A logical approach to narrative understanding
Chung Hee Hwang · ERA: Education and Research Archive (University of Alberta) · 1992
It is argued that a theory bottleneck encountered in the 70's and early 80's in attempts to build comprehensive NLU systems led to a fragmentation of NLU research, which still persists. NLU is an organic phenomenon, and enough has been learned about the vexing problems of the 80's to try to integrate these insights and build more comprehensive theories and extensible implementations. On that premise, a new comprehensive framework for narrative understanding has been developed. Its centerpiece is a new situational logic called Episodic Logic (EL), a highly expressive knowledge and semantic representation well-adapted to the interpretive and inferential needs of general NLU. EL is Montague-inspired and influenced by situation semantics. It provides an easily computed first order logical form for English. It allows propositional attitudes, unreliable generalizations, and other non-standard constructs, including ones involving events, actions, facts, kinds and donkey sentences. It incorporates a DRT-like treatment of indefinites, and makes systematic use of episodic variables in the representation of episodic sentences, using them to capture temporal and causal relationships. The rules of inference in EL include probabilistic versions of deduction rules resembling forward and backward chaining rules in expert systems. Also developed is a uniform, compositional approach to interpretation in which a parse tree leads directly (in rule-to-rule fashion) to a preliminary, indexical logical form, and this indexical LF is deindexed with respect to the current context (a well-defined structure). The initial translation is obtained using a GPSG-style grammar; the latter transformation is accomplished by a new recursive deindexing mechanism. Deindexing simultaneously transforms the LF and the context: context-dependent constituents of the LF, including tense, aspect and temporal adverbials, are replaced by explicit relations among quantified episodes, bringing the context information into the LF, thus removing context dependency; and new structural components and episode tokens are added to the context. The relevant context structures are called tense trees. The mechanism allows reference episodes to be correctly identified even for embedded clauses. Finally, a pilot implementation is able to make many (though not all) of the inferences described in this thesis, and has been successfully used in several domains.