Q-TOD: A Query-driven Task-oriented Dialogue System

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 761 - 779 December 7-11, 2022 ©2022 Association for Computational Linguistics Reasoning Like Program Executors Xinyu Pi♢∗, Qian Liu§∗, Bei Chen†, Morteza Ziyadi ♡, Zeqi Lin† Qiang Fu†, Yan Gao†, Jian-Guang Lou†, Weizhu Chen♡ ♢University of Illinois Urbana-Champaign, Urbana, USA; §Sea AI Lab, Singapore †Microsoft Research Asia, Beijing, China; ♡Microsoft Azure AI, Redmond, WA, USA xinyupi2@illinois.edu; liuqian@sea.com; {beichen, morteza.ziyadi, zeqi.lin, qifu, yan.gao, jlou, wzchen}@microsoft.com Abstract Reasoning over natural language is a long- standing goal for the research community. However, studies have shown that existing lan- guage models are inadequate in reasoning. To address the issue, we present POET, a novel reasoning pre-training paradigm. Through pre- training language models with programs and their execution results, POET empowers lan- guage models to harvest the reasoning knowl- edge possessed by program executors via a data-driven approach. POET is conceptually simple and can be instantiated by different kinds of program executors. In this paper, we showcase two simple instances POET-Math and POET-Logic, in addition to a complex in- stance, POET-SQL. Experimental results on six benchmarks demonstrate that POET can sig- nificantly boost model performance in natural language reasoning, such as numerical reason- ing, logical reasoning, and multi-hop reason- ing. POET opens a new gate on reasoning- enhancement pre-training, and we hope our analysis would shed light on the future research of reasoning like program executors. 1 Introduction Recent breakthroughs in pre-training illustrate the power of pre-trained Language Models (LM) on a wide range of Natural Language (NL) tasks. Pre- training on self-supervised tasks, such as masked language modeling (Devlin et al., 2019; He et al., 2021) using large amounts of NL sentences, boosts the language understanding of models by a large margin (Wang et al., 2018a). However, existing pre-training paradigms have primarily focused on language modeling and paid little attention to ad- vanced reasoning capabilities (Table 1). As a re- sult, though reaching near-human performance on several tasks, pre-trained LMs are still far behind expectations in reasoning-required scenarios (Rae ∗ The first two authors contributed equally. Work done during internship at Microsoft Research Asia. Language Model Program (e.g., SQL query, python code) Program Context (e.g., database, variables in python) Program Executor (e.g., MySQL, python interpreter) Sentence (e.g., question in reading comprehension) Answer Language Model Natural Context (e.g., passage in reading comprehension) Fine-tuning for Natural Language Reasoning Pre-training by Program Execution Execution Result Transfer Reasoning Knowledge into Language Model Figure 1: Given a program context and a program as input, POET pre-trains LMs to output the execution result. After fine-tuning on downstream tasks, POET can boost LMs on reasoning-required scenarios. Ex- planations about program context, program, program executor and execution result can be found in § 3. More examples of natural context and sentence are in Table 1. et al., 2021), such as numerical reasoning (Wallace et al., 2019; Ravichander et al., 2019) and logical reasoning (Yu et al., 2020; Liu et al., 2020). To alleviate the deficiency, reconciling NL un- derstanding …

Content truncated to 500 words