Session 2: Natural Language I 
Damaris M. Ayuso, Chair 
BBN Systems and Technologies Corporation 
Cambridge, MA 02138 
This first session on natural language covered a wide 
variety of topics, from formal treatments of semantics to 
research in multi-media interactions. 
The first paper, by L. Schubert and C. Hwang, con- 
centrated on compositionally defining the mapping be- 
tween a sentence and an appropriate meaning represen- 
tation that takes tense and aspect into account. In order to 
derive non-indexical formulae from indexical ones as a 
function of context, they make use of a novel type of con- 
text component, a tense tree which evolves throughout the 
discourse. Occurrences of tense-aspect operators are 
represented as tree branches corresponding to episode 
(event) relationships. An interesting feature of these trees 
is that one can "read off" the tree the relationships implicit 
in the tense-aspect operators and surface ordering of sen- 
tences. 
K. McKeown from Columbia University gave a mul- 
timedia presentation showing examples of COMET's 
capabilities. The presentation focused on the recent ad- 
vances in COMET, in particular, improvements in media 
coordination, text generation, and graphics generation. In 
experiments where both pictures and text are generated in 
an explanation, users were found to strongly prefer that 
sentence breaks correspond to picture breaks. Con- 
sequently the text and graphics generators are now coor- 
dinated to ensure appropriate sentence breaks where pos- 
sible by influencing lexical choice and syntactic structure. 
In text generation, a framework for lexical choice was 
developed using the Functional Unification Formalism, 
which allows the integration of various types of constraints 
uniformly. Finally, the generation of graphics was made 
context-dependent in certain areas, e.g., constraints from 
previous pictures apply to the current picture in order to 
prevent disconcerting, unnecessary changes in viewpoint. 
Although the Y. Schabes and A. Joshi paper covered 
two recent advances in TAGs, synchronous TAGs and LR- 
style parsers for TAGs, Schabes focused in his talk on the 
use of synchronous tags for interpretation. By representing 
semantic expressions as TAG trees, simple mappings can 
be defined between trees in the grammar and these seman- 
tic trees so that via the synchronous TAG process, a logical 
form can be derived as a parse tree is built. This approach 
is similar to the classic syntax-rule/semantic-rule composi- 
tional mappings of other NL systems. However, Schabes 
pointed out that the advantages offered by the extended 
domain of locality of TAGs in syntax are also useful for 
semantic purposes, e.g., in treating idioms and long- 
distance WH-dependencies. The use of synchronous TAGs 
for machine translation was also mentioned, where TAG 
trees in the grammars of two languages are related 
synchronously. 
E. Hovy from ISI/USC reported on a new experiment 
based on an old idea: using a KL-ONE-Iike knowledge 
representation system to perform tasks now normally done 
via unification in unification-based NL systems. The idea 
is that a single knowledge base should encode both syntac- 
tic and semantic knowledge, so that via a single operation, 
classification, the system will deduce the correct 
syntactic/semantic description of the sentence, given an in- 
itial skeletal description. Although knowledge represen- 
tation (KR) systems were not powerful enough when this 
approach was first proposed by Bobrow and Webber in 
1980, the KR system Loom provides the additional needed 
features: more complete inference of disjointness and in- 
ference with respect to coverings. Hovy argued that sub- 
stantial efficiency improvements over unification-based ap- 
proaches could be obtained by this classification approach 
because of inherent structure sharing, consistency main- 
tenance, and explicit type-hierarchy representations. He 
presented an optimistic view of the future cooperation pos- 
sibilities between KR and NL researchers, who have tradi- 
tionally kept their distance. This view is based on the fact 
that Loom developers have already implemented sugges- 
tions useful for NL processing. 
J. Hobbs from SRI reported on his continuing search for 
the optimal point in the efficiency versus power scale for 
the abduction component of TACITUS. Given that abduc- 
tion is combinatoriaUy explosive, an empirical investiga- 
tion is being done exploring 3 different techniques for con- 
trolling the computational cost of abduction: the use of a 
type hierarchy, the unwinding of transitivity axioms, and 
various heuristics for reducing the branch factor of the 
search. Large speedups were reported due to the incor- 
poration of a type hierarchy for pre-f'dtering axioms and 
eliminating some inconsistent assumptions. Where some 
transitivity axioms were creating arbitrary recursion, they 
were "unwound" to limit recursion depth. Finally, several 
heuristics, such as proving easy (specific) conjuncts first 
and propagating those results, were presented for the pur- 
pose of reducing branching of the search. 
33 
