Conversational Multi-Hop Reasoning with Neural Commonsense Knowledge and Symbolic Logic Rules

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7404–7418 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 7404 Conversational Multi-Hop Reasoning with Neural Commonsense Knowledge and Symbolic Logic Rules Forough Arabshahi∗ Facebook forough@fb.com Jennifer Lee∗ Facebook jenniferlee98@fb.com Antoine Bosselut EPFL antoine.bosselut@epfl.ch Yejin Choi University of Washington yejin@cs.washington.edu Tom Mitchell Carnegie Mellon University tom.mitchell@cmu.edu Abstract One of the challenges faced by conversational agents is their inability to identify unstated presumptions of their users’ commands, a task trivial for humans due to their common sense. In this paper, we propose a zero- shot commonsense reasoning system for con- versational agents in an attempt to achieve this. Our reasoner uncovers unstated presump- tions from user commands satisfying a gen- eral template of if-(state ), then-(action ), because-(goal ). Our reasoner uses a state-of- the-art transformer-based generative common- sense knowledge base (KB) as its source of background knowledge for reasoning. We pro- pose a novel and iterative knowledge query mechanism to extract multi-hop reasoning chains from the neural KB which uses sym- bolic logic rules to significantly reduce the search space. Similar to any KBs gathered to date, our commonsense KB is prone to missing knowledge. Therefore, we propose to conversationally elicit the missing knowledge from human users with our novel dynamic question generation strategy, which generates and presents contextualized queries to human users. We evaluate the model with a user study with human users that achieves a 35% higher success rate compared to SOTA. 1 Introduction Conversational agents are becoming prominent in our daily lives thanks to advances in speech recog- nition, natural language processing and machine learning. However, most conversational agents still lack commonsense reasoning, preventing them from engaging in rich conversations with humans. Recently, Arabshahi et al. (2021) proposed a commonsense reasoning benchmark task for conversational agents that contains natural lan- guage commands given to an agent by humans. These commands follow a general template of: ∗ Equal contribution Figure 1: ����’s conversation with a human. The multiple choices 1 through 5 are commonsense knowl- edge obtained from COMET using our multi-hop rea- soner, ranked by the reasoner’s confidence. Since the user chooses Option 2 (Because I stay dry), ���� se- lects this reasoning path as the final reasoning chain. “If (state holds), Then (perform action ), Be- cause (I want to achieve goal )”. We refer to commands satisfying this template as if-then- because commands. As stated in Arabshahi et al. (2021), humans often under-specify conditions on the if-portion ( state ) and/or then-portion ( action ) of their commands. These under- specified conditions are referred to as common- sense presumptions. For example, consider the command, “If it’s going to rain in the afternoon ( · ) Then remind me to bring an umbrella ( · ) Be- cause I want to remain dry”, where ( · ) indicates the position of the unstated commonsense presump- tions. The presumptions for this command are (and I am outside) and (before I leave the house), re- spectively. …

Content truncated to 500 words