Toward Implicit Reference in Dialog: A Survey of Methods and Data

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 587–600 November 20–23, 2022. ©2022 Association for Computational Linguistics 587 Toward Implicit Reference in Dialog: A Survey of Methods and Data Lindsey Vanderlyn Talita Anthonio Daniel Ortega Michael Roth Ngoc Thang Vu University of Stuttgart Institute for Natural Language Processing {lindsey.vanderlyn,talita.anthonio,daniel.ortega, michael.roth,ngoc-thang.vu}@ims.uni-stuttgart.de Abstract Communicating efficiently in natural language requires that we often leave information im- plicit, especially in spontaneous speech. This frequently results in phenomena of incom- pleteness, such as omitted references, that pose challenges for language processing. In this survey paper, we review the state of the art in research regarding the automatic processing of such implicit references in dialog scenarios, discuss weaknesses with respect to inconsis- tencies in task definitions and terminologies, and outline directions for future work. Among others, these include a unification of existing tasks and evaluation metrics, addressing data scarcity, and taking into account model and an- notator uncertainties. 1 Introduction In natural language conversations, speakers often leave out parts of the conversation which are un- derstood by the other party through the shared con- text, as exemplified in Figure 1. This can either serve as a way to add variance to a conversation, to make the dialog more efficient by not repeating information, or to accomplish a specific conver- sational goal, such as displaying skepticism (Car- berry, 1989). These omissions can take the form of syntactically correct sentences that leave out im- portant semantic information or even incomplete sentence fragments (Fernández et al., 2007; Raghu et al., 2015). Figure 1 shows an example of a dialog which contains both of these. Turn 2 demonstrates a syntactically correct sentence where the user asks about the capital, but leaves out which country they are referring to. Turn 3 shows an example of a syntactically incomplete sentence, where the user leaves out both the country and the verbal phrase. In this paper we refer to these omitted entities as implicit references because while there is no direct reference, e.g., a pronoun, it is still understood that the user is referring to a specific entity. We propose Turn Utterance 1 USR Who is the Chancellor of Germany? SYS Olaf Scholz is the current Chancellor of Germany. 2 USR And what is the capital _? [of Germany] SYS The capital of Germany is Berlin. 3 USR And _ the population _? [what is], [of Germany] Figure 1: Dialog between a user USR and a system SYS, with examples of implicit references (in Turn 2 and 3) and another implicit element (in Turn 3) indi- cated by underscores in red. The correct resolution of each implicit element is shown in brackets in red. implicit reference as unifying term encapsulating this type of phenomena and including implicit ar- guments, zero-anaphora, and certain types of noun ellipsis, which we expand on in section 2. While parsing such sentences is a simple task …

Content truncated to 500 words