Generating Bilingual Pragmatic Color References

Proceedings of NAACL-HLT 2018, pages 2155–2165 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Generating Bilingual Pragmatic Color References Will Monroe Computer Science Department Stanford University wmonroe4@cs.stanford.edu Jennifer Hu Department of Mathematics Harvard University jenniferhu@college.harvard.edu Andrew Jong Department of Computer Science San Jose State University andrewjong.cs@gmail.com Christopher Potts Department of Linguistics Stanford University cgpotts@stanford.edu Abstract Contextual influences on language often ex- hibit substantial cross-lingual regularities; for example, we are more verbose in situations that require finer distinctions. However, these regularities are sometimes obscured by seman- tic and syntactic differences. Using a newly- collected dataset of color reference games in Mandarin Chinese (which we release to the public), we confirm that a variety of construc- tions display the same sensitivity to contextual difficulty in Chinese and English. We then show that a neural speaker agent trained on bilingual data with a simple multitask learning approach displays more human-like patterns of context dependence and is more pragmat- ically informative than its monolingual Chi- nese counterpart. Moreover, this is not at the expense of language-specific semantic under- standing: the resulting speaker model learns the different basic color term systems of En- glish and Chinese (with noteworthy cross- lingual influences), and it can identify syn- onyms between the two languages using vector analogy operations on its output layer, despite having no exposure to parallel data. 1 Introduction In grounded communication tasks, speakers face pressures in choosing referential expressions that distinguish their targets from others in the con- text, leading to many kinds of pragmatic mean- ing enrichment. For example, the harder a target is to identify, the more the speaker will feel the need to refer implicitly and explicitly to alterna- tives to draw subtle contrasts (Zipf, 1949; Horn, 1984; Levinson, 2000). However, the ways in which these contrasts are expressed depend heav- ily on language-specific syntax and semantics. x x x 鲜绿 xi¯an l`¨u ‘bright green’ x x x 不亮的橙色 bu-li`ang de ch´engs`e ‘not-bright orange’ x x x 紫红色 zˇı h´ongs`e ‘purple-red’ Figure 1: Reference game contexts and utterances from our Chinese corpus. The boxed color is the target. Some color terms show differences be- tween Chinese and English, such as 绿 l`¨u ‘green’ in the first example for a color that might be re- ferred to with ‘blue’ or ‘aqua’ in English. In this paper, we seek to develop a model of contextual language production that captures language-specific syntax and semantics while also exhibiting responsiveness to contextual differ- ences. We focus on a color reference game (Rosenberg and Cohen, 1964; Dale and Reiter, 1995; Krahmer and van Deemter, 2012) played in both English and Mandarin Chinese. A reference game (Figure 1) involves two agents, one desig- nated the “speaker” and the other the “listener”. The speaker and listener are shown the same set of k colors C = {c1, . . . , ck} (in our experiments, k = 3), and one of these colors ct is indicated secretly to the speaker as the “target”. Both play- …

Content truncated to 500 words