Achieving Bidirectionality 
Paul S. Jacobs 
Artificial Intelligence Program 
GE Research and Development Cenl;er 
Schenectady, NY 12301 USA 
Abstract 
The topic of BIDIltECT1ONAL1TY~ using common knowledge 
in language processing for both anMysis and generation, is 
of both practical and theoretical concern. Theoretically, it 
is important to determine what knowledge structures can 
be applied to both. Practically, it is important that a cmn- 
petent natural language system be ~ble to generate out- 
puts tilat are relevant to the inputs it understands, with- 
out exce~:sive redundancy. Tiffs problem revolves around 
the ~bility to relate linguistic structures declaratively to 
their mc~ming. 
1 Introduction 
BIDIRECTIONALITY, or the ability to use a cormnon knowledge 
base for both language anMysis mid generation, is a desirable 
feature of a reM language proccssing system. A natural lan- 
guage "front end" must not only perform syntactic anMysis, 
but must derive a suitable representation of a meaning or in- 
tention from a linguistic input. A natural language generator 
performs the: inverse task of producing a linguistic utterance 
fi'om a nmm~ing or intention. A bidirectional system performs 
both tusks using as much shaxed knowledge as possible. 
Two praeticM concerns motivate this work: (1) A system 
that uses shax'ed knowledge for analysis and generation will 
produce output ix* the subset of laa.lgu~ge that it understands, 
thus avoiding inconsistences between the input and output, and 
(2) Using shared knowledge avoids the inefficiency of having 
distinct encodings of the same linguistic information. 
The first concern, having a naturM language interface 
"speak" the same language it understands, is more than a con- 
veniencc. Responses in a dialogue often use a word or phrase 
that has been mentioned by another speaker. This cannot be 
done effectively unless the word or phrase is common to both 
the input and output language. A computer user will expect 
the system to understand a phrase or construct that the sys- 
tem has itsel\[ used; this aggravates the consequences of incon- 
sistencies between input and output language. Moreover, if 
an interface is to be traalsportable across domains, a distinct 
subset of language will be applicable to each domain. The bidi- 
rectional knowledge base allows both the input and output to 
be constrained simultaneously. 
The second concern, efficiency of knowledge representa- 
tion, becomes more compelling as the lexical and semantic ca- 
pabilities of natural language systems increase. While there 
is ample motivatioz, for having a common grammar fox" anal- 
ysis and ge:oeration, the need for a common lexicon is even 
stronger. H~ving two lexicons is counterintuitive; what makes 
practical sense is to have a single lexicon indexed differently 
for gene.ration fl'om analysis. Now that many systems have 
more mid more knowledge built into their lexicons, the effects 
of redundancy become more &'astie. When more information 
is required of the lexicon, however, the difficulties in developing 
a shared lexicon are more pronounced. 
The principal concern in designing a natural hmguage sys- 
tem that performs both analysis and generation, thereibre, is 
a bidirectional lexicon. The main issue to be eonsidered here 
is wt,at information nmst be included in this lexicon and how 
bidirectional lexiea./knowledge should be structured. 
2 Issues Regarding Bidlrectionality 
There h~s been very little reseaxch in language generation rela- 
tive to language understanding and syntactic analysis. A negli- 
gible amount of research has addressed the t)roblem of bidirec- 
tionality. Some work has touched on shared knowledge of lexi- 
cal semantics \[aacobs, 1985, Steinacker and Buchberger, 19831 
and on grammatical frameworks suitable for bidirectional sys- 
tems \[Kay, 1984\]. At the recent TINLAP (Theoretical Issues 
in Natural Language Processing) conference \[Wilks, 1987\], po- 
sition papers brought out a number of points concerning bidi- 
rectionality that had not previously appeared in the literature. 
The positions largely embraced the need for knowledge shared 
between a.nalysis and generation while laying out the practicsl 
reasons why bidirectional systems are not prevMent. 
A good summary of' issues in bidireetionality is found in 
\[Mann, 1.987\]. Erich aspect of the generation process can be 
related to some part of language analysis that seems to draw 
fl'om common knowledge. However, the processes themselves 
as well as the problems involved in building actuM language 
processing systems differ, to such an extent that scientists do 
not find the time to attend to the common issues. Another 
point is that both fields, especially generation, largely ignore 
the problem of lexical semantics \[Maxcus, 1987\], a problem that 
might help to bring the tasks closer together. 
It is a mistake to treat analysis and generation as como 
pletely independent tasks. Given that the goal of much of natu- 
ral language research is to program computers to communicate 
in the way people do, the ideal natural language program must 
use natural language us both a "front end" and a "back end". 
Knowledge that has tfistorieally been used more in generation, 
pertaining to text structure, coherence, and constraints on lex-. 
ical choice, influences the analysis task. Knowledge primarily 
applicable to analysis, such as vocabulary and grammatical 
coverage, and information applied to ambiguity mtd vagueness~ 
can be applied to generation as well. The problem of linguis- 
tic knowledge base design is thus fundamentally different for a 
bidirectional system. 
3 The Bidirectional Lexicon 
Several characteristics are essential to a lexicon that can be 
used effectively in both analysis and generation: 
267 
1. Principally, the lexicon and knowledge base of the system 
must be declarative; all the material must take the form 
of data structures rather than rules or program code. 
2. The semantic component of the lexicon; i. e. the rep- 
resentation of word meanings and word senses, must be 
sufficient to guide lexical choice in generation and to re- 
solve vague or ambiguous words in analysis. 
3. Lexical collocations, phrasal lexemes, and grammatical 
constructions must be represented. This compound lexical 
knowledge is necessary in generation because the selection 
of a particular word influences the selection of other words 
in a phrase, even when the phrase is internally grammati- 
cal. The knowledge is important in analysis in so far as it 
can aid in handling multiple word senses. 
Most systems satisfy the declarative requirement above, 
although the degree to which knowledge is proceduralized 
varies greatly from one model to another. The second and third 
requirements, the richness of lexical semantics and the need for 
compound knowledge , are more often overlooked. In genera- 
tion, a lexical entry that lists a word stem and a corresponding 
set of linguistic and semantic features is not enough; what is 
needed is a relationship between the lexlcal item and a knowl- 
edge representation structure \[Jacobs, 1986\] and a means of 
selecting the lexical item from among the other possible words 
\[Mathiessen, 1981\]. A word choice is not made independently 
from other choices; lexical choices have a direct influence on 
other lexical choices \[Jacobs, 1985\]. 
Lexical knowledge used primarily for generation can im- 
pact the way language analysis is performed, and vice versa. 
The following simple examples help to illustrate how complex 
lexical knowledge required for generation can also affect under- 
standing: 
• (la) Hit the "return" key. 
• (lb) Hit the "x" key. 
• (2a) Type "return". 
• (2b) Type "x". 
• (3a)Hit "return". 
• (3b) Hit "x". 
A generation system needs a variety of lexical knowledge 
in order to produce utterances such as (3a), which is natural for 
most native speakers. In addition to knowledge about the word 
sense of "hit", the system must know what keys are suitable 
for "hitting", as well as that "hit" is used to describe striking 
a single key. This detailed lexical knowledge should also avoid 
using (2a) in place of (3a), since one cannot use "type" for 
a key that does not produce a character or text. Now, given 
that this knowledge is required for the appropriate generation 
of the utterances above, it makes sense that it should be used 
in determining the difference in meaning between (2a) and (3a) 
(the former means "Hit the sequence of keys r-e-t-u-r-n). In 
designing a system strictly for analysis, one would tend to dis- 
tingtiish (2a) from (3a) by assnming "hit" to have a different 
meaning fl'om "type", and thus produce two incorrect but rela- 
tively subtle effects: First, the meanings of (2b) and (3b) would 
also be different, and second, (3b) would be equally acceptable 
to (3a). 
Because a generation system must have enough informa- 
tion in the lexicon to make appropriate lexical choices, it must 
have lexical knowledge that relates the specific word senses 
above to the linguistic context in which they are used. A lin- 
guistic analyzer can then use this knowledge to make more 
accurate interpretations of the same words. This is a typical 
way in which lexical choice and word sense determination are 
related. 
4 FLUSH 
An example of a lexicon designed with the three characteristics 
described in the previous section is FLUSH (Flexible Lexicon 
Using Structured Hierarchical knowledge) \[Besemer and Ja- 
cobs, 1987\]. FLUSH combines a hierarchical phrasal lexicon 
\[Wilensky and Arens, 1980, Jacobs, 1985, Dyer and Zernik, 
1986\] with declarative relations between language and mean- 
ing \[Jacobs and Rau, 1985\]. For example, figure 1 shows part 
of the lexical knowledge about the preposition "to", used in 
a prepositional phrase modifying either a verb or noun. The 
lexical relation to-pmod represents this linguistic category, and 
constrains how it can be used in a surface structure, based on 
its membership in the more general rood-tel (modifying rela- 
tion) category. 
Figure 2 shows how the to-pmod relation is associated 
with a generalized transfer event (either a physical transfer 
or a transfer of possession), with the object of the preposition 
describing the destination of the transfer. The link marked 
"REF" in figure 2 represents this sort of association between 
a linguistic and a conceptual structure. More specific trans- 
fers, as well as metaphorical "VIEWs" of transfers, are also 
explicitly represented in this diagram. Knowledge about senses 
of "sell", "tell", and "send?, as well as constructs using such 
verbs, is thus represented in a neutral fashion. 
Conceptual Structures Linguistic Structures 
~-t r a~"~ ~.. ~"l~ 
transfer-event .r ~ ~'~ ,-, ~w!ver~-iodirl , ¢ 
z~ I re¢ipient~ '~ -~ 
nerch-tran~,.~ selling 
Figure 2: Relating linguistic mad conceptual structures 
Compound lexical knowledge , often involving flgm'ative 
expressions, is also represented declaratively in FLUSH. Fig- 
Ure 3 shows how such knowledge is encoded: It-give-hug, the 
lexical category for "giving a hug" and other variations on 
the same expression, belongs to a general category, linguis~ 
tic/conceptual, which accounts for its linguistic flexibility such 
as its potential use in the passive voice. A "REF" association 
links Ic-give-hug to the hugging concept, indicating declara- 
tively that these expressions describe a hugging action rather 
2'68 
I whole-verb 
___~D 
base-va 
\[ mod-va 
\[ compound-lexeme j I prep-phrase \] 
D ~ npm -- ' rood "mod-'~l mqdynpm _J 
prep-root 
Figure 1: The modifying-relation compound-lexeme hierardiy. 
than a literal sense of "give". 
-- DI ' m "~ '"lexeme I 
I 
~1 "1 e~le-fiive-xxx \] 
\[ /o I 
Figure 3: The linguistic/conceptual relation Icr-give-hug. 
Th$.,~e examples, while only touching upon the lexical rep- 
resentatio:a of FLUSH, shows some of the characteristics of 
a birectional lexicon. The hierarchy of linguistic structures al- 
lows access to these structures for both analysis and generation. 
Declarative links between linguistic and conceptual entities al- 
low specific knowledge about linguistic expression to be used 
in hoth processes. The current task is to encode enough infor- 
mation in this form so that analysis and generation alike can 
be robustly performed. 
5 Conclusion 
Using certMn knowledge for both analysis and generation is de- 
sirable in ;~ natural language system, for both theoretical and 
practical reasons. This bidirectionality aids efficiency as well as 
insuring compatibility between analysis and generation compo- 
nents. A lexicon designed for bidirectionality differs distinctly 
fi:om one designed for either generation or analysis alone, and 
often develops aspects of each process that might otherwise be 
overlooked. 

References 

\[Besemer and Jacobs, 1987\] David Besemer and Paul S. Ja- 
cobs. FLUSH: a flexible lexicon design. In Proceedings 
of the 25th .Meeting of the Association for Computational 
Linguistics, Palo Alto, California, 1987. 

\]Dyer and Zernik, 1986\] Michael G. Dyer and Uri Zernik. En- 
coding and acquiring meanings for figurative phrases. In 
Proceediugs of the 2~\[th Annual Meeting of the Association 
for Computational Linguistics, New York, 1986. 

\[Jacobs, 1985\] Paul S. Jacobs. PHRED: t~ generator for natu- 
ral language interfaces. Computational Linguistics, 11(4), 
1985. 

\[Jacobs, 1986\] Paul S. Jacobs. Knowledge structures for nat- 
ural language genen~tion. In Proceedings of the Eleventh 
International Conference on Computational Lingui.~tic G 
Bonn, Germany, 1986. 

\[Jacobs and Rau, 1985\] Paul S. Jaeobs and Lisa F. Rau. Ace: 
associating language with meaning. In Tim O'Shca, ed- 
itor, Advances in Artificial Intelligence, pagcs 295 304, 
North Holland, Amsterdam, 1985. 

\[Kay, 1984\] M. Kay. Functional Unification Grammar: a for- 
malism for machine translation. In Proceedings of the 
Tenth International Conference on Computational Lin- 
guistics, Palo Alto, California, 1984. 

\[Mann, 1987\] W. C. Mann. What is special about natural lan- 
guage generation research? In TINLAP-3: Theoretical 
Issues in Natural Language Processing-3, pages 206- 210, 
Computing Research Laboratory, New Mexico State Uni- 
versity, Las Cruces, New Mexico, 1987. 

\[Marcus, 1987\] M. Marcus. Generation systems should choose 
their words. In TINLAP-3: Theoretical Issues in Nat- 
ural Language Processing-3, pages 211--214, Computing 
Research Laboratory, New Mexico State University, Las 
Cruces, New Mexico, 1987. 

\[Mathiessen, 1981\] C. Mathiessen. A grmnmar and a lexicon 
for a text production system. In Proceedings of the 19th 
Annual Meeting of the Association for Computational 
Linguiztics, pages 49 56, Stanford, California, 1981. 

\[Steinacker and Buchberger, 1983\] I. Steinacker and E. Buch- 
berger. Relating syntax and semantics: the syntactico- 
semantic lexicon of the system VIE-LANG. In Proceed- 
lugs of the First European Meeting of the A CL, Pisa, Italy, 
1983. 

\[Wilensky and Arens, 1980\] R. Wilensky and Y. Areas. 
PHRAN-A knowledge-based natural language under- 
stander. In Proceedings of the 18th Annual Meeting of the 
Association for Computational Linguistics, Philadelphia, 
1980. 

\[Wilks, 1987\] Y. Wilks. TINLAP-3: Theoretical Issue~ in Nat- 
ural Language Processing-3. Computing Research Labo- 
ratory~ New Mexico State University, Las Cruces, New 
Mexico, 1987. 259 
