Semi-Supervised Tri-Training for Explicit Discourse Argument Expansion

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 104–111 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 104 MuDoCo: Corpus for Multidomain Coreference Resolution and Referring Expression Generation Scott Martin, Shivani Poddar, and Kartikeya Upasani Facebook, Inc. 1 Facebook Way, Menlo Park, CA 94025 USA {scottfb, shivanip, kart}@fb.com Abstract This paper proposes a new dataset, MuDoCo, composed of authored dialogs between a fictional user and a system who are given tasks to perform within six task domains. These dialogs are given rich linguistic annotations by expert linguists for several types of reference mentions and named entity mentions, either of which can span multiple words, as well as for coreference links between mentions. The dialogs sometimes cross and blend domains, and the users exhibit complex task switching behavior such as re-initiating a previous task in the dialog by referencing the entities within it. The dataset contains a total of 8,429 dialogs with an average of 5.36 turns per dialog. We are releasing this dataset to encourage research in the field of coreference resolution, referring expression generation and identification within realistic, deep dialogs involving multiple domains. To demonstrate its utility, we also propose two baseline models for the downstream tasks: coreference resolution and referring expression generation. Keywords: anaphora, coreference, dialog 1. Introduction In this work we introduce a new dataset, MuDoCo, com- posed of a total of 8,429 dialogs with an average of 5.36 turns per dialog. The dataset contains authored dialogs be- tween a ficional user and system, where the user asks the system to perform extensive tasks within multiple domains, as well as switch across a set of 6 task domains (calling, messaging, reminders, weather, news, and music). The users exhibit complex task switching behavior such as • Re-initiating a previous task in the dialog by referenc- ing the entities within it • Re-using entities from a previous task as data for a new task The re-use of entities across task domains introduces the aspect of resolving entities not only within the same do- main, but across multiple ones. For instance, resolving the name of a song, which could also match a street name, to provide coherent information to the user and enable task completion. This work is distinguished from the exist- ing literature in that, to the best of our knowledge, there does not exist any other dataset that is able to provide an insight into coreference and referring expression genera- tion (REG) across domains. We also demonstrate that our dataset can be used for these valuable downstream dialog- focused tasks, and we propose baseline models for each, both built using this new dataset. The work is organized in the following sections. In Sec- tion 2. we examine the various related datasets available in the area, as well as the downstream tasks that can draw value from our new dataset. In Section 3., we describe the proposed dataset MuDoCo, the authoring, annotation, and review processes as well as the …

Content truncated to 500 words