Fully Lexicalized Head-Driven Syntactic Generation 
Tilman Becker 
German Research Center for Artificial Intelligence (DFKI GmbH) 
Stuhlsatzenhausweg 3, 66123 Saarbriicken, Germany 
becker@dfki, de 
Abstract 
We describe a new approach to syntactic generation with Head-Driven Phrase Structure 
Grammars (HPSG) that uses an extensive off-line preprocessing step. Direct generation algo- 
• rithms apply the phra~se-structure rules (schemata) of the grammar on:line which is an com- 
putationally expensive step. Instead, we collect off-line for every lexical type of the HPSG 
grammar all minimally complete projections (called elementary trees) that can be derived with 
the schemata. This process is known as 'compiling HPSG to TAG' and derives a Lexicalized 
Tree-Adjoining Grammar (LTAG). The representation as an LTAG is 'fully lexicalized' in the 
sense that all grammatical information is directly encoded with the lexical item (as a set of 
elementary trees) and the combination operations are reduced from schema applications to the 
TAG primitives of adjunction and substitution. Given this LTAG, the generation task has a 
very different search space that Can be traversed very efficiently, avoiding the costly on-line 
applications of HPSG unification. The entire generation task from a semantic representation 
to a surface string is split into two tasks, a microplanner and a syntactic realizer. This paper 
discusses the syntactic generator and the preprocessing steps as implemented in the Verbmobil 
system. 
1 Generation in a Speech-to-Speech System 
The syntactic generation algorithm and the preprocessing steps presented in this paper are inte- 
grated into the Verbmobil system (see \[Wahlster 1993, Bub; Wahlster, and Waibel 1997\]). It is a 
system for speech-~to-speech dialog translation. The input for the generation module VM-GECO 1 
is generated by a semantic-based transfer component (see \[Dorna and Emele 1996\]). The inter- 
face language chosen comprises the encoding of target language-specific semantic information in a 
combination of Underspecified Discourse Representation Theory and Minimal Recursion Semantics 
(see \[Bos et al. 1996\] and \[Copestake, Flickinger, and Sag 1997\]). 
The internal architecture of the generation module is modularized: it is separated into two phases, a 
microplanner and a syntactic generator. Throughout the system, we emphasize declarativity, which 
is also a necessary precondition for a comprehensive off-line preprocessing of external knowledge 
bases-in particular the preprocessing of the underlying Head-Driven Phrase Structure Grammar 
(HPSG, see \[Pollard and Sag 1994\]) which has been developed at CSLI, reflecting the latest devel- 
opments in the linguistic theory and with a fairly wide coverage and also covering phenomena of 
spoken language. 
1VerbMobil GEneration COmponents 
•208 
,! 
! 
! 
! 
i 
2 Microplanning and Syntactic Generation 
Starting from the semantic representation, the microplanning component generates an annotated 
dependency structure which is used by the syntactic generation component to realize a surface 
string. The microplanner also carries out word-choice. 
One goal of this modularization is a stepwise constraining of the search-space of alternative lin- 
guistic realizations, using different views in the different modules. In each step, only an abstraction 
of the multitude of information contained in an alternative needs to be considered. 
Another aspect of this architecture is the separation into a kernel system, i.e., the language in- 
dependent core algorithms (a constraint-solver for microplanning and the search and combination 
algorithms for syntactic generation described in section'5) and declarative knowledge bases, e.g., 
the language specific word-choice constraints in microplanning and the TAG grammars used in syn- 
tactic realization. This separation allows for an easy adaptation of the system to other languages 
and domains (see \[Becker et al. 1998\]). 
3 Declarativity in the Syntactic Generator 
All modules of the generator utilize external, declarative knowledge bases. For the syntactic gen- 
erator, extensive off-line preprocessing of the highly declarative HPSG grammar for English 2 is 
applied. The grammar has not even been written exclusively as a generation grammar 3. It is 
specialized, however, in that it covers phenomena of spoken language. The high level of abstraction 
which is achieved in the hierarchically organized grammar description (see \[Flickinger 1987\]) allows 
for easy maintenance as well as off-line preprocessing. 
The off-line preprocessing steps described in the next section keep the declarative nature of the 
grammar intact, Le. they retain explicitly the phrase structures and syntactic features as defined 
by the HPSG grammar. 
In general, declarative knowledge bases allow for an easier adaptation of the system to other 
domains and languages. This is a huge benefit in the current second phase of the Verbmobil project 
\[Becker et al. 1996\] where the generator is extended to cover German, English and Japanese as well 
as additional and extended domains with a considerably larger vocabulary. 
4 Off-Line Preprocessing: HPSG to TAG Compilation 
The subtasks in a direct syntactic generator based on an HPSG grammar will always include the 
application of schemata (the HPSG equivalent of phrase structure rules) such that all syntactic 
constraints introduced by a lexical item (especially its SUBCAT list)are fulfilled. This results 
in a constant repetition of, e.g., building up the projection of a verb in a declarative sentence. 
In preprocessing the HPSG grammar we aim at computing all possible partial phrase structures 
which can be derived from the information in a lexicon entry. Given such sets of possible syn- 
tactic realization together with a set of selected lexicon entries for an utterance and finally their 
dependencies, the task of a syntactic generator is simplified considerably. Instead of exploring all 
2The HPSG grammar is being developed at CSLI, Stanford University. Development is carried out on a grammar 
development platform which is based on TDL \[Krieger and Sch£fer 1994\]. 
3In fact, most of the testing during grammar development depends on the use of a parser. 
209 
possible, computationally expensive applications of HPSG schemata, it merely has to find suitable 
precomputed syntactic structures for each lexical item and combine them appropriately. 
For this preprocessing of the HPSG grammar, we adapted the 'HPSG to TAG compilation' process 
described in \[Kasper et al. 1995\]. The basis for the compilation is an identification of syntactically 
relevant selector features which express subcategorization requirements of a lexical item, e.g. the 
VALENCE features. In general, a phrase structure is complete when these selector features are 
empty. 
Starting from the feature structure for a lexical item, HPSG schemata are applied such that the 
current structure is unified with a daughter feature of the schema. The resulting structure is again 
subject to this process: This compilation process stops when certain termination criteria are met, 
e.g., when all selector features are empty. Thus, all projections from the lexical item are collected 
as a set of minimally complete phrase structures which can also be interpreted as elementary trees 
of a Tree-Adjoining Grammar (TAG). 
Instead of actually applying this compilation •process to all lexical items, certain abstractions over 
the lexical entries are specified in the HPSG grammar. In fact, the needs of the compilation process 
have led to a clear-cut separation of lexica! types and lexical entries as shown in Figure 1. A 
typical lexical entry is shown in Figure 2 and demonstrates that only three kinds of information are 
stored: the lexical type MV_NP_TRANS_LE 4, the semantic contribution (th e relation _SUIT_REL) 
and morphological information (the stem and potentiallyirregular forms): By expanding the lexical 
type, the full feature structure can be obtained. 
Lexicon Hierarchy - Phrase Structure 
\[ syntac,,c \] \[ sem nt, c \] Types Types 
\[,Morphological L Infomation J 
ILexiCa~x 200 types 
~__._\[ Lexical \] Instance 
approx. 2,900entries 
HPSG \] Principles 
Schemata \] 
approx. 25 schemata 
Figure 1: Organization of the HPSG grammar. 
Some of the trees which result from the preprocessing of the lexical type MV_NP_TRANS_LE 
are shown in Figure 3. The figure Shows only the phrase structure and an abstraction of the 4MV_NP_TRANS_LE 
is an abbreviation for "Main Verb, NP object, TRANSitive Lexical Entry" used in sentences like 
"Monday suits me." 
210 
I 
I 
,I 
! 
I 
suit_vl := mv_np_trans_le 
\[ STEM < "suit" >, 
SYNSEM.LOCAL.CONT.STEMLISZT < ! \[ PRED _suit_tel \] !>\]. 
Figure 2: Specification of a lexical instance for the verb "suit." 
node's categories. All nodes still represent the full HPSG •feature structures. E.g., the tree 
MV_NP_TRANS_LE.2 of Figure 3 represents an imperative clause. As a consequence PERSON 
has the value SECOND and CL-MODE is set to IMPERATIVE. Note that the compilation process 
stopped at this node since the selector features are empty. 
MV NP TRANS_LE.1 MV NP_TRANS_LE.2 MV NP TRANS_LE.3 MV NP TRANS_LE.4 
vP 
V NP$ I 
• MV NP TRANS_LE 
s s s 
VP NP.S.C P NP.S.COMP $ VP 
I I I MV NP TRANS_LE MV NP TRANS_LE MV NP TRANS_LE 
Figure 3: Some of the trees for transitive verbs. They are compiled from the corresponding lexical 
type MV_NP_TRANS_/E as defined in the HPSG grammar. Trees 3 and 4 differ only with respect 
to their feature structures which are not shown in this figure. 
From these trees, two kinds of knowledge bases are built. For the microplanner, the relation between 
the lexical and syntactic realization and the semantic representation (encoded in the SYNSEM 
LOCAl CONT feature) is extracted as a constraint. For the syntactic generator, the relevant 
syntactic information is extracted in the form of a Feature-Based Lexicalized TAG (FB-LTAG) 
grammar, see \[Joshi 1987, Vijay-Shanker and Joshi 1991, Schabes, Abeill4, and Joshi 1988\]. This 
includes the phrase structure and a selected part of the feature structure (mainly the SYNSEM 
LOCAL CAT and SYNSEM NON-LOCAL features). Figure 4 shows the bottom feature structure 
extracted from the root node of MV_NP_TRANSJE.2. Note that some of the feature paths are 
abbreviated, e.g. 5LCI stands for SYNSEM LOCAL CONT INDEX. The elementary TAG trees which 
are built from the compilation result have so-called restricted •feature structures which can be 
exploited for an efficient, specialized unification algorithm. 
The node names shown in the figures represent a disjunction of possible categories, e.g. NP.S.COMP 
in tree MV_NP_TRANS_LE.3 implies that the subject of a transitive verb may be a nominal or 
sentential phrase. 
211 
Bottom Dag at selected node: 
:ROOT: \[SLC: \[HEAD: \[FRD: (- .)\] \[MOOD: (SUBJUNCTIVE MODAISUB} Ih~ICATIVE)\] 
\[VOICE: (PASSIVE ACTIVE)\] \['mNSE: (FUnma PAST PR~Sln, rO\] 
\[worv~t Bs~ Ire'v: -I 
\[AUX'.-\] \[ROOT: .\] 
\[CL-MODE: IMPERATIVE\] \[RULE: 
IMPERATIVE_RULE\] \[SLCI: NIL\] 
\[SY~EM: It,tON-LOCAL: \[Qtm: -\] 
Figure 4: The bottom feature structure of the S node of tree MV.NP_TRANS_LE.2. 
Finally, the leaf nodes of the trees (except for the lexical item itself) are marked either as substi- 
tution nodes or as a foot node, thus creating an auxiliary tree. In a TAG derivation, substitution 
nodeS are replaced with trees bearing the correct category and a Unifiable feature structure at their 
root node. Auxiliary trees can be inserted into other trees by the adjunction operation. 
5 The Syntactic Generator VM-GIFT 
The task of the syntactic generator is the construction of a sentence (or phrase, given the often 
incomplete utterances in spoken dialogs) from the microplanning result which is then sent to a 
speech-synthesis component. It proceeds in three major steps which are also depicted in Fig. 5. 
• A tree selection phase determines the set of relevant TAG trees. A first tree retrieval step 
maps every object of the dependency tree into a set of applicable elementary TAG trees. The 
main tree selection phase uses information from the microplanner output to further refine the 
set of retrieved trees. 
• A combination phase finds a successful combination of trees to build a (derived) phrase 
structure tree. 
* An inflection phase uses the information in the feature structures of the leaves (i.e. the words) 
to apply appropriate morphological functions, including the use of irregular forms as provided 
by the HPSG lexiconand regular inflection function as supplied (as LISP code) by the HPSG 
grammar. 
An initial preprocessing phase computes: the necessary auxiliary verbs from the tense, aspect, 
and sentence mood information. It also rearranges the dependency tree accordingly (e.g. subject 
arguments are moved from the main verb to become dependents of the inflected auxiliary verb). 
The two core phases are the tree selection and the tree combination phase. The tree selection 
phase consists of two steps. First, a set of possible trees is retrieved and then appropriate trees are 
selected from this set. The retrieval is driven by the HPSG instance or word class that is supplied 
by the microplanner. It is mapped to a lexical type by a lexicon that is automatically compiled from 
the HPSG grammar. The lexical types are then mapped to a tree family, i.e., a set of elementary 
TAG trees representing all possible minimally complete phrase structures that can be build from 
the instance. The additional information in the dependency tree is then used to add further feature 
212 
-t .o 
tD e- 
e- 
a) 
Preprocessing I I Tree selection I I Tree combination { I (expand auxiliaries)~\] and sorting \[--~ (adjoining and substitution) ~-~ Inflecti°n \[ 
Irregular 
Figure 5: Steps of the syntactic generator. 
t:). 
O 
la0 
t- 
values to the trees. This additional information acts as a filter for selecting appropriate trees in 
two stages: 
• Some values are incompatible with values already present in the trees. These trees can 
therefore be filtered immediately from the set. E.g., a syntactic structure for an imperative 
clause is marked as such by a feature and can be discarded if a declarative sentence is to be 
generated. 
• Additional features can prevent the combination with other trees during the combination 
phase. This is the case, for example with agreement features. 
The combination phase explores the search space of all possible combinations of trees from the 
candidate sets for each lexical item (instance). An inefficient combination phase is a potential 
drawback of using the precomputed TAG trees. However, there is sufficient information available 
fl'om the microplanner result and from the trees such that a well:guided best-first search strategy 
can be employed in the current system. The difference in run-time can be as dramatic as 24 seconds 
(comprehensive breadth-first) versus 1.5 seconds (best-first). 
As part of the tree selection phase, based on the rich annotation of the input structure, the tree sets 
are sorted locally. Then a backtracking algorithm traverses the dependency tree in a bottom-up 
fashion s. At each node, and for each subtree in the dependency tree, a candidate for the phrase 
structures of the subtree is constructed. Then all possible adjunction or substitution sites are 
computed, possibly sorted (e.g. allowing for preferences in word order) and the best candidate for 
a combined phrase structure is returned. Since the combination of two partial phrase structures 
by adjunction or substitution might fail due to incompatible feature structures , a backtracking 
5The algorithm stores intermediate results with a memoization technique. 
213 
algorithm must be used. A partial phrase structure for a subtree of the dependency is finally checked 
for completeness. These tests include the unifiability of all top and bottom feature structures and 
the satisfaction of all other constraints (e.g. obligatory adjunctions or open substitution nodes) 
since no further adjunctions or substitutions will occur in this subtree. 
The necessity of a spoken dialog translation system to produce output robustly calls for some 
relaxations in these tests. E.g., 'obligatory' arguments may be missing in the utterance and the 
tests in the syntactic generator must accept a sentence with a missing obligatory object if no other 
complete phrase can be generated. 
Figure 6 shows an example of the input of from the microplanner after the preprocessing phase has 
inserted the entity LGV1 for the auxiliary will. 
( (ENTITY LGVI 
((CAT V) (HEAD.WILL_AUX_POS) (INTENTION WH-QUESTION) (FUNC AUX) 
(TENSE• FUTUP~) (MOOD INDICATIVE) (VOICE ACTIVE) (FORM OP, DINARY) 
(VPORM FIN))) 
(ENTITY LS-WORK_ACCEPTABLE 
. ((FORM OKI)INARY) (VFOKM BSE) (CAT V) (GOVE~-BY WH-SENTENCE) 
(OPTIONAL-AGENT NO) (HEAD (OR SUIT_VI SUIT_V2)) (REALIZED LOCAL) 
(KEG LGVl))) 
(ENTITY LI3-PRON 
((REALIZED LOCAL) (CAT PPRON) (PERS 3) (NUM SG) (GENDER NTR) 
(TYPE NORMAL) (GOVERNED-BY V) (IS-COMPLEMENT T) (FORM CONTINUOUS) 
(KEG LGVI) (FUNC AGENT))) 
(ENTITY LI0-PRON 
((REALIZED LOCAL) (CAT PPRON) (PERS 2A) (NUM SG) (GENDER FEM) (TYPE NORMAL) 
(GOVERNED-BY (0R V PREP SENTENCE)) (FORM CONTINUOUS) (KEG L5-WORK_ACCEPTABLE) 
(FUNC • PATIENT) ) ) 
(ENTITY L6-TEMP_LOC 
((CA T ADV) (REAL WH_QUEST) (SORT TIME) (POINTED'BY TEMP_LOC) 
(GOVERNED-BY (0R V N ADV SENTENCE)) (PRED TIME) (HEAD WHEN1) 
• (REALIZED L0CAL) (WH-FOCUS T) (KEG L5-WORE_ACCEpTABLE) (FUNC TEMP-SPEC))) 
(ENTITY LI5-TEMP_LOC 
((CAT ADV) (BEAD THEN_ADV) (REALIZED GRouP-TIME-DEMONSTRATiVE) 
(REAL (0R ADV WH_QUEST YOFC)) (SORT (SUBSORT TIME)) (POINTED-BY TEMP_LOC) 
(GOVERNED-BY (OR V N ADV SENTENCE)) (BEG LS-WORK_ACCEPTABLE) (FUNC TEMP-SPEC)))) 
• . ..~ : 
Figure 6: Example of the input from microplanning after preprocessing for auxiliaries 
In the tree retrieval phase for L5-WORK_ACCEPTABLE, first the HEAD information is used to deter- 
mine the lexical types of the possible realizations SUIT_V1 and SUIT_V2, namely MV_NP_TRANS_LE and 
MV_EXPL_PREP_TRANSIE respectively. These types are then mapped to their respective sets of ele- 
mentary trees, a total of 25 trees. In the tree selection phase (as described above), this number 
is reduced to six. For example, the tree MV_NP_TRANS_L£.2 in Figure 3 has a feature CL-MOD£ 
with the value IMPERATIVE Now, the microplanner output for the root entity LGV1 contains the 
information (INTENTION WH-QUESTION) The NTENTION information is unified with all appropriate 
Ck-MOD£ features, which in this case fails. Thereforethe tree MV_NP_TRANS_k£.2 can be dis- 
carded in the tree selection phas e . 
214 
The combination phase uses the best-first bottom-up algorithm described above to determine one 
suitable tree for every entity and also a target node in the tree that is selected for the governing 
entity. For the above example, the selected trees and their combination nodes are shown in Figure 
7 6 , 
A0 " " % % ,¢ "~ % 
, V VP/ADV VP VP 
I ~''~ 1% 
I" % " l. % 
ADV Y NP $ .. " " NP Y NP J, ,' .NP VP ADV 
I I '-",- I I ''1,-, I 
when will it suit you then 
L6-TEMP_LOC LGVI LI3-PRON L5-SUIT LI0-PRON LI5-TEMP_LOC 
Figure 7: The trees finally selected for the entities Of the example sentence. The dashed lines 
connect to suitable substitution or adjunction nodes. They correspond to the dependency tree. 
The inflection function finally uses attribute values like verb-form, number and person from the 
final tree to derive the correct inflections. Information about the sentence mode WH-QUESTION 
can be used to annotate the resulting string for the speech-synthesis module. 
6 Conclusion and Comparison 
We have shown how preprocessing an HPSG grammar can be used to avoid the costly on-line ap- 
plication (unification) of HPSG schemata in a modularized generation system with a microplanner 
and a separate syntactic generator. The compilation of an HPSG grammar to TAG grammar allows 
the use of an efficient syntactic generator without sacrificing the declarative nature of the HPSG 
grammar. 
It is important to compare the generation strategy presented here with Semantic-head-driven gen- 
eration \[Shieber et al. 1990, van Noord 1990\] which is a direct generation algorithm froni logical 
form encodings. It improves previous algorithms in efficiency and in imposing less restrictions on 
the type of grammar. It is also applicable to HPSG and proceeds by applying the HPSG schemata 
in a bottom-up fashion, driven from the lexical heads of the schemata. 
To a large ex.tend, the TAG-based generation algorithm presented here goes through the same steps 
as semantic-head-driven generation. However, most of •these steps will have been made during the 
off-line preprocessing and are encoded in the elementary trees of the TAG grammar thns resulting 
6Note that the node labels shown in Figures 7 are only a concession to readability. The TAG requirement that 
in an auxiliary tree the foot node must have the same category label as the root node is formally fulfilled in our 
implementation. 
215 
in an important gain in efficiency. Note though, that the generation task in the algorithm presented 
here is shared between the micr0planner and the syntactic generator,-so a formal comparison must 
include both components. 
Work on generation with TAG generally assumes that there is a one, to~-one mapping between the in- 
formation in the generator input and the choice of elementary tree \[Mcdonald and Pustejovsky 1985, 
Yang, McCoy, and Vijay-Shanker 1991, D0ran and Stone 1997\]. In general, this will not be the 
case. In particular, in our system the input is not always sufficiently analyzed and the preprocess- 
ing froman HPSG grammar potentially •creates more than one elementary tree that fits the input 
parameters. 
One possible approach are choice nets-see \[Yang, McCoy, and Vijay-Shanker 1991\] who interpret 
systemic grammar in this way. Our approach has some similarity, though we have provided a more 
general algorithm that does not require the specification of grammar specific choice nets but rather 
executes tree Selection and combination from more declarative knowledge bases. Tree selection is 
implemented mainly by unification (adding feature values from the input specification to the trees 
where unifiable) and the best-first search algorithm is a general framework for handling sets of 
possible elementary trees, including backtracking steps when non-local tests (e.g. unification in the 
resulting derived tree) fail. This approach is also a precondition in our system since we have no 
direct access to the TAG grammar as it is automatically preprocessed from an HPSG grammar. 
VM-GECO is fully implemented (in Common Lisp) and integrated into the speech-to-speech 
translation system Verbmobil for Enghsh and German. For example, the underlying English HPSG 
grammar has almost 3000 iexical entries with over 200 lexical types. The resulting lexicalized TAG 
consists of about 2800 trees. The average overall generation time per sentence (up to length 24) is 
0.7 cpu •seconds on a SUN ULTRA-1 machine, 68% of the runtime are used for tile microplanning 
while tile remaining 32% of the runtime are used for syntactic generation. 
7 Current Work 
In general, the task of finding appropriate elementary trees for the chosen words and consequently 
a consistent phrase structure tree can exhibit constraints between an), two elementary trees in 
the utterance (as expressed through feature equations). •However, most of these constraints exist 
between elementary trees that are combined directly with each other (adjoined or substituted). To 
exploit thiSi we are currently experimenting with various well established binary constraint-solving 
algorithms to preselect elementary trees that are pairwise consistent w.r.t, feature equations. 

References 
\[Becket et al. 1996\] Becket, T. W. Finkler, A. K!lger, and W. Wahlster. 1996. Vorhabensbeschreibung zur 
Sprachgenerierung innerhalb des Teilprojektes 5 (Sprachgenerierung und -synthese) in Verbmobil, Phase 
2. Document, German Research Center for Artificial Intelligence (DFKI GmbH), Saarbriicken, Germany, 
AUgust. 
\[Becke r et al. 1998\] Becker, Tilman, Wolfgang Finkler, Anne Kilger, and Peter Poller. 1998. An efficient 
kernel for multilingual generation in speech-to-speech dialogue translation. In Proceedings of COLING- 
ACL 98, Montreal, Canada. 
\[Bos et al. 1996\] Bos, J., B. Gamb~ick, C. Lieske, Y. Mori, M. Pinkal, and K. Worm. 1996. Composi- 
tional semantics in verbmobil. Technical report, University of the Saarland, Computational Linguistics, 
Saarbrficken, July. Verbmobil Report 135. 
\[Bub, Wahlster, and Waibel 1997\] Bub, Th. W. Wahlster, and A. Waibel. 1997. Verbmobil: The combina- 
.tion of deep and shallow processing for spontaneous speech translation. In Proceedings of ICASSP '97. 
(forthcoming). 
\[Copestake, Flickinger, and Sag 1997\] Copestake, Ann, Dan Flickinger, and Ivan A. Sag. 1997. Minimal 
recursion semantics: An introduction, available at ftp://csli-ftp, stanford, edu/linguistics/sag/- 
mrs.ps, gz. ' 
\[Doran and Stone 1997\] Doran, Christy and Matthew Stone. 1997. Sentence planning as description using 
tree adjoining grammar. In ACL-EACL, Madrid, Spain, July. 
\[Dorna and Emele 1996\] Dorna, M. and M. Emele. 1996. Semantic-based transfer. In Proceedings of the 
16th International Conference on Computational Linguistics (COLING '96). 
\[Flickinger 1987\] Flickinger, Daniel P. 1987. Lexical Rules in the Hierarchical Lexicon. Ph.D. thesis, Stanford 
University. 
\[Joshi 1987\] Joshi, Aravind K. 1987. An introduction tO Tree Adjoining Grammars. In A. Manaster-Ramer, 
• editor, Mathematics of Language. John Benjamins, Amsterdam. 
\[Kasper et al. 1995\] Kasper, R., B. Kiefer, K. Netter, and K. Vijay-Shanker. 1995. Compilation of HPSG to 
TAG. In Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics, pages 
92-99, Cambridge, Mass. 
\[Krieger and Sch~fer 1994\] Krieger, Hans-Ulrich and Ulrich Sch~ifer. 1994. 779£--a type description language 
for constraint-based grammars. In Proceedings of the 15th International Conference on Computational 
Linguistics, COLING-9g, pages 893-899. 
\[Mcdonald and Pustejovsky 1985\] Mcdonald, David D. and James D. Pust'ejovsky. 1985. Tags as a gram- 
matical formalism for generation. In Proc. of the 23 th ACL, Chicago, IL. 
\[Pollard and Sag 1994\] Pollard, Carl and Ivan A. Sag. 1994. Head-Driven Phrase Structure Grammar. 
Studies in Contemporary Linguistics. University of Chicago Press, Chicago. 
\[Schabes, Abeill4, and Joshi 1988\] Schabes, Y., A. Abeill4, and A.K. Joshi. 1988. Parsing strategies with 
'lexicalized' grammars: Application to Tree Adjoining Grammars. In Proc. 12th bzternational Conference 
on Computational Linguistics (COLING-88}, pages 578-583, Budapest, August. 
\[Shieber et al. 1990\] Shieber, Stuart, Gertjan Van Noord, Fernando Pereira, and Robert Moore. 1990. 
Semantic-head-driven generation. Computational Linguistics Vol. 16 No. 1, 16(1):30-43. 
\[van Noord 1990\] van Noord, Gertjan. 1990. An overview of head-driven bottom-up generation. In Robert 
Dale, Chris Mellish, and Michael Zock, editors, Current Research in Natural Language Generation. Aca- 
demic Press, New York, pages 141-165. 
\[Vijay-Shanker and Joshi 1991\] Vijay-Shanker, K. and Aravind K. Joshi. 1991. Unification Based Tree 
Adjoining Grammars. In J. Wedekind, editor, Unification-based Grammars. MIT Press, Cambridge, Mas- 
sachusetts. 
\[Wahlster 1993.\] Wahlster, W. 1993. Verbmobil: Translation of face-to-face dialoges. In MT Summit IV, 
Kobe, Japan. 
\[Yang, McCoy, and Vijay-Shanker 1991\] Yang, Gijoo, Kathleen F. McCoy, and K. Vijay-Shanker. 1991. 
From functional specification to syntactic structures: Systemic grammar and tree adjoining grammar. 
Computational Intelligence, 7(4):207-219, November. 
