📄 Bidirectional Grammars and the Design of Natural Language Generation Systems

Summary

This is a position paper on the architecture of systems that both understand and generate natural language. The author examines what changes in such a system when it relies on a bidirectional grammar: a single declarative specification of the correspondence between sentences and their meanings that both the parser and the generator use. This approach has several advantages: simpler implementation, a single lexicon, and no need to keep two separate components synchronized. The technical basis is the PATR-II formalism. In PATR-II, unification of feature structures is order-independent, so the rules are not tied to any particular process. For generation, the author proposes a symmetric scheme. A strategic module decides what to say and outputs a logical form with functional features. A tactical module decides how to say it and turns the specification into a sentence. The main finding is the problem of logical-form equivalence. The strategic module cares only about the content of a formula, while the tactical module cares about its syntax. As a result, logically equivalent formulas may yield different sentences or none at all. The paper states this key obstacle to reversible grammars early and clearly.

Key points

• A bidirectional grammar has two basic requirements. First, it must be declarative. Second, it must define a one-to-one mapping between strings and a meaning representation (logical form plus functional features). For example, an active sentence and a passive sentence share a logical form but differ in their features. DIAGRAM is given as an example of a procedural grammar that cannot be used for generation.
• In PATR-II, bidirectionality comes from the associativity and commutativity of unification. Shieber's generator (still work in progress at the time) takes the logical form the parser would produce and reproduces the same sentence.
• An example from electrical circuit repair: the formula Resistance-of(R1, ohm(500)) maps poorly onto the constituents of the sentence, whereas a version with an equality predicate (the verb be) fits better. Swapping the arguments of the equality yields a different sentence. If this conflicts with the functional features, the generator produces nothing.
• The grammars of KAMP, TEXT, MUMBLE and NIGEL/PENMAN have no explicitly represented formal semantics. Their semantics is built into procedures, so these systems avoid the problem but cannot be considered bidirectional.
• The author sees a possible way out in allowing the tactical component limited logical transformations. For example, unification could treat equality and the logical connectives as associative and commutative.