Resolving Indirect Referring Expressions for Entity Selection

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12313–12335 July 9-14, 2023 ©2023 Association for Computational Linguistics Resolving Indirect Referring Expressions for Entity Selection Mohammad Javad Hosseini Filip Radlinski Silvia Pareti Annie Louis Google Research {javadh,filiprad,spareti,annielouis}@google.com Abstract Recent advances in language modeling have enabled new conversational systems. In par- ticular, it is often desirable for people to make choices among specified options when using such systems. We address this problem of ref- erence resolution, when people use natural ex- pressions to choose between the entities. For example, given the choice ‘Should we make a Simnel cake or a Pandan cake?’ a natural response from a dialog participant may be in- direct: ‘let’s make the green one’. Such nat- ural expressions have been little studied for reference resolution. We argue that robustly understanding such language has large poten- tial for improving naturalness in dialog, rec- ommendation, and search systems. We create AltEntities1 (Alternative Entities), a new public dataset of 42K entity pairs and expres- sions (referring to one entity in the pair), and develop models for the disambiguation prob- lem. Consisting of indirect referring expres- sions across three domains, our corpus enables for the first time the study of how language models can be adapted to this task. We find they achieve 82%-87% accuracy in realistic settings, which while reasonable also invites further advances. 1 Introduction Natural dialog often requires resolving referring expressions (REs), not only within and across texts, but also for grounding natural language expres- sions to specific entities or images. We focus on a specific conversational setting where a speaker’s utterance intends to disambiguate between known named entities. While many aspects of RE resolu- tion have been studied extensively, past work has focused on pragmatic reasoning (Dale and Reiter, 1995; Frank and Goodman, 2012), influence of dis- course (Orita et al., 2015), and multimodal (e.g., image) context (Zhang et al., 2018). 1Our dataset can be found at https://github.com/ google-research-datasets/AltEntities Did you mean a Simnel or Pandan cake? It looks surprisingly green in color Without any frosting or fruit It is made from some leaf Comes from Indonesia Isn’t the Easter one Table 1: Responses to the question which intend to choose Pandan cake over the alternative. In the specific case of dialog, when people make choices, the natural REs are not always item names, spatial locations or attributes present in the ques- tion. For instance when the choice is among items with similar names (perhaps disambiguating au- tomatic speech recognition errors), or items with difficult to pronounce names, or where the user does not even recall which name is correct but in- stead recalls some higher level attribute, the user may choose an indirect expression (Table 1). Most related to our work, Celikyilmaz et al. (2014) pre- viously studied REs in response to a set of related items (e.g., Harry Potter movies) shown in a user interface. Their work both contains direct (using entity name), indirect, as well as …

Content truncated to 500 words