marker is used to identify truly conflated 
positions, not just those positions without the * marker. 
614 
Bonnie J. Dorr Machine Translation Divergences 
4.6 Categorial Divergence 
Unlike the previous five divergence types, categorial divergence affects the operation 
of the C$/~, not the ~£T4. It is characterized by a situation in which CAT(0) is forced 
to have a different value than would normally be assigned to T(~ ~) by means of 
the mapping specified in Figure 7. Thus, categorial divergence is formally described 
as follows: a lexical-semantic type T(~') is related to a syntactic category CAT(~), 
where CAT(~) ~ CST4(T(~')). In such a case, CAT(~) must be specified through lexical 
parameterization. 
An example of categorial divergence is the case given earlier in (6). The syntactic 
structures and corresponding CLCS are shown here: 
(29) \[C-MAX \[I-MAX \[N-MAX I\] \[V-MAX \[V am\] \[A--MAX hungry\]\]\]\] 
\[State BEIdent (\[Thing I\], 
\[Position aTIdent (\[Thing I\], \[Property HUNGRY\])\])\] 
\[C-MAX \[I-MAX \[N-MAX Ich\] \[V-MAX IN-MAX Hunger\] \[v habe\]\]\]\] 19 
Here, the predicate is adjectival (hungry) in English but nominal (Hunger) in German. 
Categorial divergence is resolved by means of the :CAT parameter. The lexical 
entries for be and haben illustrate the difference in the use of this parameter: 2° 
(30) (i) 
(ii) 
Lexical entry for be: 
\[State BEIdent (\[Thing W\], 
\[Position aTIdent (\[Thing W\], \[Property Y\])\])\] 
Lexical entry for haben: 
\[State BEIdent (\[Thing W\], 
\[Position aTIdent (\[Thing W\], \[Property :CAT(N) Y\])\])\] 
Because the English entry does not contain a :CAT marker in the position correspond- 
ing to \[Property Y\], this constituent is realized in the syntactic structure as C$/~(Property) 
Adjective (i.e., hungry). By contrast, the :CAT(N) marker specified in the German 
entry forces the default C$T4 category to be overridden and the Property argument is 
realized as a noun form (i.e., Hunger). 
19 The German structure shown here is the base form of the following surface syntactic tree: 
\[C-MAX IN-MAX Ich\]i 
\[C Iv babe\]j\] 
\[I--MAX \[N--MAX tli \[V-MAX \[N-MAX Hunger\] \[v t\]jI \[I e\]\]\] 
This form is derived by the syntactic processor after lexical-semantic processing is complete (see Dorr 
1993a, for details). 
20 As might be expected, there are two lexical entries for the verb haben, only one of which is listed here. 
The one not shown here corresponds to the possessional sense of haben, i.e., the meaning 
corresponding to the word have in English. This entry does not contain a :CAT marker. 
615 
Computational Linguistics Volume 20, Number 4 
The general solution to categorial divergence is diagrammed as follows: 
(31) RLCS 1: \[T(X,) X' (\[T(W') W'\], IT(Z,) :Z'\] \[T(Q') Q'\])\] 
RLCS 2: It(x,) X' (\[7(w') W'\], IT(Z,) (:CAT 6) Z'\] \[~(Q,) Q'\])\] 
Translation: \[Y-MAX \[Y-MAX W IX-MAX X Z\]\] Q\] 
\[T(X') X' (\[T(W') W'\], \[T(Z') Z'\] \[T(Q') Q'\])I 
\[Y-MAX \[Y-MAX W \[X-MAX X a\]\] QI 
where CAT(Z) = &MAX. 
4.7 Lexical Divergence 
Lexical divergence arises only in the context of other divergence types. 21 This is be- 
cause the choice of lexical items in any language relies crucially on the realization and 
composition properties of those lexical items. Because the six preceding divergences 
potentially alter these properties, lexical divergence is viewed as a side effect of other 
divergences. Thus, the formalization thereof is considered to be some combination of 
those given above. 
Unlike the first six divergence types, lexical divergence is solved during the pro- 
cess of lexical selection. 22 Thus, there is no specific override marker that is used for 
this type of divergence. For example, in the lexical divergence (7), a conflational di- 
vergence forces the occurrence of a lexical divergence. The syntactic structures and 
corresponding CLCS for this example are shown here: 
(32) \[C-MAX \[I-MAX \[N-MAX John\] 
IV--MAX IV broke\] \[P-MAX into IN--MAX the room\]\]\]\]\] 
\[Event CAUSE 
(\[Thing JOHN\], 
\[Event GOLoc 
(\[Thing JOHN\], 
\[Path TOLoc 
(\[Position INcoc (\[Thing JOHN\], \[Location ROOMI)\])I)\] 
\[Manner FORCEFULLY\])\] 
\[C-MAX \[I-MAX \[N-MAX Juan\] 
IV-MAX IV forz6\] IN-MAX la entrada\] \[P-MAX al cuarto\]\]\]\] 
Because the word-particle pair break into subsumes two concepts (forceful spatial mo- 
tion and entry to a location), it is crucial that the word forzar (literally, force) be selected 
in conjunction with entrada (literally, entry) for the underlying break-into concept. 
21 As noted by a reviewer, this is not strictly true, since there are many cases in which a source-language 
word maps to more than one target-language word without the simultaneous occurrence of another 
divergence type. An example of such a case is the English word eat, which maps to essen (for humans) 
or fressen (for animals). Such cases are considered to be outside of the classes of lexical-semantic 
divergences considered here (see Footnote 3). However, a simple approach to resolving such cases 
would be to use featural restrictions during syntactic processing. 
22 The solution to lexical divergence is trivial for transfer machine translation systems, since transfer 
entries map source-language words directly to their target-language equivalents. In general, lexical 
selection is not seen as a problem in these systems. 
616 
Bonnie J. Dorr Machine Translation Divergences 
There is also a structural divergence in this example, since the prepositional phrase 
into the room must be translated into a noun phrase entrada al cuarto. This divergence 
compounds the lexical divergence problem, since it is necessary to choose the target- 
language word a in the absence of a source-language counterpart. 
Lexical divergence also shows up in three previously presented examples, (12), 
(29), and (26), owing to the presence of thematic, categorial, and conflational diver- 
gences, respectively: in (12) the word like is chosen for the word gustar (literally, to 
please); in (29) the word haben (literally, to have) is chosen for the word be; and in (26) 
the word dar (literally, to give) is chosen for the word stab. 
5. Discussion 
This section discusses certain issues of relevance to the formal classification and reso- 
lution of translation divergences. In particular, we will discuss (1) the limits imposed 
on the range of repositioning possibilities; (2) the justification for distinguishing be- 
tween promotional and demotional divergences; (3) the notion of full coverage in the 
context of lexical selection; and (4) the resolution of interacting divergence types. 
5.1 Limits on Repositioning Divergences 
In Section 4.1 we made the claim that the thematic, promotional, and demotional 
divergences account for the entire range of repositioning possibilities. We will now 
explore the validity of this claim. 
There are two potential types of syntactic relations that exist between a head and 
a satellite: the first is complementation (i.e., involving the internal argument), and 
the second is adjunction. 23 Given these two types of relations, there are only a small 
number of ways syntactic entities may be repositioned. The three CLCS positions that 
are involved in these relations are X', Z ~, and Q'. If we compute the repositionings 
combinatorically, there are 33 = 27 configurations (i.e., X ~, Z', and Q~ would map into 
any of three positions). However, we can eliminate 15 of these (since a CLCS must 
contain exactly one head), thus leaving only 12 possible configurations. One of these 
corresponds to the default G£~ mapping (i.e., the logical head, logical argument, and 
logical modifier map into canonical positions). The remaining 11 configurations can 
be factored into three cases as follows: 
1. X' 4=> X 
1.1 Q' ~ Z; Z 1 4=~ Z. 
1.2 Z' ~ Q; Q' 4=~ Q. 
1.3 Q' <~ Z; Z' ~ Q. 
2. Q' 4=¢ X 
2.1 X ~ 4=~ Z; Z' ¢¢ Z. 
2.2 X ~Z;Z ~<=~Q. 
23 We have left out the possibility of an external argument as a participant in the head-satellite relation. 
Of course, the external argument is a satellite with respect to the head, but it turns out that the external 
argument, which corresponds to the logical subject in the CLCS, has a special status and does not have 
the same repositioning potential that internal arguments and syntactic adjuncts have. In particular, the 
external argument has the unique property that it never participates as the incorporated argument of a 
conflational verb. Hale and Keyser (1989) provide evidence that this property holds across all 
languages. Thus, we take the external argument to have a special status (universally) that exempts it 
from participating in divergences other than thematic divergence. 
617 
Computational Linguistics Volume 20, Number 4 
(a) Case 1.1 CLCS: 
X~ Y-MAX\] 
Y-MAX 
71 X-MAX 
IN 
x Q 
(b) Case 1.2 CLCS: 
Syntax: 
Y-MAX 
I\ Y-MAX Z 
/I X-MAX 
I 
x 
(e) Case 2.3 CLCS: Syntax: 
IN Y-MAX X 
/I Z~, X-MAX 
t\ 
• Q z 
(d) Case 3.2 CLCS: Syntax: 
Y-MAX 
X-MAX 
IN Z X 
Q 
Figure 10 
Illegal translation mappings for natural language. 
. 
2.3 X' ~ Q; Z' 4~ Z. 
2.4 X'~Q;Z'~Q. 
Z'~X 
3.1 X'~Z;Q'~Z. 
3.2 x' <=~ Z; Q' ~=~ Q. 
3.3 X' ~ Q; Q' <=~ Z. 
3.4 X' <=~ Q; Q' <=~ Q. 
We will discuss in detail how four of these cases, i.e., the ones characterized in Fig- 
ure 10, are ruled out. Of the remaining, 2.1 and 3.4 correspond to the definitions of 
promotional and demotional divergences illustrated in Figures 9b and 9c, respectively; 
cases 1.3, 2.2, 2.4, 3.1, and 3.3 are ruled out for the same reasons that cases 1.1 and 1.2 
are ruled out (as we will see shortly), namely, that an internal argument Z can never 
be associated with a logical modifier Q' and that a syntactic adjunct (Q) can never be 
associated with a logical argument Z'. 
It cannot be the case that a logical modifier maps to an internal argument position 
(case 1.1) because a logical modifier is an optional participant of a particular action, 
i.e., it need not be "governed" by the lexical item that it modifies. Internal argument 
positions are reserved for cases in which a government relation must hold; thus, logical 
modifiers must necessarily be mapped into syntactic adjunct positions. 
Similarly, it also cannot be the case that a logical argument maps to a syntactic 
adjunct position (case 1.2). A logical argument is a necessary participant of a particular 
action, and as such, it must be "governed" by the lexical item that selects it. By contrast, 
adjunct positions are reserved for optional modifying participants that do not need to 
be governed by the lexical item that they are modifying; thus, logical arguments must 
necessarily be mapped into internal argument positions. 
618 
Bonnie J. Dorr Machine Translation Divergences 
Another case that is eliminated is the renaming of a logical head as a syntactic 
adjunct whose head corresponds to a logical modifier (case 2.3). The idea is simply 
that modification is a one-way relation. If a logical head has a modifier, the head 
cannot become an adjunct of that modifier because the modifying relation would be 
reversed (i.e., the logical head would modify the syntactic head rather than the other 
way around). In contrast, a logical head of a CLCS can be mapped to an internal 
argument position in cases in which a logical modifier is mapped to a syntactic head 
(i.e., the case of promotional divergence presented earlier), since there is no violation 
of the one-way relation. 
A similar argument is used to eliminate the case in which a logical head is mapped 
to an internal argument whose head corresponds to a logical argument (case 3.2). The 
idea is that heads and arguments participate in a one-way relation. If a logical head 
has an argument, the head cannot become an internal argument of that argument 
because the head-argument relation would be reversed (i.e., the logical head would be 
an argument of the syntactic head rather than the other way around). In contrast, the 
logical head of a CLCS can be mapped to an adjunct (i.e., modifier) position in cases 
in which a logical argument is mapped to a syntactic head (i.e., the case of demotional 
divergence presented earlier), since these is no violation of the one-way relation. 
The argument for the elimination of the last two cases could be viewed as an appeal 
to a constraint that is analogous to the 0-criterion in syntax. Essentially, this constraint 
states that all arguments and modifiers must be licensed (see Chomsky 1986a; Abney 
1989) in order to appear either in the syntactic structure or in the conceptual structure. 
In the context of conceptual structure, a logical modifier may license the realization of 
a logical head in an internal argument position, but not in an adjunct position, since 
the modifier relation is already satisfied by virtue of the relation between the head 
and the modifier. Similarly, a logical argument may license the realization of a logical 
head in a syntactic adjunct position, but not in an internal argument position, since 
the head-argument relation is already satisfied by virtue of the relation between the 
head and the argument. Having eliminated the meaningless possibilities, we are left 
with the promotional and demotional cases presented above. 
5.2 Promotional versus Demotional Divergences 
We will now provide justification for the earlier claim that promotional and demo- 
tional divergence should be classified differently, even though they exhibit some of 
the same properties. It might be argued that these divergences are essentially the same, 
since both cases involve an association of a main verb with an adverbial satellite or 
vice versa. In the examples given earlier, the promotional divergence referred to a map- 
ping between the adverbial usually and the main verb soler and demotional divergence 
referred to a mapping between the adverbial gern and the main verb like. However, 
as mentioned in Section 4.3, these are taken to be in distinct classes: the difference 
between these two cases is determined by the "triggering" element (i.e., promotion is 
triggered by a main verb such as soler, whereas demotion is triggered by an adverb 
such as gern). 
Another factor that distinguishes between promotional and demotional diver- 
gences is the fact that verbs such as like and verbs such as soler do not have parallel 
syntactic distributions, nor do they have analogous logical interpretations. The verb 
like may take a sentential complement that has its own event structure (as in I like to 
eat), or it may take a nominal complement without an event structure (as in I like the 
car). In either case, the verb like generally means the same thing (i.e., it describes a 
state in which an event or a thing is somehow desirable to that person). By contrast, 
the verb soler is a modal verb that contributes an aspectual component of meaning 
619 
Computational Linguistics Volume 20, Number 4 
that crucially relies on a verbal complement with an event structure; in a sense, soler 
is analogous to the modal must in English in that it cannot be used in isolation, but 
requires the presence of a verbal complement in order for it to be interpretable. In such 
a configuration, the modal soler allows the event to be interpreted as being habitual 
in nature. 
Given these distinctions, it would not be appropriate to consider the head switch- 
ing mapping to be the same for the soler-usually and like-gern cases. Not only do they 
have different triggering elements (i.e., the main verb in the former and the adverb in 
the latter), but they do not have identical syntactic distributions and their logical in- 
terpretations are not analogous. Thus, they are taken to be two independent mappings 
with entirely different syntactic and lexical-semantic ramifications. 
The handling of promotional and demotional divergences is a topic that has re- 
ceived recent attention, although it has been labeled differently, depending on how it 
is manifested. An example of such a case is the way construction. This phenomenon 
has been studied by Jackendoff (1990) in his extended version of the original LCS 
framework: 
(33) Bill belched his way out of the restaurant 
In such cases, Jackendoff claims that belching is subordinated to a higher predicate like 
GO, "in effect demoting the meaning of the lexical verb to a subordinate accompani- 
ment or means 'modifier" (Jackendoff 1990, p. 214). This characterization is essentially 
equivalent to that of the soler-usually example (i.e., promotional divergence) given 
earlier. 24 
5.3 Lexical Selection: Full Coverage Constraint 
Because of the compositional nature of the LCS representation, the current framework 
automatically imposes a full coverage constraint during the lexical selection process. 
Formally, this constraint is defined as follows: 
(34) Full coverage constraint: 
A RLCS R matches a CLCS C if and only if R fully covers C. 
where R fully covers C under the following conditions: 
(35) A RLCS R fully covers a CLCS C if and only if: 
(a) there is no portion of R that does not match C; 
24 Jackendoff's approach to handling the way construction has been criticized by Marantz (1992) for its 
use of arbitrary exceptions to the "usual mappings." Marantz takes issue with the characterization of 
such cases as an idiosyncratic relation between syntax and semantics and proposes, instead, that the 
conceptual structure looks different from what Jackendoff envisions. Whichever of these proposals is 
correct, neither Jackendoff nor Marantz considers their proposals in the context of interlingual machine 
translation. If the exceptional mappings are indeed arbitrary, then one needs to explain how this affects 
the handling of different languages. Moreover, neither Jackendoff nor Marantz mentions the possibility 
that the number of exceptional mappings might not be arbitrarily large, but that there might be a fixed 
number of exceptions, delineated in such a way that only a handful need to be considered at any time 
for any given language. This is why the formalization described in this paper is a valuable resource: it 
provides a means for proving that only certain types of exceptions are allowed and that the number of 
such exceptions is actually quite small. Finally, neither Jackendoff nor Marantz considers keeping 
Jackendoff's version of the LCS intact and using a single parameterized mapping along the lines 
proposed in the current framework. 
620 
Bonnie J. Dorr Machine Translation Divergences 
(b) either R completely matches C (i.e., there is no portion of C that 
does not match R) or R matches all of C except some portion C' 
(i.e., a subcomponent of C) that is fully covered by some other 
RLCS R/. 
In cases in which more than one lexical entry matches a current concept, this constraint 
is used to determine which possibilities should be ruled out (if any). 
One of the main advantages to the formalization defined in this paper is that it 
allows one to judge whether a target-language concept fully covers the concept underly- 
ing the source-language sentence and, thus, to make an evaluation of the status of the 
system. As an illustration of this point, consider the stab-dar example of Section 4.5. 
The notion of full coverage is manifested through the lexical-selection process. The basic 
idea is that RLCSs are chosen such that they entirely cover the underlying concept. 
In terms of the mapping from the source-language sentence to the CLCS, this implies 
that the RLCSs must compose in such a way as to provide a full cover (i.e., there must 
be a path from the root to the leaf nodes that includes all the \[content\] words of the 
sentence). In terms of the mapping from the CLCS to the target-language sentence, 
this implies that the CLCS must be decomposed into (potentially overlapping) RLCSs 
whose "union" covers the entire CLCS. 
From the English sentence in (26), the RLCSs must be chosen for each word in the 
sentence such that they provide a coherent CLCS. The RLCS for stab is 25 
(36) \[Event CAUSE 
(\[Thing * W\], 
\[Event GOposs 
(\[Thing Y KNIFE-WOUND :CONFLATED\], 
\[Path TOWARDposs 
(\[Position ATposs (\[Thing Y KNIFE-WOUND\[, \[Thing * Z\])\])\])\])\[ 
Because Y does not have a * specification, only W and Z need to be filled in. Once 
these positions are filled in, the resulting CLCS fulfills the full coverage requirement, 
since there is a path from the root to the leaves that covers all of the words of the 
source-language sentence. The resulting CLCS is 
(37) \[Event CAUSE 
(\[Thing I\], 
\[Event GOposs 
(\[Thing KNIFE-WOUND\[, 
\[Path TOWARDposs 
(\[Position ATposs (\[Thing KNIFE-WOUND\[, \[Thing JOHN\])\])\])\])\] 
To complete the translation, this CLCS must be decomposed into RLCSs that satisfy 
the full coverage requirement. The RLCS that is selected as a match for this top-level 
CLCS is that of the word dar: 
25 In (27) we abbreviated the lexical entries for stab and dar, showing the * marker only in the position 
that was relevant to example (26). In 36 and 38, we show the complete form of the lexical entries to 
illustrate the notion of full coverage, 
621 
Computational Linguistics Volume 20, Number 4 
(38) \[Event CAUSE 
(\[Thing * W\], 
\[Event GOposs 
(\[Thing * YI, 
\[Path * TOWARDposs 
(\[Position ATposs (\[Thing Y\], \[Thing ZI)\])I)\])I 
Unlike the RLCS for stab, the Y position is associated with a • marker; thus, it is 
necessary to find RLCSs for all three positions W, Y, and Z. Positions W and Y are 
filled at the leaf level and Z is filled at the TOWARDposs level. Once these positions 
are filled, the combination of the RLCSs for yo, dar, pu~aladas, a, and Juan covers the 
entire concept. Thus, the full coverage requirement is satisfied. 
It should be noted that a number of other systems have attempted to tackle diver- 
gences similar to those discussed in this paper without appealing to the notion of full 
coverage. Three examples of such systems are (1) GETA/ARIANE (Vauquois and Boitet 
1985; Boitet 1987); (2) LMT (McCord 1989); and (3) METAL (Alonso 1990; Thurmair 
1990). In particular, these approaches address the problem of thematic divergence by 
means of transfer rules of the following form, respectively: 
(39) 
(40) 
(41) 
like(SUBJ(ARG2:GN),OBJI(ARGI:GN)) 
plaire(SUBJ(ARGI:GN),OBJI(ARG2:PREP, GN)) 4~ 
gverb (like (dat : ,, nom: X), ge+f all, • : X) 
like V ~ gustar V 
NP (\[ROLE SUBJD ~ NP (\[ROLE IOBJ\]) 
NP (\[ROLE DOBJ\]) ~ NP (\[ROLE SUBJ\]) 
One problem with these approaches is that surface syntactic decisions are, in a sense, 
performed off-line by means of lexical entries and transfer rules that specifically en- 
code language-specific syntactic information. Such a scheme is limited in that it has 
no potential for relating thematic divergence to the rest of the space of divergence 
possibilities. Moreover, although transfer rules might be deemed suitable for local di- 
vergences such as simple subject-object reversal, it is well known that simple transfer 
rules of this type do not readily accommodate more complicated divergences. 
Consider a more complicated case such as the following promotional divergence: R6 
(42) Promotional divergence: 
E: The baby just fell ~ F: Le b6b6 vient de tomber 
'The baby just (verb-past) of fall' 
Here, the English adverbial just is translated as the French main verb venir, which 
takes the falling event as its complement de tomber. 
At first glance, it might seem difficult to construct transfer rules that handle such 
cases. However, the LFG-MT system by .Kaplan et al. (1989) does, in fact, handle 
such cases by means of mappings between source and target functional structures (f- 
structures). The f-structures that correspond, respectively, to the English and French 
sentences in this example are the following: 
26 This example was taken from Kaplan et al. (1989). 
622 
Bonnie J. Dorr Machine Translation Divergences 
(43) (i) 
(ii) 
PRED 
ARG 
'JUST((T ARG))' 
PRED 'FALL((T SUBJ))' 
TENSE PAST 
PRED 'BABY' 
NUM SG SUBJ 
DEF 
SPEC PRED 
+ 
'THE' \] 
PRED 
SUBJ 
XCOMP 
'VENIR((T 
PRED 
GENDER 
NUMB 
SPEC 
PRED 
COMPL 
TENSE 
SUBJ 
SUBJ)(T XCOMP)}' 
'Bt~Bt~' 
MASC 
SG 
PRED 'LE' 
'TOMBER((T SUBJ)}' 
DE 
INF 
The translation mapping is performed by a transfer equation that relates the 
source- and target-language f-structures: 
(44) (T T PRED 'JUST((T ARG))') = VENIR (T T XCOMP) = T (T ARG) 
This equation identifies venir as the corresponding French predicate, and it maps the 
argument of just to a complement that is headed by the prepositional complementizer 
de. 
Although such a case is handled in the LFG-MT system, there are a number of 
problems with this approach. A serious flaw concerns the handling of divergences in 
the context of embedded clauses. (For additional discussion, see Sadler and Thompson 
1991.) In particular, if the English sentence in (42) were realized as an embedded 
complement such as I think that the baby just fell, it would not be possible to generate 
the French output. The reason for this is that the LFG-MT system is not designed 
to handle an interaction between a (divergent) matrix clause and a (nondivergent) 
embedded clause. This sentence is broken down into predicate-argument relations 
that conform (roughly) to the following logical specification: 
(45) think(I,fall(baby)) 
just(fall(baby)) 
Because the logical constituent fall(baby) is viewed as an argument of two logical 
heads, "think" and "just," the LFG-MT generator cannot determine how to compose 
these concepts and produce an output string. 
The divergence solution proposed in the LCS framework overcomes this difficulty 
by imposing the full coverage constraint. In particular, specific relations are set up 
623 
Computational Linguistics Volume 20, Number 4 
between logical heads and their associated arguments and modifiers so that there 
could never be any question of how two concepts are composed, even for embedded 
cases. For the current example, the LCS approach would reduce the logical relations 
to a single specification for both French and English: 
(46) think(I, fall(baby, just)) 
That is, the "just" component of meaning is a modifier of the "falling" action, regard- 
less of how this constituent is realized on the surface. 
The full benefit of this approach is further demonstrated when one considers the 
extent to which the full coverage constraint carries over to other divergence categories. 
Consider the following (frequently cited) conflational divergence: 
(47) Conflational divergence: 
E: John swam across the river 
F: John a travers6 la rivi~re ~ la nage 
'John crossed the river by swimming' 
In this example, the English path component (across) is translated as the French main 
verb (traverser), and the English main verb (swim) is translated as the French manner 
component (a la nage). In a system like TAUM (Isabelle 1987), transfer rules impose 
lexical transformations in order to establish correspondences between source and target 
structures, i.e., rules of the following form: 
(48) (X swim (across Y)) 4~ (X traverser Y (a la nage)) 
This rule maps the path component (across) to the French main verb (traverser), and 
the English main verb (swim) to the French adverbial (a la nage). 27 
The disadvantage to this approach is that it requires a new transfer rule for every 
adverbial that might potentially participate in the traverser construction (e.g., ~ pied, en 
courant, en marchant, etc.). The LCS approach, on the other hand, resolves this type of 
divergence compositionally by relying on the full coverage constraint and the :CON- 
FLATED marker, analogous to the handling of example (26). The underlying LCS for 
the two sentences in (47) is the following: 
(49) \[Event GOLoc 
(\[Thing JOHN\], 
\[Path ACROSSLoc (\[Position ATLoc (\[Thing JOHN\], \[Location RIVER\])\[)\], 
\[Manner SWIMMINGLY\])\] 
The solution to this example relies on the assumption that the distinction between 
the English and French exists by virtue of the fact that the word swim includes the 
manner component swimmingly, whereas the word traverser does not. Because of this 
conflational distinction, the manner component is suppressed in English, but is overtly 
realized (as ~ la nage) in French. 2s The important point is that this entire concept is fully 
27 A related, but more general, strategy would be to handle such cases in bilingual lexical entries (see, for 
example, Beaven 1992a, 1992b; Whitelock 1992; Trujillo 1992). 
28 The use of the word traverser (i.e., cross) instead of nager (i.e., swim) is independently determined by the 
fact that the path component ACROSS is present in the conceptual representation (i.e., there would be 
no way to realize the path component in conjunction with the word nager). 
624 
Bonnie J. Dorr Machine Translation Divergences 
covered (in the sense of \[34\]) by both the source- and target-language sentences, even 
though these two sentences do not have the same structural representation. That is, 
as long as all conceptual components of (49) are somehow retrievable, the suppres- 
sion/realization of the individual components in the surface structure may be forced 
by the presence/absence of the :CONFLATED marker in the relevant lexical entries. 
5.4 Interacting Divergence Types 
We now turn to another important issue that has only recently received the attention 
it deserves, namely, that of handling interacting divergence types. In particular, there 
have been criticisms (see, for example, Lindop and Tsujii 1991) of systems that perform 
transfer on relatively shallow analyses (e.g., early METAL \[Alonso 1990, Thurmair 
1990\] and LTAG \[AbeillG Schabes, and Joshi 1990\]) owing to the fact that such systems 
are not likely to be able to handle divergence interactions (although they may be able to 
handle each divergence type in isolation). The solution adopted in the current approach 
does not appeal to a shallow analysis (i.e., it does not use a set of already-coded canned 
"frames" with predetermined argument structure). Rather, the syntactic structures are 
derived compositionally on the basis of two pieces of information: the structure of 
the CLCS (i.e., the language-independent predicate-argument information) and the 
lexical entries (i.e., the RLCSs and their associated language-dependent information). 
It would not be possible to handle interacting divergence types in an approach that 
maps directly from a set of hard-wired source-language frames to a set of hard-wired 
target-language frames. This is because an argument that occurs in a divergent phrasal 
construction might itself be a divergent phrasal construction. 
Consider the following example: 
(5O) Promotional and thematic divergence: 
S: Leer libros le suele gustar a Juan 
'Reading books (him) tends to please (to) John' 
E: John usually likes reading books 
This example exhibits a simultaneous occurrence of two types of divergences: the verb 
soler exhibits a promotional divergence with respect to its internal argument gustar a 
Juan, which itself exhibits a thematic divergence. The recursive nature of the GGT4 is 
crucial for handling such cases. 
The CLCS for (50) is the following: 
(51) \[State BEcirc 
(\[Thing JOHN\], 
\[Position ATcirc 
(\[Thing JOHN\], \[Event READ (\[Thing JOHN\], \[Thing BOOK\])\])\], 
\[Manner LIKINGLY\], 
\[Manner HABITUALLY\])\] 
Note that there are two modifiers, LIKINGLY and HABITUALLY. It is the job of the 
~£T4 function to determine the appropriate decomposition of the event on the basis 
of the language-specific requirements of the RLCSs involved in the mapping. In the 
current example, the LIKINGLY component is, in a sense, an "inherent" modifier, since 
it appears in the RLCS of both like and gustar. In contrast, the HABITUALLY modifier 
is an independent constituent that corresponds to independent RLCSs for usually and 
soler. 
625 
Computational Linguistics Volume 20, Number 4 
We will now formally analyze how this example is handled. The two relevant 
override mappings are specified in (15) and (11), repeated here for convenience: 
(52) Promotional override: 
1.' X' ~=~ Z 
4.' Q' ~ X 
(53) Thematic override: 
1.' W' ~Z 
4/Z' ¢¢ W 
In terms of the logical constituents that participate in the divergence mapping, X' 
corresponds to \[State BEcirc -..\], W' corresponds to \[Thing JOHN\[, Z' corresponds to 
\[Event READ ...\], and Q' corresponds to \[Manner HABITUALLY\]. 
Formally, the structure of the English sentence in (50) has the default syntactic 
representation 
(54) \[Y-MAX \[Y-MAX W \[X-MAX X Z\]l Q\] 
In contrast, the equivalent Spanish sentence has an entirely different syntactic repre- 
sentation: 
(55) \[Y-MAX \[Y-MAX Z \[X-MAX Q \[ ..- X W\]\]\]\] 
Note that this structure differs from the default representation in that the external 
argument is Z, not W, and that the syntactic head is Q, not X. Also, because the 
promotional mapping forces X into an internal argument position, additional structure 
is created so that X retains its status as a head. This head selects an internal argument 
that, in this case, is W rather than Z because of the interaction with the thematic 
divergence. 
Suppose we were to generate the Spanish sentence for this example. The RLCSs 
for the Spanish case conform to the following formal specifications: 
(56) gustar: \[T(X') X' (\[T(W') :INT W'\], \[T(Z') :EXT Z'\])\] 
soler: IT(Q,) :PROMOTE Q'\] 
When the ~£T4 is applied to the CLCS of (51), the promotional override (52) is im- 
mediately triggered by the :PROMOTE marker in the RLCS for soler; this invocation 
crucially precedes the invocation of the thematic override (53). 29 The promotional over- 
ride forces Q' (i.e., \[Manner HABITUALLY\[) to be realized as the syntactic head soler. 
This head takes an internal argument corresponding to X' (\[State BEcirc ...\]) that is 
29 This seems to indicate that there is some notion of prioritization during the resolution of interacting 
divergence types. In particular, the head swapping (promotional and demotional) divergences appear to 
take priority over the "argument swapping" cases (thematic). The decision to impose this prioritization 
is not entirely unprincipled. It would not be possible to apply these two overrides in the opposite 
order, since a head must be properly positioned (potentially via a promotional override) in order to 
identify the relative positions for its satellites (potentially via a thematic override). Although the formal 
ramifications of this ordering have not yet been established, it should be noted that the prioritization 
fits in naturally in the current framework, given that the syntactic realization process starts by realizing 
"outer" phrases, but then recursively realizes "inner" phrases before any attachments are made. 
626 
Bonnie J. Dorr Machine Translation Divergences 
realized as the verb gustar. Recall that the structural relation between the subordi- 
nated verb and its internal argument is not changed by the promotional mapping (see 
Footnote 13). However, the promotional operation does change the relation between 
the subordinated verb and its external argument by realizing the argument in an ex- 
ternal position relative to the main verb. Normally, this would mean that the CLCS 
constituent \[Thing JOHN\] would become an external argument of soler such as in the 
noninteracting case, John suele leer libros. In the current example, however, the attach- 
ment of the external argument is delayed until the recursive application of the ~£T4 
on \[State BEcirc ...\]. 
At this point, the :INT and :EXT markers trigger the thematic interchange, and 
the logical subject \[Thing JOHN\] is realized as the internal argument a Juan. The \[Event 
READ ...\] constituent is then taken to be the external argument of gustar, except 
that this constituent cannot be attached inside of the subordinate phrase owing to 
the promotional divergence. Instead, the current phrase is completed and the external 
argument is "passed up" to the higher phrase, which then attaches it in an external 
position relative to the verb soler. The final structure is then generated: 
(57) \[C-MAX 
\[I-MAX 
\[C-MAX leer libros\] 
\[V-MAX \[Vle suele\] 
IV-MAX \[V gustar\] \[P-MAX a Juan\]\]\]\]\] 
Thus, we have shown how the current framework provides a formal means for demon- 
strating how interacting divergence types are handled. 
6. Limitations and Conclusions 
The current approach has been implemented in a system called UNITRAN (Dorr 1990a, 
1990b, 1993b). 3° Many of the problems associated with the direct replacement and 
transfer approaches of previous systems have been eliminated in this new design. In 
particular, UNITRAN does not make use of analysis/synthesis rules that are meticu- 
lously tailored to each of the source and target languages, nor does it require detailed 
source-to-target transfer rules. On the other hand, because the system is designed (de- 
liberately) to operate on one sentence at a time, it has a number of inherent limitations. 
In particular, the lack of a theory of multisententiality makes high quality transla- 
tion difficult, since it is often the case that a single sentence in one language should 
be translated as two or more sentences in the other language; currently, UNITRAN 
does not allow for a mismatch between the number of source- and target-language 
sentences. 31 Since generation has not been the primary focus of the current research, 
this and other generation problems have not yet been addressed in this framework. 
Such problems include cohesion (Granville 1983), selecting propositional and rhetorical 
goals (McKeown 1985), selecting open-class items from "deep knowledge" (Goldman 
30 The name UNITRAN stands for UNiversal TRANslator, that is, the system serves as the basis for 
translation across a variety of languages, not just two languages or a family of languages. 
31 In general, systems should be designed so that sentences are realized differently in different languages, 
depending on the speaker's intended effect. In the words of a reviewer, "What's a long-winded and 
boring sentence to an American may be precisely a very fine formal sentence to a German." Even for 
languages as close as French and English, sentence boundaries differ about 10% of the time (see Brown 
et al. 1991). 
627 
Computational Linguistics Volume 20, Number 4 
1975; Jacobs 1985; Kittredge, Iordanskaja, and Polgu6re 1988; Nirenburg and Niren- 
burg 1988; Nirenburg et al. 1992; among others), ordering propositions for producing 
coherent text (Hovy 1988), resolving anaphora (Derr and McKeown 1984; Sondheimer, 
Cumming, and Albano 1990; Werner and Nirenburg 1988), and many others. In fact, 
not all of these issues (e.g., selecting propositional and rhetorical goals) are directly 
relevant to the task of machine translation, which already has the advantage (at least 
from the generation point of view) that the source-language sentence and, in the cur- 
rent model, the conceptual analysis underlying this sentence are available at the onset 
of the generation process. 
Instead, the current research focuses on demonstrating the utility of the LCS- 
based interlingua and the associated parameters for resolving translation divergences 
while maintaining the systematic relation between the interlingua and the syntax. The 
tasks involved in achieving this objective have been reduced to what might be con- 
sidered the standard "what" and "how" questions of generation: (1) lexical selection, 
i.e., the task of deciding what target-language words accurately reflect the meaning 
of the corresponding source-language words; and (2) syntactic realization, i.e., the task 
of determining how target-language words are mapped to their appropriate syntactic 
structures. In the context of the current model, the first task consists of matching the 
LCS-based interlingua (the CLCS) against the LCS-based entries (the RLCS) in the 
dictionary in order to select the appropriate word, and the second task consists of re- 
alizing the positions marked by * (and other parametric markers) into the appropriate 
syntactic structure. 
The question of lexical choice is one that deserves further discussion. The details 
of the matching process that achieves lexical selection of the target-language RLCSs 
(e.g., the selection of the RLCSs for like \[or gustar\] from the underlying CLCS shown 
in Definition 1) have not been presented here, but see Dorr (1993a) for a discussion 
with examples. Roughly, lexical selection in UNITRAN is a "reverse" unification-like 
process that matches the CLCS to the RLCS templates in the lexicon and chooses 
the associated lexical words accordingly. One of the important problems that must 
be considered with respect to lexical selection is that of overgeneration. 32 In particu- 
lar, one might ask how the matcher knows whether it should try to choose phrases 
that restate the source language more succinctly in the target language, or whether it 
should be allowed to restate the source-language phrases more verbosely in the target 
language. This comes up in cases such as the stab example given earlier. It turns out 
that the word stab can be translated into Spanish as the succinct form apu~alar, or as 
the more verbose form dar pu~aladas. Similarly, in the reverse direction, the transla- 
tion of dar pu~aladas is the more succinct form stab; however, one could conceive of 
a more verbose translation (e.g., inflict knife wounds or even give knife wounds). Cur- 
rently, there is no preference assignment during lexical selection (as in Nirenburg and 
Nirenburg 1988; Wilks 1973); instead, the system requires an exact match of the CLCS 
to the target-language RLCS (or some combination of RLCSs). If there is more than 
one way of matching the CLCS, multiple forms will be generated (although we have 
discussed only one target-language form, dar pu~aladas, for the stab example). A first- 
pass approach to resolving such cases of overgeneration (based on aspectual features) 
is discussed in Dorr (1992a) and in more detail in Dorr (1993b). In addition, a model 
32 The complexity of the lexical selection process is a well-studied problem. See, for example, the work by 
Reiter (1990), which shows that the selection of the optimal set of adjectives for a noun phrase under 
some very strong conditions is NP-complete. Presumably, the general lexical selection problem is considerably harder. 
628 
Bonnie J. Dorr Machine Translation Divergences 
of generation based on theories of tense by Allen (1983, 1984), Hornstein (1990), and 
Reichenbach (1947) is discussed in Dorr and Gaasterland (1992) and Dorr (1993b). 
The difficulty of the lexical choice problem will become more important as the 
system grows beyond its current prototypical state. Research is currently underway 
to extend the system by developing automatic lexical acquisition procedures that 
make use of a small set of conceptual structures, as a starting point, and then ac- 
quire syntactic and semantic information on the basis of these initial representations 
plus machine-readable definitions from the Longman's Dictionary of Contemporary English 
(Proctor 1978). (LDOCE is useful because it includes collocations and sense frequency, 
thus making it possible to determine the argument structures for different words.) 
This investigation will benefit from the work of several researchers in the field of au- 
tomatic lexicon construction, most notably, Brent (1993), Boguraev and Briscoe (1989), 
Boguraev and Pustejovsky (1990), Briscoe and Copestake (1990), Byrd et al. (1987), Far- 
well, Guthrie, and Wilks. (1992), Montemagni and Vanderwende (1992), Pustejovsky 
(1987), Pustejovsky and Bergler (1987), and Pustejovsky, Bergier, and Anick (1993), 
among others. In particular, it has been argued convincingly by Farwell, Guthrie, and 
Wilks (1992) that resources such as the LDOCE are useful for constructing dictionary 
representations for languages other than English, thus paving the way for scaling up 
interlingual machine translations so that they have broader coverage. Once this ex- 
tension is complete, we intend to scale up the UNITRAN system and test the LCS 
approach to lexical choice on a broader set of phenomena using a larger lexicon. 
The current framework provides a systematic classification of machine transla- 
tion divergences. We have shown how this classification can be formally defined and 
systematically resolved through the use of general mapping relations and a small set 
of cross-linguistic parameters. Because the parameters are used to factor out "trans- 
fer" information, the current approach obviates the need for transfer rules. We have 
provided evidence that supports the view that the lexical-semantic divergence classifi- 
cation proposed in the current framework covers all lexical-semantic divergences that 
arise during translation (i.e., divergences based on properties associated with lexical 
entries that are not based on purely syntactic information, idiomatic usage, aspectual 
knowledge, discourse knowledge, domain knowledge, or world knowledge). Since the 
characterization of the range of potential divergences is manageably small, the task of 
accommodating divergences is immensely simplified. We have also demonstrated the 
usefulness of a full coverage requirement as a tool that allows one to judge whether a 
particular target-language sentence fully covers the concept that underlies the corre- 
sponding source-language sentence. Finally, we have shown, formally, that the current 
model accommodates interacting divergence types. 
Acknowledgments 
This paper describes research done at the 
University of Maryland Institute for 
Advanced Computer Studies. Support for 
this research has been provided in part by 
the National Science Foundation under a 
Young Investigator Award IRI-9357731 and 
grant IRI-9120788, by the DARPA Basic 
Research Program under grant 
N00014-92-J-1929, by the Army Research 
Office under contract DAAL03-91-C-0034 
through Batelle Corporation, and by the 
Army Research Institute under contract 
MDA-903-92-R-0035 through 
Microelectronics and Design, Inc. Useful 
guidance and commentary during the 
research and preparation of this document 
were provided by Bob Berwick, Bruce 
Dawson, Ken Hale, Clare Voss, and Amy 
Weinberg. The author would also like to 
thank three anonymous reviewers for their 
helpful comments during the preparation of 
this document. 
References 
AbeillG Anne; Schabes, Yves; and Joshi, 
Aravind K. (1990). "Using lexicalized tags 
for machine translation." In Proceedings, 
13th International Conference on 
629 
Computational Linguistics Volume 20, Number 4 
Computational Linguistics, Helsinki, 
Finland, 1-6. 
Abney, S. (1989). "A computational model 
of human parsing." Journal of 
Psychol in guistic Research 18:129-144. 
Allen, James E (1983). "Maintaining 
knowledge about temporal intervals." 
Communications of the ACM 26(11):832-843. 
Allen, James E (1984). "Towards a general 
theory of action and time." Artificial 
Intelligence 23(2):123-160. 
Alonso, Juan Alberto. (1990). "Transfer 
interstructure: designing an 'interlingua' 
for transfer-based MT systems." In 
Proceedings, Third International Conference 
on Theoretical and Methodological Issues in 
Machine Translation of Natural Languages, 
Linguistics Research Center, The 
University of Texas, Austin, Texas, 
189-201. 
Arnold, Doug, and des Tombe, Louis. 
(1987). "Basic theory and methodology in 
Eurotra." In Machine Translation: Theoretical 
and Methodological Issues, edited by Sergei 
Nirenburg, 114-135. Cambridge: 
Cambridge University Press. 
Barnett, Jim; Mani, Inderjeet; Martin, Paul; 
and Rich, Elaine (1991a). "Reversible 
machine translation: What to do when the 
languages don't line up." In Proceedings, 
Workshop on Reversible Grammars in Natural 
Language Processing, ACL-91, University of 
California, Berkeley, California, 61-70. 
Barnett, Jim; Mani, Inderjeet; Rich, Elaine; 
Aone, Chinatsu; Knight, Kevin; and 
Martinez, Juan C. (1991b). "Capturing 
language-specific semantic distinctions in 
interlingua-based MT." In Proceedings, 
Machine Translation Summit, Washington, 
DC, 25-32. 
Beaven, John (1992a). "Lexicalist 
unification-based machine translation." 
Doctoral dissertation, University of 
Edinburgh, Edinburgh, UK. 
Beaven, John (1992b). "Shake and bake 
machine translation." In Proceedings, 14th 
International Conference on Computational 
Linguistics, Nantes, France, 603-609. 
Boguraev, Branimir, and Briscoe, Ted (1989). 
Computational Lexicography for Natural 
Language Processing. London: Longman. 
Boguraev, Branimir, and Pustejovsky, James 
(1990). "Lexical ambiguity and the role of 
knowledge representation in lexical 
design." In Proceedings, 13th International 
Conference on Computational Linguistics, 
Helsinki, Finland, 36-41. 
Boitet, Christian (1987). "Research and 
development on MT and related 
techniques at Grenoble University 
(GETA)." In Machine Translation: The State 
of the Art, edited by Margaret King, 
133-153. Edinburgh University Press, 
Edinburgh. 
Brent, Michael (1993). "From grammar to 
lexicon: Unsupervised learning of lexical 
syntax." Computational Linguistics 
19(2):243-262. 
Briscoe, E. J., and Copestake, A. A. (1990). 
"Enjoy the paper: Lexical semantics via 
lexicology." In Proceedings, 13th 
International Conference on Computational 
Linguistics, Helsinki, Finland, 42-47. 
Byrd, Roy J., Calzolari, Nicoletta; 
Chodorow, Martin S.; Klavans, Judith L.; 
Neff, Mary S.; and Rizk, Omneya A. 
(1987). "Tools and methods for 
computational linguistics." Computational 
Linguistics 13(3-4):219-240. 
Carbonell, Jaime G., and Tomita, Masaru 
(1987). "Knowledge-based machine 
translation, the CMU approach." In 
Machine Translation: Theoretical and 
Methodological Issues, edited by Sergei 
Nirenburg, 68-89. Cambridge: Cambridge 
University Press. 
Chomsky, Noam A. (1981). Lectures on 
Government and Binding. Dordrecht, 
Holland: Foris Publications. 
Chomsky, Noam A. (1982). Some Concepts 
and Consequences of the Theory of Government 
and Binding. Cambridge: MIT Press. 
Chomsky, Noam A. (1986a). Barriers. 
Cambridge: MIT Press. 
Chomsky, Noam A. (1986b). Knowledge of 
Language: Its Nature, Origin and Use. 
Cambridge: MIT Press. 
Copeland, C.; Durand, J.; Krauwer, S.; and 
Maegaard, B. (1991). "The Eurotra 
linguistic specifications." In Studies in 
Machine Translation and Natural Language 
Processing, Volume 1, edited by Erwin 
Valentini. Brussels: Commission of the 
European Communities. 
Derr, M., and McKeown, K. (1984). "Using 
focus to generate complex and simple 
sentences." In Proceedings, Tenth 
International Conference on Computational 
Linguistics, Stanford, California, 319-326. 
Dorr, Bonnie J. (1990a). "A cross-linguistic 
approach to machine translation." In 
Proceedings, Third International Conference 
on Theoretical and Methodological Issues in 
Machine Translation of Natural Languages, 
Linguistics Research Center, The 
University of Texas, Austin, Texas, 13-32. 
Dorr, Bonnie J. (1990b). "Solving thematic 
divergences in machine translation." In 
Proceedings, 28th Annual Conference of the 
Association for Computational Linguistics, 
University of Pittsburgh, Pittsburgh, 
Pennsylvania, 127-134. 
630 
Bonnie J. Dorr Machine Translation Divergences 
Dorr, Bonnie J. (1992a). "Lexical semantics 
for interlingual machine translation." 
Machine Translation 7(3). 
Dorr, Bonnie J. (1992b). "A parameterized 
approach to integrating aspect with 
lexical-semantics for machine translation." 
In Proceedings, 30th Annual Conference of the 
Association of Computational Linguistics, 
University of Delaware, Newark 
Delaware, 257-264. 
Dorr, Bonnie J. (1993a). "lnterlingual 
machine translation: A parameterized 
approach." Artificial Intelligence 63(1,2). 
Dorr, Bonnie J. (1993b). Machine Translation: 
A View from the Lexicon. Cambridge: MIT 
Press. 
Dorr, Bonnie J., and Gaasterland, Terry 
(1992). "Reflecting time in generated text: 
Tense, aspect and temporal connecting 
words." Technical report UMIACS TR 
92-92, CS TR 2950, Department of 
Computer Science, University of 
Maryland, College Park, Maryland. 
Dorr, Bonnie J., and Voss, Clare R. (1993a). 
"Constraints on the space of MT 
divergences." In Building Lexicons for 
Machine Translation, Papers from the 1993 
Spring Symposium, Technical report 
SS-93-02, Stanford University, Stanford, 
California. 
Dorr, Bonnie J., and Voss, Clare R. (1993b). 
Machine translation of spatial 
expressions: Defining the relation 
between an interlingua and a knowledge 
representation system." In Proceedings, 
Twelfth Conference of the American 
Association for Artificial Intelligence, 
Washington, DC. 
Farwell, David; Guthrie, Louise; and Wilks, 
Yorick (1992). "The automatic creation of 
lexical entries for a multilingual MT 
system." In Proceedings, 14th International 
Conference on Computational Linguistics, 
Nantes, France, 532-538. 
Goldman, Neil M. (1975). "Conceptual 
memory and inference." In Conceptual 
Information P~ocessing, edited by Roger C. 
Schank, 289-371. Amsterdam, Holland: 
Elsevier Science Publishers. 
Granville, Robert (1983). "Cohesion in 
computer text generation: Lexical 
substitution." Technical Report, LCS 
Technical Report 310, Massachusetts 
Institute of Technology, Cambridge, 
Massachusetts. 
Hale, Kenneth, and Keyser, S. Jay (1986a). 
"Some transitivity alternations in 
English." Technical report, Lexicon Project 
Working Paper 7, Center for Cognitive 
Science, Massachusetts Institute of 
Technology, Cambridge, Massachusetts. 
Hale, Kenneth, and Keyser, S. Jay (1986b). 
"A view from the middle." Technical 
report, Lexicon Project Working Paper 10, 
Center for Cognitive Science, 
Massachusetts Institute of Technology, 
Cambridge, Massachusetts. 
Hale, Kenneth, and Keyser, S. Jay (1989). 
"On some syntactic rules in the lexicon." 
Technical Report, Center for Cognitive 
Science, Massachusetts Institute of 
Technology, Cambridge, Massachusetts. 
Hale, Kenneth, and Laughren, Mary (1983). 
"Warlpiri lexicon project: Warlpiri 
dictionary entries." Technical report, 
Warlpiri Lexicon Project, Massachusetts 
Institute of Technology, Cambridge, 
Massachusetts. 
Hornstein, Norbert (1990). As Time Goes By. 
Cambridge: MIT Press. 
Hovy, Eduard (1988). Generating Natural 
Language Under Pragmatic Constraints. New 
Jersey: Lawrence Erlbaum Associates. 
Isabelle, Pierre (1987). "Machine translation 
at the TAUM group." In Machine 
Translation: The State of the Art, edited by 
Margaret King, 247-277. Edinburgh: 
Edinburgh University Press. 
Jackendoff, Ray S. (1983). Semantics and 
Cognition. Cambridge: MIT Press. 
Jackendoff, Ray S. (1990). Semantic 
Structures. Cambridge: MIT Press. 
Jacobs, Paul S. (1985). "PHRED: A generator 
for natural language interfaces." 
Computational Linguistics 11(4):219-242. 
Johnson, Rod; King, Maghi; and des Tombe, 
Louis (1985). "Eurotra: A multilingual 
system under development." 
Computational Linguistics 11(2,3):155-169. 
Kameyama, Megumi; Ochitani, Ryo; Peters, 
Stanley; and Sirai, Hidetoshi (1991). 
"Resolving translation mismatches with 
information flow." In Proceedings, 29th 
Annual Meeting of the Association for 
Computational Linguistics, University of 
California, Berkeley, California, 193-200. 
Kaplan, Ronald M., and Bresnan, Joan 
(1982). "Lexical-functional grammar: A 
formal system for grammatical 
representation." In The Mental 
Representation of Grammatical Relations, 
edited by Joan Bresnan, 173-28l. 
Cambridge: MIT Press. 
Kaplan, Ronald M.; Netter, Klaus; 
Wedekind, Jurgen; and Zaenen, Annie 
(1989). "Translation by structural 
correspondences." In Proceedings, Fourth 
Conference of the European Chapter of the 
Association for Computational Linguistics, 
Manchester, UK, 272-281. 
Kay, Martin (1984). "Functional unification 
grammar: A formalism for machine 
631 
Computational Linguistics Volume 20, Number 4 
translation." In Proceedings, lOth 
International Conference on Computational 
Linguistics, Stanford University, Stanford, 
California, 75-78. 
Kinoshita, Satoshi; Phillips, John; and Tsujii, 
Jun-ichi (1992). Interaction between 
structural changes in machine 
translation." In Proceedings, 14th 
International Conference on Computational 
Linguistics, Nantes, France, 679-685. 
Kittredge, Richard I.; Iordanskaja, Lidija; 
and Polgu6re, Alain (1988). "Multi-lingual 
text generation and the meaning-text 
theory." In Proceedings, Conference on 
Theoretical and Methodological Issues in 
Machine Translation of Natural Languages, 
Carnegie Mellon University, Pittsburgh, 
Pennsylvania. 
Levin, Beth, and Rappaport, Malka (1986). 
"The formation of adjectival passives." 
Linguistic Inquiry 17:623-662. 
Lindop, Jeremy, and Tsujii, Jun-ichi (1991). 
"Complex transfer in MT: A survey of 
examples." Technical report, CCL/UMIST 
Report 91/5, Center for Computational 
Linguistics, UMIST, Manchester, UK. 
Marantz, Alec (1992). "The 
way-constructions and the semantics of 
direct arguments in English: A reply to 
Jackendoff." Technical report, Department 
of Linguistics and Philosophy, 
Massachusetts Institute of Technology, 
Cambridge, Massachusetts. 
McCord, Michael C. (1989). "Design of 
LMT: A prolog-based machine translation 
system." Computational Linguistics 
15(1):33-52. 
McKeown, Kathleen (1985). Text generation: 
Using discourse strategies and focus 
constraints to generate natural language text. 
Cambridge: Cambridge University Press. 
Melby, A. K. (1986). "Lexical transfer: 
Missing element in linguistic theories." In 
Proceedings, 11th International Conference on 
Computational Linguistics, Bonn, Germany. 
Meyer, Ingrid; Onyshkevych, Boyan; and 
Carlson, Lynn (1990). "Lexicographic 
principles and design for 
knowledge-based machine translation." 
Technical report, CMU CMT Technical 
Report 90-118, Carnegie Mellon 
University, Pittsburgh, Pennsylvania. 
Montemagni, Simonetta; and Vanderwende, 
Lucy (1992). "Structural patterns vs. 
string patterns for extracting semantic 
information from dictionaries." In 
Proceedings, 14th International Conference on 
Computational Linguistics, Nantes, France, 
546-552. 
Nirenburg, Sergei; Carbonell, Jaime; Tomita, 
Masaru; and Goodman, Kenneth (1992). 
Machine Translation: A Knowledge-Based 
Approach. San Mateo, California: Morgan 
Kaufmann. 
Nirenburg, Sergei, and Goodman, Kenneth 
(1990). "Treatment of meaning in MT 
systems." In Proceedings, Third International 
Conference on Theoretical and Methodological 
Issues in Machine Translation of Natural 
Languages, Linguistics Research Center, 
The University of Texas, Austin, Texas, 
171-187. 
Nirenburg, Sergei, and Levin, Lori (1989). 
"Knowledge representation support." 
Machine Translation 4(1):25-52. 
Nirenburg, Sergei, and Nirenburg, Irene 
(1988). "A framework for lexical selection 
in natural language generation." In 
Proceedings, 12th International Conference on 
Computational Linguistics, Budapest, 
Hungary, 471-475. 
Nirenburg, Sergei; Raskin, Victor; and 
Tucker, Allen B. (1987). "The structure of 
interlingua in translator." In Machine 
Translation: Theoretical and Methodological 
Issues, edited by Sergei Nirenburg, 90-113. 
Cambridge: Cambridge University Press. 
Perlmutter, David M. (1983). Studies in 
Relational Grammar 1. Chicago: The 
University of Chicago Press. 
Proctor, P. (1978). Longman Dictionary of 
Contemporary English. London: Longman. 
Pustejovsky, James (1987). "On the 
acquisition of lexical entries: The 
perceptual origin of thematic relations." 
In Proceedings, 25th Annual Conference of the 
Association for Computational Linguistics, 
Stanford University, Stanford, California, 
172-178. 
Pustejovsky, James, and Bergler, Sabine 
(1987). "The acquisition of conceptual 
structure for the lexicon." In Proceedings, 
Sixth Conference of the American Association 
of Artificial Intelligence, Seattle, 
Washington, 566-570. 
Pustejovsky, James; Bergler, Sabine; and 
Anick, Peter (1993). "Lexical semantic 
techniques for corpus analysis." 
Computational Linguistics 19(2):331-358. 
Reichenbach, H. (1947). Elements of Symbolic 
Logic. London: Macmillan. 
Reiter, Ehud (1990). "Generating 
appropriate natural language object 
descriptions." Doctoral dissertation, 
Harvard University, Cambridge, MA. 
Sadler, Louisa, and Thompson, Henry S. 
(1991). "Structural non-correspondence in 
translation." In Proceedings, Fifth Conference 
of the European Chapter of the Association for 
Computational Linguistics, Berlin, Germany, 
293-298. 
Shieber, Stuart M.; van Noord, Gertjan; 
632 
Bonnie J. Dorr Machine Translation Divergences 
Moore, Robert C.; and Pereira, Fernando 
C. N. (1989). "A semantic-head-driven 
generation algorithm for unification-based 
formalisms." In Proceedings, 27th Annual 
Conference of the Association for 
Computational Linguistics, University of 
British Columbia, Vancouver, British 
Columbia, Canada, 7-17. 
Shieber, Stuart M.; van Noord, Gertjan; 
Moore, Robert C.; and Pereira, Fernando 
C. N. (1990). "Semantic-head-driven 
generation." Computational Linguistics 
16(1). 
Sondheimer, N.; Cumming, S.; and Albano, 
R. (1990). "How to realize a concept: 
Lexical selection and the conceptual 
network in text generation." Machine 
Translation 5(1):57-78. 
Thurmair, Gregor (1990). "Complex lexical 
transfer in metal." In Proceedings, Third 
International Conference on Theoretical and 
Methodological Issues in Machine Translation 
of Natural Languages, Linguistics Research 
Center, The University of Texas, Austin, 
Texas, 91-107. 
Trujillo, Arturo (1992). "Locations in the 
machine translation of prepositional 
phrases." In Proceedings, Fourth 
International Conference on Theoretical and 
Methodological Issues in Machine Translation 
of Natural Languages, Montreal, Canada, 
13-20. 
Tsujii, Jun-ichi, and Fujita, Kimikazu (1991). 
"Lexical transfer based on bilingual signs: 
Towards interaction during transfer." In 
Proceedings, European Chapter of the 
Association for Computational Linguistics, 
Berlin, Germany, 275-280. 
Vauquois, Bernard, and Boitet, Christian 
(1985). "Automated translation at 
Grenoble University." Computational 
Linguistics 11(1):28-36. 
Werner, P., and Nirenburg, S. (1988). "A 
specification language that supports the 
realization of intersentential anaphora." 
In Proceedings, AAAI Workshop on Natural 
Language Generation, Saint Paul, 
Minnesota. 
Whitelock, Pete (1992). "Shake-and-bake 
translation." In Proceedings, 14th 
International Conference on Computational 
Linguistics, Nantes, France, 784-791. 
Wilks, Yorick (1973). "An artificial 
intelligence approach to machine 
translation." In Computer Models of Thought 
and Language, edited by Roger C. Schank 
and K. M. Colby, 114-151. San Francisco: 
Freeman. 
Zubizarreta, Maria Luisa (1982). "On the 
relationship of the lexicon to syntax." 
Doctoral dissertation, Department of 
Linguistics and Philosophy, 
Massachusetts Institute of Technology, 
Cambridge, Massachusetts. 
Zubizarreta, Maria Luisa (1987). Levels of 
Representation in the Lexicon and in the 
Syntax. Dordrecht, Holland: Foris 
Publications. 
633 

