TACITUS: 
The Abductive Commonsense Inference-based 
Text Understanding System 
Jerry R. FIobbs 
SR;I International 
333 Ravenswood Ave. 
Menlo Park, CA 94025 
TACITUS is a natural language processing system 
that is intended to be general and domain-independent. 
It performs a syntactic analysis of the sentences in 
the text, producing a logical form. Next, inferential 
pragmatics processing is applied to the logical form to 
solve problems of schema recognition, reference reso- 
lution, metonymy resolution, and the interpretation of 
vague predicates. An analysis component then produces 
the desired output for the application. TACITUS has 
been applied to several quite different domains, includ- 
ing naval equipment failure reports, naval operations re- 
ports, and terrorist reports. 
Given a text to be analyzed, TACITUS first calls a 
preprocessor to do spelling correction and to translate 
domain-peculiar expressions of times, ranges, tables, and 
so on, into a canonical form. For the terrorist reports, 
we have also implemented a keyword-based statistical 
filter to select the relevant Sentences in the text. This 
eliminates 75% of the sentences from deeper analysis, 
but only eliminates about 10% of the relevant sentences. 
The syntactic component is the DIALOGIC system, 
developed originally for the TEAM transportable, nat- 
uraI language interface to databases. The parser is 
bottom-up and produces all the parses at once, together 
with their logical forms. Its grammar is among the 
largest computer grammars of English in existence, giv- 
ing nearly complete coverage of such phenomena as sen- 
tential complements, relative clauses, adverbials, sen- 
tence fragments, and the most common varieties of con- 
junction. Selectional constraints are applied, and there 
are a large number of heuristics for selecting the pre- 
ferred parses of ambiguous sentences. The logical form 
produced is an "ontologically promiscuous" version of 
first-order predicate calculus, in which relations of gram- 
matical subordination are represented. Optionally and 
where possible, the logical forms for different parses are 
merged into a neutral representation. Experiments are 
now being conducted to determine ways to speed up the 
parsing process, using probabilistic and other criteria for 
scheduling the parser. We are also experimenting with 
ways of extracting information from failed parses. 
Pragmatics processing is based on abductive inference, 
implemented in the Prolog Technology Theorem Prover 
(PTTP), using a knowledge base encoding commonsense 
and domain-specific knowledge in the form of predicate- 
426 
calculus axioms. The fundamental idea is that the in- 
terpretation of a sentence is the minimal proof from the 
knowledge base of the logical form of the sentence to- 
gether with the constraints predicates impose on their 
arguments, allowing for coercions, where one merges re- 
dundancies where possible and makes assumptions where 
necessary. This formulation leads to an elegant, unified 
solution to the problems of schema recognition, reference 
resolution, metonymy resolution, and the interpretation 
of vague predicates. The output of this component is 
an elaborated logical form with the relevant inferences 
drawn and the identities of entities explicitly encoded. 
Finally, an analysis component takes the interpreta- 
tion produced by the pragmatics component and gen- 
erates the required output. For the equipment failure 
reports, this is a diagnosis of the problem described. For 
the naval operations reports and the terrorist reports, 
this is entries for a database. With very little effort, 
analysis components could be constructed for a number 
of other applications, such as message routing and mes- 
sage prioritizing. 
A number of convenient knowledge-acquisition facili- 
ties have been implemented for TACITUS. These include 
a menu-based lexical acquisition component, a sort hier- 
archy editor, and a component allowing entry of axioms 
in a subset of English. 
The people working on this project include Douglas 
Appelt, John Bear, Jerry ttobbs, David Magerman, 
Mark Stickel, and Mabry Tyson. 
References 
\[1\] Hobbs, Jerry R., Mark Stickel, Douglas Appelt, and 
Paul Martin, 1988. "Interpretation as Abduction", 
SRI Technical Note 499, SP,.I International, Menlo 
Park, California. December 1990. 
