Pragmatically Informative Text Generation

Proceedings of NAACL-HLT 2019, pages 4060–4067 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 4060 Pragmatically Informative Text Generation Sheng Shen† Daniel Fried† Jacob Andreas‡ Dan Klein† †Computer Science Division, UC Berkeley ‡Computer Science and Artificial Intelligence Laboratory, MIT {sheng.s,dfried,klein}@berkeley.edu, jda@mit.edu Abstract We improve the informativeness of models for conditional text generation using techniques from computational pragmatics. These tech- niques formulate language production as a game between speakers and listeners, in which a speaker should generate output text that a lis- tener can use to correctly identify the original input that the text describes. While such ap- proaches are widely used in cognitive science and grounded language learning, they have re- ceived less attention for more standard lan- guage generation tasks. We consider two prag- matic modeling methods for text generation: one where pragmatics is imposed by informa- tion preservation, and another where pragmat- ics is imposed by explicit modeling of distrac- tors. We find that these methods improve the performance of strong existing systems for ab- stractive summarization and generation from structured meaning representations. 1 Introduction Computational approaches to pragmatics cast lan- guage generation and interpretation as game- theoretic or Bayesian inference procedures (Gol- land et al., 2010; Frank and Goodman, 2012). While such approaches are capable of modeling a variety of pragmatic phenomena, their main ap- plication in natural language processing has been to improve the informativeness of generated text in grounded language learning problems (Monroe et al., 2018). In this paper, we show that prag- matic reasoning can be similarly used to improve performance in more traditional language genera- tion tasks like generation from structured meaning representations (Figure 1) and summarization. Our work builds on a line of learned Rational Speech Acts (RSA) models (Monroe and Potts, 2015; Andreas and Klein, 2016), in which gen- erated strings are selected to optimize the behav- Input meaning representation (i): NAME[FITZBILLIES], EATTYPE[COFFEE SHOP], FOOD[ENGLISH], PRICERANGE[CHEAP], CUSTOMERRATING[5 OUT OF 5], AREA[RIVERSIDE], FAMILYFRIENDLY[YES] Human-written A cheap coffee shop in riverside with a 5 out of 5 customer rating is Fitzbillies. Fitzbillies is family friendly and serves English food. Base sequence-to-sequence model (S0) Fitzbillies is a family friendly coffee shop located near the river. Distractor-based pragmatic system (SD 1 ) Fitzbillies is a family friendly coffee shop that serves English food. It is located in riverside area. It has a customer rating of 5 out of 5 and is cheap. Reconstructor-based pragmatic system (SR 1 ) Fitzbillies is a family friendly coffee shop that serves cheap English food in the riverside area. It has a customer rating of 5 out of 5. Figure 1: Example outputs of our systems on the E2E generation task. While a base sequence-to-sequence model (S0, Sec. 2) fails to describe all attributes in the input meaning representation, both of our prag- matic systems (SR 1 , Sec. 3.1 and SD 1 , Sec. 3.2) and the human-written reference do. ior of an embedded listener model. The canoni- cal presentation of the RSA framework …

Content truncated to 500 words