Hybrid Semantics for Goal-Directed Natural Language Generation

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1936 - 1946 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Hybrid Semantics for Goal-Directed Natural Language Generation Connor Baumler Case Western Reserve University baumler@umd.edu Soumya Ray Case Western Reserve University sray@case.edu Abstract We consider the problem of generating natu- ral language given a communicative goal and a world description. We ask the question: is it possible to combine complementary meaning representations to scale a goal-directed NLG system without losing expressiveness? In par- ticular, we consider using two meaning rep- resentations, one based on logical semantics and the other based on distributional semantics. We build upon an existing goal-directed gener- ation system, S-STRUCT, which models sen- tence generation as planning in a Markov de- cision process. We develop a hybrid approach, which uses distributional semantics to quickly and imprecisely add the main elements of the sentence and then uses first-order logic based semantics to more slowly add the precise de- tails. We find that our hybrid method allows S-STRUCT’s generation to scale significantly better in early phases of generation and that the hybrid can often generate sentences with the same quality as S-STRUCT in substantially less time. However, we also observe and give insight into cases where the imprecision in dis- tributional semantics leads to generation that is not as good as using pure logical semantics. 1 Introduction We consider the problem of goal-directed natural language generation (NLG) (Gatt and Krahmer, 2018). Here, the agent intends to communicate some information about its world to another entity. It has semantic representations for its world, its goal, and a grammar to realize the language. Given this input, the goal is to generate (realize) a syntac- tically correct representation of the semantic goal without omissions or additions (see Figure 1). This task is different from open ended text generation that fills in text after a prompt or the problem of filling in a blank given some context. Many previous systems for goal-directed NLG use first-order logic (FOL) extended with the λ- calculus to represent semantics (Church, 1985). dog(d1), brown(d1), World dog(d2), golden(d2), cat(c1), white(c1), cat(c2), black(c2), chase(d1, c1), chase(d2, c2) dog(d1), cat(c1), chase(d1,c1) Goal Grammar NP N dog NP_r N_f A brown λv.brown(v) λQ.∃x(dog(x) ∧ Q(x)) etc. S VP NP NP N cat A white D the V chased NP NP N dog A brown D the ∃z1, z2(dog(z1) ∧ brown(z1) ∧ cat(z2) ∧white(z2) ∧ chase(z1, z2)) [z1 : d1, z2 : c1] Figure 1: Example of the goal-directed NLG task. Given a world of facts, a grammar, and a communica- tive goal, generate a grammatical sentence that unam- biguously expresses the goal. This semantic representation allows for very pre- cise generation. However, the process is usually slow, primarily because each step of the generation process needs to check that the semantics of the partially realized text is compatible with the even- tual goal. This step typically involves checking all possible compatible bindings, which is combi- …

Content truncated to 500 words