Learning Symmetric Collaborative Dialogue Agents with Dynamic Knowledge Graph Embeddings

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1766–1776 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1162 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1766–1776 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1162 Learning Symmetric Collaborative Dialogue Agents with Dynamic Knowledge Graph Embeddings He He and Anusha Balakrishnan and Mihail Eric and Percy Liang Computer Science Department, Stanford University {hehe,anusha28,meric,pliang}@cs.stanford.edu Abstract We study a symmetric collaborative dia- logue setting in which two agents, each with private knowledge, must strategically communicate to achieve a common goal. The open-ended dialogue state in this set- ting poses new challenges for existing di- alogue systems. We collected a dataset of 11K human-human dialogues, which exhibits interesting lexical, semantic, and strategic elements. To model both struc- tured knowledge and unstructured lan- guage, we propose a neural model with dy- namic knowledge graph embeddings that evolve as the dialogue progresses. Au- tomatic and human evaluations show that our model is both more effective at achiev- ing the goal and more human-like than baseline neural and rule-based models. 1 Introduction Current task-oriented dialogue systems (Young et al., 2013; Wen et al., 2017; Dhingra et al., 2017) require a pre-defined dialogue state (e.g., slots such as food type and price range for a restau- rant searching task) and a fixed set of dialogue acts (e.g., request, inform). However, human conversa- tion often requires richer dialogue states and more nuanced, pragmatic dialogue acts. Recent open- domain chat systems (Shang et al., 2015; Serban et al., 2015b; Sordoni et al., 2015; Li et al., 2016a; Lowe et al., 2017; Mei et al., 2017) learn a map- ping directly from previous utterances to the next utterance. While these models capture open-ended aspects of dialogue, the lack of structured dialogue state prevents them from being directly applied to settings that require interfacing with structured knowledge. In order to bridge the gap between the two types Friends of agent A: Name School Major Company Jessica Columbia Computer Science Google Josh Columbia Linguistics Google ... ... ... ... A: Hi! Most of my friends work for Google B: do you have anyone who went to columbia? A: Hello? A: I have Jessica a friend of mine A: and Josh, both went to columbia B: or anyone working at apple? B: SELECT (Jessica, Columbia, Computer Science, Google) A: SELECT (Jessica, Columbia, Computer Science, Google) Figure 1: An example dialogue from the Mutual- Friends task in which two agents, A and B, each given a private list of a friends, try to identify their mutual friend. Our objective is to build an agent that can perform the task with a human. Cross- talk (Section 2.3) is italicized. of systems, we focus on a symmetric collabora- tive dialogue setting, which is task-oriented but encourages open-ended dialogue acts. In our set- ting, two agents, each with a private list of items with attributes, must communicate …

Content truncated to 500 words