A Unified Approach to Entity-Centric Context Tracking in Social Conversations

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1275–1285 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1275 A Unified Approach to Entity-Centric Context Tracking in Social Conversations Ulrich R¨uckert, Srinivas Sunkara, Abhinav Rastogi, Sushant Prakash, Pranav Khaitan Google Research 1600 Amphiteatre Pkwy, Mountain View, CA 94043, USA {rueckert, srinivasksun, abhirast, sush, pranavkhaitan}@google.com Abstract In human-human conversations, Context Tracking deals with identifying important entities and keeping track of their properties and relationships. This is a challenging problem that encompasses several subtasks such as slot tagging, coref- erence resolution, resolving plural mentions and entity linking. We approach this problem as an end-to-end modeling task where the conversational context is represented by an entity repository containing the entity references mentioned so far, their properties and the relationships between them. The repository is updated turn-by-turn, thus making training and inference computationally efficient even for long conversations. This paper lays the groundwork for an investigation of this framework in two ways. First, we release Contrack, a large scale human-human conversation corpus for context tracking with people and location annotations. It contains over 7000 conversations with an average of 11.8 turns, 5.8 entities and 15.2 references per conversation. Second, we open-source a neural network architecture for context tracking. Finally we compare this network to state-of-the-art approaches for the subtasks it subsumes and report results on the involved tradeoffs. Keywords: Context Tracking, Coreference Resolution, Entity Linking 1. Introduction Computers and mobile phones have changed how peo- ple communicate. A large amount of today’s interper- sonal communication happens in messaging apps on mobile devices or on chat and discussion services on the internet. Consequently, this has piqued the inter- est of the research community in developing assistive technologies for human-human conversations. Repre- senting the current status of a conversation in a succinct and semantically complete way is a central component of such technologies. At the core of this endeavor lies the task of tracking the entities mentioned in a conver- sation, their properties and the relationships that are be- ing expressed about them. In this paper, we frame this task, which we call Context Tracking, as an online ma- chine learning problem, where the model is expected to track the current status of the conversation at any time. This formulation extends and complements existing re- search in three key areas. First of all, in this framework the model ingests the messages of a conversation turn by turn and updates a growing repository of detected entity references in each turn. This allows for fast inference because the model only needs to ingest a single message and the repos- itory instead of the entire conversation history. That is important particularly for long conversations, where ingesting the entire dialogue history is not feasible. Second, the repository of entity references serves as a condensed storage of the semantic knowledge con- veyed in the conversation. In this sense, context track- ing serves the same purpose for open-domain conver- sations that dialog state …

Content truncated to 500 words