Learning from Omission

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 619–628 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 619 Learning from omission Bill McDowell Duolingo Pittsburgh, PA 15206 mcdowell@duolingo.com Noah D. Goodman Stanford University Stanford, CA 94305 ngoodman@stanford.edu Abstract Pragmatic reasoning allows humans to go be- yond the literal meaning when interpreting language in context. Previous work has shown that such reasoning can improve the per- formance of already-trained language under- standing systems. Here, we explore whether pragmatic reasoning during training can im- prove the quality of learned meanings. Our experiments on reference game data show that end-to-end pragmatic training produces more accurate utterance interpretation models, es- pecially when data is sparse and language is complex. 1 Introduction We often draw pragmatic inferences about a speaker’s intentions from what they choose to say, but also from what they choose not to say in con- text. This pragmatic reasoning arises from listen- ers’ inferences based on speakers’ cooperativity (Grice, 1975), and prior work has observed that such reasoning enables human children to more quickly learn word meanings (Frank and Good- man, 2014). This suggests that pragmatic reason- ing might allow modern neural network models to more efficiently learn on grounded language data from cooperative reference games. As a motivating case, consider an instance of the color reference task from Monroe et al. (2017)— shown in the first row of Table 1. In this task, a speaker communicates a target color to a lis- tener in a context containing two distractor colors; the listener picks out the target based on what the speaker says. In the first instance from Table 1, the speaker utters “dark blue” to describe the tar- get. Whereas “dark” and “blue” also apply to the target, they lose their informativity in the presence of the distractors, and so the speaker pragmatically opts for “dark blue”. A listener who is learning the language from such examples might draw several inferences from the speaker’s utterance. First, under the assump- tion that the speaker is informative, a “literal” learner might infer that “dark blue” applies to the target shade more than the distractors. Second, a “pragmatic” learner might consider the cheaper alternatives–“dark” and “blue”–that have occurred in the presence of the same target in prior contexts, and infer that these alternative utterances must also apply to the distractors given the speaker’s fail- ure to use them. The pragmatic learner might thus gain more semantic knowledge from the same training instances than the literal learner: prag- matic reasoning can reduce the data complexity of learning. The pragmatic learning effects just described depend on the existence of low cost alternative ut- terances that the learner already knows can apply to the target object. The existence of short alter- natives will be more likely when the target objects are more complex (as in row 2 of Table 1), be- cause these objects require longer utterances (with therefore more short alternatives) to individuate. Thus, we further …

Content truncated to 500 words