Logical Form Generation via Multi-task Learning for Complex Question Answering over Knowledge Bases

Proceedings of the 29th International Conference on Computational Linguistics, pages 1687–1696 October 12–17, 2022. 1687 Logical Form Generation via Multi-task Learning for Complex Question Answering over Knowledge Bases Xixin Hu, Xuan Wu, Yiheng Shu, Yuzhong Qu State Key Laboratory for Novel Software Technology, Nanjing University, China {xixinhu,xuanwu,yhshu}@smail.nju.edu.cn, yzqu@nju.edu.cn Abstract Question answering over knowledge bases (KBQA) for complex questions is a challenging task in natural language processing. Recently, generation-based methods that translate natural language questions to executable logical forms have achieved promising performance. These methods use auxiliary information to augment the logical form generation of questions with unseen KB items or novel combinations, but the noise introduced can also leads to more in- correct results. In this work, we propose GMT- KBQA, a Generation-based KBQA method via Multi-Task learning, to better retrieve and utilize auxiliary information. GMT-KBQA first obtains candidate entities and relations through dense retrieval, and then introduces a multi- task model which jointly learns entity disam- biguation, relation classification, and logical form generation. Experimental results show that GMT-KBQA achieves state-of-the-art re- sults on both COMPLEXWEBQUESTIONS and WEBQUESTIONSSP datasets. Furthermore, the detailed evaluation demonstrates that GMT- KBQA benefits from the auxiliary tasks and has a strong generalization capability.1 1 Introduction Question answering over knowledge bases (KBQA) is the task of answering natural language questions based on the facts stored in knowledge bases (KBs). In recent years, an increasing number of KBQA methods arise with the emergence of large-scale KBs, such as Freebase (Bollacker et al., 2008), DBpedia (Lehmann et al., 2015) and Wikidata (Vrandeˇci´c and Krötzsch, 2014). A mainstream paradigm of KBQA methods is semantic parsing (SP) (Berant et al., 2013), where natural language questions are parsed into logical forms such as λ- DCS (Liang, 2013), SPARQL (Pérez et al., 2009), S-expression (Gu et al., 2021), etc. However, com- plex questions that involve reasoning over multiple 1Our code is available at https://github.com/ HXX97/GMT-KBQA. entities, relations, or constraints remain a challenge for SP-based methods. Most of the SP-based meth- ods use a pipeline including entity/relation link- ing, constraint detection, and logical form building (Singh et al., 2018; Hu et al., 2021). As complex questions require multiple entities and relations, errors introduced by previous linkers reduce the performance of the pipeline. With the success of applying natural language generation to various tasks (Raffel et al., 2019; Rothe et al., 2020), recent KBQA methods (Cao et al., 2022; Yin et al., 2021; Gu et al., 2021) cast semantic parsing to a logical form generation task in a sequence-to-sequence (Seq2Seq) manner, fine- tuning pre-trained encoder-decoder models to gen- erate logical forms from natural language questions. However, it is impractical for a simple Seq2Seq model to generate unseen entities and relations that never appear in the training set. To allevi- ate such cases, more generation-based methods (Huang et al., 2021; Das et al., 2021; Ye et al., 2021) augment logical form generation with auxil- iary information such as linked entities (Huang et al., 2021), similar question-query pairs (Das et al., 2021), candidate logical forms (Ye …

Content truncated to 500 words