Learning Language Games through Interaction

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 2368–2378, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Learning Language Games through Interaction Sida I. Wang Percy Liang Christopher D. Manning Computer Science Department Stanford University {sidaw,pliang,manning}@cs.stanford.edu Abstract We introduce a new language learning setting relevant to building adaptive nat- ural language interfaces. It is inspired by Wittgenstein’s language games: a hu- man wishes to accomplish some task (e.g., achieving a certain configuration of blocks), but can only communicate with a computer, who performs the actual actions (e.g., removing all red blocks). The com- puter initially knows nothing about lan- guage and therefore must learn it from scratch through interaction, while the hu- man adapts to the computer’s capabilities. We created a game called SHRDLURN in a blocks world and collected interactions from 100 people playing it. First, we an- alyze the humans’ strategies, showing that using compositionality and avoiding syn- onyms correlates positively with task per- formance. Second, we compare computer strategies, showing that modeling prag- matics on a semantic parsing model accel- erates learning for more strategic players. 1 Introduction Wittgenstein (1953) famously said that language derives its meaning from use, and introduced the concept of language games to illustrate the fluid- ity and purpose-orientedness of language. He de- scribed how a builder B and an assistant A can use a primitive language consisting of four words— ‘block’, ‘pillar’, ‘slab’, ‘beam’—to successfully communicate what block to pass from A to B. This is only one such language; many others would also work for accomplishing the cooperative goal. This paper operationalizes and explores the idea of language games in a learning setting, which we call interactive learning through language games Figure 1: The SHRDLURN game: the objective is to transform the start state into the goal state. The human types in an utterance, and the computer (which does not know the goal state) tries to in- terpret the utterance and perform the correspond- ing action. The computer initially knows nothing about the language, but through the human’s feed- back, learns the human’s language while making progress towards the game goal. (ILLG). In the ILLG setting, the two parties do not initially speak a common language, but nonethe- less need to collaboratively accomplish a goal. Specifically, we created a game called SHRD- LURN,1 in homage to the seminal work of Wino- grad (1972). As shown in Figure 1, the objective is to transform a start state into a goal state, but the only action the human can take is entering an utterance. The computer parses the utterance and produces a ranked list of possible interpretations according to its current model. The human scrolls through the list and chooses the intended one, si- multaneously advancing the state of the blocks and providing feedback to the computer. Both the hu- man and the computer wish to reach the goal state 1Demo: http://shrdlurn.sidaw.xyz 2368 (only known to the human) with as little scrolling as possible. For the …

Content truncated to 500 words