Unified Pragmatic Models for Generating and Following Instructions

Proceedings of NAACL-HLT 2018, pages 1951–1963 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Unified Pragmatic Models for Generating and Following Instructions Daniel Fried Jacob Andreas Dan Klein Computer Science Division University of California, Berkeley {dfried,jda,klein}@cs.berkeley.edu Abstract We show that explicit pragmatic inference aids in correctly generating and following natural language instructions for complex, sequential tasks. Our pragmatics-enabled models reason about why speakers produce certain instruc- tions, and about how listeners will react upon hearing them. Like previous pragmatic mod- els, we use learned base listener and speaker models to build a pragmatic speaker that uses the base listener to simulate the interpretation of candidate descriptions, and a pragmatic lis- tener that reasons counterfactually about al- ternative descriptions. We extend these mod- els to tasks with sequential structure. Eval- uation of language generation and interpreta- tion shows that pragmatic inference improves state-of-the-art listener models (at correctly interpreting human instructions) and speaker models (at producing instructions correctly in- terpreted by humans) in diverse settings. 1 Introduction How should speakers and listeners reason about each other when they communicate? A core in- sight of computational pragmatics is that speaker and listener agents operate within a cooperative game-theoretic context, and that each agent ben- efits from reasoning about others’ intents and ac- tions within that context. Pragmatic inference has been studied by a long line of work in linguistics, natural language processing, and cognitive sci- ence. In this paper, we present a technique for lay- ering explicit pragmatic inference on top of mod- els for complex, sequential instruction-following and instruction-generation tasks. We investigate a range of current data sets for both tasks, showing that pragmatic behavior arises naturally from this inference procedure, and gives rise to state-of-the- art results in a variety of domains. Consider the example shown in Figure 1a, in which a speaker agent must describe a route to (a) Behavior Base% Speaker Rational% Speaker ✔ walk forward four times go forward four segments to the intersection with the bare concrete hall (b) Instruction walk along the blue carpet and you pass two objects Base% Listener Rational% Listener ✗ ✔ Figure 1: Real samples for the SAIL navigation en- vironments, comparing base models, without explicit pragmatic inference, to the rational pragmatic infer- ence procedure. (a) The rational speaker, which rea- sons about listener behavior, generates instructions which in this case are more robust to uncertainty about the listener’s initial orientation. (b) The base listener moves to an unintended position (even though it cor- rectly passes two objects). The rational listener, which reasons about the speaker, infers that a route ending at the sofa would have been described differently, and stops earlier. a target position in a hallway. A conventional learned instruction-generating model produces a truthful description of the route (walk forward four times). But the pragmatic speaker in this paper, which is capable of reasoning about the listener, chooses to also include additional information (the intersection with the bare concrete hall), to reduce potential ambiguity and …

Content truncated to 500 words