Generating Contrastive Referring Expressions

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 678–687 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1063 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 678–687 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1063 Generating Contrastive Referring Expressions Mart´ın Villalba and Christoph Teichmann and Alexander Koller Department of Language Science and Technology Saarland University, Germany {villalba|cteichmann|koller}@coli.uni-saarland.de Abstract The referring expressions (REs) produced by a natural language generation (NLG) system can be misunderstood by the hearer, even when they are semantically correct. In an interactive setting, the NLG system can try to recognize such misun- derstandings and correct them. We present an algorithm for generating corrective REs that use contrastive focus (“no, the BLUE button”) to emphasize the information the hearer most likely misunderstood. We show empirically that these contrastive REs are preferred over REs without con- trast marking. 1 Introduction Interactive natural language generation (NLG) systems face the task of detecting when they have been misunderstood, and reacting appropriately to fix the problem. For instance, even when the sys- tem generated a semantically correct referring ex- pression (RE), the user may still misunderstand it, i.e. resolve it to a different object from the one the system intended. In an interactive setting, such as a dialogue system or a pedestrian navigation sys- tem, the system can try to detect such misunder- standings – e.g. by predicting what the hearer un- derstood from their behavior (Engonopoulos et al., 2013) – and to produce further utterances which resolve the misunderstanding and get the hearer to identify the intended object after all. When humans correct their own REs, they rou- tinely employ contrastive focus (Rooth, 1992; Krifka, 2008) to clarify the relationship to the orig- inal RE. Say that we originally described an object b as “the blue button”, but the hearer approaches a button b′ which is green, thus providing evidence that they misunderstood the RE to mean b′. In this case, we would like to say “no, the BLUE button”, with the contrastive focus realized by an appropri- ate pitch accent on “BLUE”. This utterance alerts the hearer to the fact that they misunderstood the original RE; it reiterates the information from the original RE; and it marks the attribute “blue” as a salient difference between b′ and the object the original RE was intended to describe. In this paper, we describe an algorithm for gen- erating REs with contrastive focus. We start from the modeling assumption that misunderstandings arise because the RE rs the system uttered was corrupted by a noisy channel into an RE ru which the user “heard” and then resolved correctly; in the example above, we assume the user literally heard “the green button”. We compute this (hypo- thetical) RE ru as the RE which refers to b′ and has the lowest edit distance from rs. Based on this, we mark the contrastive words in rs, …

Content truncated to 500 words