Discourse Processing in MT: 
Problems in Pronominal Translation 
Hajime Wada : wadah@yvax.byu.edu 
Department of Linguistics, The University of Texas at Austin 
and 
ECS, 455 North University, Suite 202, Provo, UT 84601, U.S.A. 
Keywords: Machine Translation, Anaphora, Discourse Representation Theory, Focus 
1. Intrcnluction 
Translation of anaphoric expressions has been 
problematic in most of the MT systems (Key 1986). 
One ot' the main reasons for the difficulties lies in the 
lack of discourse information representation in the MT 
systems. In this paper, we report an implementation of 
the Discourse Representation Theory in an LFG-based 
English-to-Japanese MT program, and discuss 
problems in translating anaphoric expressions in this 
system. 
2. Problems in Translating Anaplmra 
Problems in translation of anaphoric expressions 
can be seen on three different but interactive levels of 
linguistic information: lexical, syntactic and 
pragmatic. 
The main problem on the lexical level is due to a 
difference in the language specific parameters in the 
pronominal system such as F features (person, gender, 
number, etc.). Surface forms of pronominals depend 
on the F features of their antecedents, so that the 
translation of a particular pronominal form cannot be 
determined sorely by the pronominal itself. For 
example, 'ship' in English is feminine, but its 
translation, 'hune', is neutral in Japanese. Thus, a 
preform 'she' for 'ship' should not be translated as 
'kanojo' (3rd, sing, fem), but as 'sore' (3rd, sing, neut). 
Problems on the syntactic level are mainly due to a 
difference in the distribution of anaphoric expressions. 
Mapping relations between English pronominals and 
their Japanese counterparts are shown below: 
E~!~ish ~oanese 
Overt Pronoun~-.,~ Overt Pronoun 
Reflexives ~ zibun 
X-zishin 
Empty Pronoun --~ Control-Zero 
\ Topic-Zero 
The most crucial case is English overt pronouns. As 
Kuroda noted (1965:107), overt pronouns are not 
repeatable in Japanese, while they are in English. (1) 
a. He does Iris work when he feels like doing it. 
b. *Kare-wa kare-ga sitai toki-ni kare-no 
he 'FOP he NOM like to do when his GEN 
sigoto-wo suru. 
work ACC do 
Instead of repeating overt pronouns, zero pronominals 
and 'zibun' should be used though there are some 
variations in their use. 
c. Kare-wa 0 sitai toki-ni ~ sigoto-wo suru. 
d. Kare-wa zibun-ga sitai toki-ni 0 sigoto-wo sum. 
e. Kare-wa 0 sitai toki-ni zil)un-no sigoto-wo suru. 
Resumptive pronouns are also problematic. 
Possessive relative pronoun 'whose' in English must be 
translated as a resumptive pronoun whose antecedent 
is the head NP of a relative clause in Japanese. 
The last set of problems is on the level o1' 
pragmatics. It has been known that in many languages 
certain types of pronominal expressions represent a 
focus in a discourse. For example, in English, overt 
pronouns play this role while zero pronouns (topic- 
zero) do the job in Japanese. In a certain discourse 
context, a discourse focus should be translated as a 
reflexive 'zibun' instead of a zero pronoun. 
Some cases may be induced to lexical properties of 
predicates and pronominals such as functional and 
psuedo-functional control cases and reflexivity. 
Nonetheless, most of the above problems are not 
easily separable; they should better be solved in a 
unified model. 
3. Treatment of Pronominal Translation 
In this paper, we will focus on the problems of 
translating English overt non-reflexive pronominals 
into Japanese. These l)ronominals must be mapped 
into 1) overt l)ronominals, 2) 'zibun', 3) pro, or 4) 
discourse pro in Japanese. 
3.1. Role of Discourse 
To explain some exceptions on locality conditions 
and ambiguous cases in anaphora resolution, recent 
studies have focused on the role of discourse (Grosz 
1981; Sidner 1983; Joshi & Weinstein 1981; 
Kameyama 1985; Roberts 1986) and world 
knowledge. Discourse factors affecting anaphora 
resolution are discourse structure and discourse 
focus/center. This report concerns only the latter; in 
particular, the selection of focus and its maintenance. 
Following Grosz, we assume that a focus is the 
most salient NP in a sentence, and every sentence has 
one and only one focus 1. Often a focus is marked by 
the use of pronominal, stress, thematic role and 
particular syntactic constructions such as clefting, 
question, passivization, etc. However, there seems no 
clear generalization on the organization of the focus 
marking factors. We simply assume the most marked 
NP with respect to die number of found factors (asher 
& Wada 1989; Alshawi 1987). 
We have adopted the Discourse Representation 
Structure (Kamp 1981) to represent a current 
discourse structure which will be appended to an f- 
structure representing a sentence. DRS+f-structure 
gives a ground for examining logical conditions based 
1 We concern a local focus only, and multiple foci cases are 
excluded from our discussion. 
1 73 
on the accessibility condition, and locality conditions 
such as disjoint reference on each anaphoric 
expression (cf. Roberts 1986). 
Now consider 4) discourse pro. Its characteristic is 
that it may not be bound in the current sentence, but 
must be bound by the focus in the previous discourse. 
In other words, when a pronominal cannot find its 
antecedent in the current DRS, we assume that it must 
be bound by the previous discourse focus unless the 
focus has been shifted, and in such a case we translate 
it as ~, a lexical zero pronominal. 
3.2. Proform Selection by FU 
On the other hand, when an antecedent is found in 
the DRS, the anaphoric link will be noted not only in 
DRS, but also in relevant f-structures via features such 
as Ant(ecedent) and Desc(endent) with a common 
index. However, we still need to determine which 
form should be used: pro, 'zibun', or overt 
pronominals. The selection should be syntactically 
made on the Japanese side since no additional 
information is available for the selection in the English 
f-structures. 
Let us discuss the syntactic mechanism for the 
proforrn selection. Suppose that a proform be lexically 
presented in each noun in the lexicon such as 
follows2: I 
~ENDER msc GENDER msc 
)ROFORI'I 'he' PROFORid 'he' 
When the two shown above are anaphorically 
linked in a sentence, LFG offers a syntactic solution, 
namely, the Functional Uncertainty (Kaplan & 
Maxwell 1988) 3. The link is represented by a 
functional-application equation (FU) such as follows: 
(2) < ^/GF* \GF PREFORM > = < ^ PREFORM > 
In the above FU, the second GF after the right slash 
represents an f-structure that dominates a possible 
antecedent, and the first GF* after the left slash, f- 
commanding path to the antecedent from the current f- 
structure 4. That is, the FU imposes a constraint such 
that the right-hand side of the equation in (2) 
represents an f-structure containing a pronominal, and 
if there is an f-structure within the path represented by 
the left-hand side of the equation, then it must be the 
case that their PROFORMs are unifiable 5. 
For the generation of surface PROFORMs, we 
assume that a pronominal element possesses a 
lexically unspecified PREFORM, and it will be 
specified by its antecedent. In other words, given that 
a pronominal and its antecedent are coindexed, an FU 
such as follows should be satisfied. 
(3) < ^ / GF* \ GF* Ant > =c < ^ Desc > 
2plural PROFORMs should be assigned to NPs syntactically. 
3The stone mechanism has been used in other cases such as 
the noun classifier selection for numeral expressions and 
negative scope domain setting by 'wa' in Japanese (Wada 
forthcoming). 
4 This has been done by pointing the antecedent from the 
current l'-structure. 
5 Note that the equation is far more powerful than this 
description. 
This equation can be solved only when the indices are 
identical. Once the equation (3) is solved, PREFORM 
will be unified between the antecedent and the 
pronominal 6 
Selection on pro and 'zibun' can be considered as 
solving particularly restricted FU equations. For 
example, 'zibun' will be selected when the following 
FU equation is satisfied 7. 
(4) <^/GF(ADJUNCT)GF* ',SUBJ Ant> =c <A Desc > 
This equation says that if there exists a SUBJ that f- 
commands GF(s) that may contain an adjunct which 
includes an f-structure in which the current pro 
resides, then it must be the case that the Ant value of 
that SUBJ and Desc value of the pro are identical. 
Each unspecified 'pro' will carry a set of equations, 
each of which selects a particular PREFORM 
disjunctively. 
(5) 
{ \[<A/GF(ADJUNCT)GF* KSUBJ Ant> =c <A DOSC > 
< ^ PREFORM > = 'zibun' \] 
\[ <A/GF* ADJUNCTx SUBJ Ant> =c <A Desc > 
< ^ PREFORM > = 0 \] 
\[<^/GF* \GF* Ant> =c <^ Desc > 
< A PREFORM > = <!A PREFORM > \] } 
4. Implementation of the Treatment 
The anaphora resolution mechanism in our 
program consists of three functional units: the DRS 
construction (DRSCONS), the salient element storing 
(SLSTOR), and the antecedent searching (ANSR). 
4.1. DRSCONS 
This module constructs DRSs compositionally, 
following Montagovian tradition (Wada 8,: Asher 
1986, Reyle 1988). Thus, we assume that a semantic 
representation, SMS, is appended to each F-structure. 
SMS is a pair < SemType, DRS >, where SemType is 
a semantic type. 
4. 2. SLSTOR 
SLSTOR's main function is to store the most 
salient, focused element in the current discourse for 
the next sentence processing (Alshawi 1987; Asher & 
Wada 1989). In order to find the most salient NP, 
SLSTOR sets three kinds of filters among others such 
as grammatical function, use of pronominal, and 
syntactic construction (See Asher & Wada 1989), and 
check all of the NPs appeared in the current sentence 
with respect to the three filters. 
After the most salient element has been selected, 
SLSTOR checks whether the current focus has been 
shifted or retained by comparing the current SLSTOR 
value and the newly selected one. If their gender, 
number, person, etc. are unifiable, the discourse focus 
is considered to retain and SLSTOR keeps the 
previous NP in the storage. If they are not unifiable, 
the focus has been shifted, and SLSTOR stores the 
newly selected element as a current focus. 
4.3. ANSR 
6Resumptive pronouns are treated in the same way shown 
above, except for the co-indexation. Since an antecedent for a 
resumptive pronoun is always the head NP of the relative 
clause and therefore, the co-indexation is not necessary. 
7This is the condition that Kuroda (1965) proposed, and still 
has been a basically correct generalization. 
74 2 
This module consists of three functions. The first 
function is SEARCH. SEARCH searches antecedents 
by testing the accessibility on the DRS and morpho- 
syntactic constraints such as gender, number, and 
binding features (sub, nuclear, etc) on the current f- 
structure. If a search in the DRS fails, SEARCH 
further searches in SLSTOR, and check a focused 
element in SLSTOR. In effect, we distinguish 
pronominals into three classes according SEARCH's 
result; 1) an antecedent found in the current DRS, 2) 
an antecedent not found in the current DRS, but 
controlled by a discourse focus, and 3) an antecedent 
not found in either in the current DRS or in the 
previous focus. 
The second function is LINK. In the case in 1) 
above, \]LINK will set a unique anaphoric index in both 
F-structm'es of the antecedent and its pronominal and 
assigns an unspecified 'pro' to the pronominal. 
The last function is SETPRO, which assigns 0 to 
the pronominal in the case of 2), and assigns a default 
word-for-word translation in the case of 3). 
5. Examples 
Let us take some examples. First, consider the 
following donkey sentence. 
(6-a) 
Every man who loves a ship treats her tenderly. 
'Her' will be linked with 'a ship' by SEARCH due to 
the accessibility of 'a ship' to 'her' on the DRS. Then, 
the same anaphoric index is assigned to the transfer f- 
structure of 'a ship' and of 'her' by LINK. 'Her' is 
translated as an unspecified 'pro' as shown below. 
(6-b) 
Hune we aisuru otoko wa mina sore we itukushimu. 
ship ACC love man TOP every it ACC treat-tender 
In this result, an overt pronominal, 'sore', is selected 
for 'pro' since the f-structure for 'hune' and 'pro' share 
an anphoric index, and the last clause in the FU 
equation in (5) is satisfied. 
Suppose that the next sentence appeared discourse- 
initially. 
(7-a) John believes that she is a genius. 
Since there is neither a sentence-internal antecedent 
nor a previous focus, 'she' gets a default translation, 
'kanojo'. However, if the following sentence precedes 
the sentence in (7-a), a lexical ~ will be selected as 
the target for 'she' because 'Mary' is the focus of (7-b) 
since it is a SUBJ and appears in a passive 
construction. 
(7-b) Mary was aw,'uded as a distinguished artist. 
Now, consider the next example. 
(8-a) John works when he wants to do his work. 
In the above example, 'his' should be linked to 'he' and 
'he' to 'John'. Each anaphoric index has a different 
value, but both of the pronominals will be translated as 
unspecified 'pro's. However, the equation (5) gives 
them different PROFORMs, 'his' 'zibun' and 'he' ~. 
This is indeed a desired result as shown below. 
(8-b) 
John-wa 0 zibun-no sigoto-wo sitai-toki-ni hataraku. 
John TOP POSS work ACC want-to-do when work 
6. Remaining Problems 
A note on defects and shortcomings is in order. 
It is no doubt that world knowledge plays an 
important role in anaphora resolution. We assume that 
in the near future, a world knowledge inferencing 
module should become available, and that it will be 
incorporated into our system. 
We assumed that an input text as a whole consists a 
discourse segment; however, this is due to the lack of 
a treatment in discourse segmentation in our program. 
One related problem is on tim distinction of the global 
focus and local loci. Since our program considers 
only one segment at a time, there is no way to handle 
complex foci. 
Another ,area which we need to handle but has been 
neglecting is number related expressions such as 
plural antecedents and split antecedents. 
Also, the salient element selection mechanism in 
our program is still based on a crude stipulation. We 
need to conduct further research on this topic. 
Finally, it is desirable to map a source DRS to the 
target DRS and handle language-specific discourse- 
based expressions. These will be our next task. 

References 

Alshawi, H. (1987) Memory and Context for Language 
Interpretation, Cambridge University Press. 

Asher, Nicholas (1986) "Belief in Discourse 
Representation Theory", in Journal of 
Philosophical Logic, 5, pp. 127-189. 

Asher, N. & H. Wada (1989). "Computational Account 
of Anaphora", ms, CGS, Univ. of Texas. 

Gmsz, B. (1978). "Discourse Knowledge", in 
Understanding Spoken Language, ed. by Donakt 
Walker, pp.228-345,. 

Grosz, B. (1981). "Focusing in Dialogue", in TINLAP- 
2 :Theoretical Issues in Natural Language 
Processing. pp.96-103. ACM anti ACL. N.Y. 

Joshi, A. & S. Weinstein (1981) "Control of 
Inference:Role of Some Aspects of Discourse 
Structure-Centering", in proceedings of 7th IJCAI, 
pp.385-387 

Kmneyama, Megumi. (1985). Zero Anaphora." Ttw 
Case of Japanese. Ph.D. dissertation.Stanford 
University. 

K,'unp, H. (1981) "A Theory of Truth and Semantic 
Representation", in Groenendijk and Janssen (eds), 
Formal Methods in the Study of Language, 
Gronningen:Amsterdam, pp.l-41. 

Kmnp, H. (1983) "Situation in Discourse", ms, CSLI. 

Kaphm, R. & J. Maxwell (1988). "An Algorithm for 
Functional Uncertainty", in the proceedings of 
COL1NG-88, pp.297-302. 

Kay, M (1986) "Machine Translation will not work", in 
the proceedings of ACL, p.268. 

Kuroda, S-Y. (1965) Generative Grammatical Studies in 
the Japanese Language, Ph.D. dissertation, MIT. 

Roberts, C. (1986) Modal Subordination, Anaphora, 
and Distributivity, Ph.D.dissertation, UMASS. 

Reyle, W. (1988)"Computational Sem~mtics for LFG", in 
Natural Language Parsing and Linguistics Theories, eds 
by U. Reyle and C Rohrer, pp. 448-47& 

Sidner, Candice (1983). "Focusing in the 
Comprehension of Definite Anaphora", in 
Computational Models of Discourse, eds by R. 
Berwick anti M. Brady, pp.267-330. 

Wada, Hajime (forthcoming). "Applications of Functional 
Uncertainty in ECS System", ms. 

Wada, Hajime & N. Asher (1986). "BUILDRS: An 
hnplementation of DR Theory and LFG", in the 
proceedings of COLING-86, pp.540-546. 
