Structured Context and High-Coverage Grammar for Conversational Question Answering over Knowledge Graphs

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8813–8829 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 8813 Structured Context and High-Coverage Grammar for Conversational Question Answering over Knowledge Graphs Pierre Marion∗ Sorbonne Université, CNRS Laboratoire de Probabilités, Statistique et Modélisation, LPSM F-75005, Paris, France pierre.marion@sorbonne-universite.fr Paweł Krzysztof Nowak and Francesco Piccinno Google Research {pawelnow,piccinno}@google.com Abstract We tackle the problem of weakly-supervised conversational Question Answering over large Knowledge Graphs using a neural semantic parsing approach. We introduce a new Log- ical Form (LF) grammar that can model a wide range of queries on the graph while re- maining sufficiently simple to generate super- vision data efficiently. Our Transformer-based model takes a JSON-like structure as input, al- lowing us to easily incorporate both Knowl- edge Graph and conversational contexts. This structured input is transformed to lists of em- beddings and then fed to standard attention layers. We validate our approach, both in terms of grammar coverage and LF execution accuracy, on two publicly available datasets, CSQA and ConvQuestions, both grounded in Wikidata. On CSQA, our approach increases the coverage from 80% to 96.2%, and the LF execution accuracy from 70.6% to 75.6%, with respect to previous state-of-the-art results. On ConvQuestions, we achieve competitive re- sults with respect to the state-of-the-art. 1 Introduction Graphs are a common abstraction of real-world data. Large-scale knowledge bases can be repre- sented as directed labeled graphs, where entities correspond to nodes and subject-predicate-object triplets are encoded by labeled edges. These so- called Knowledge Graphs (KGs) are used both in open knowledge projects (YAGO, Wikidata) and in the industry (Yahoo, Google, Microsoft, etc.). A prominent task on KGs is factual conversational Question Answering (Conversational KG-QA) and ∗Work done while an intern at Google Research. it has spurred interest recently, in particular due to the development of AI-driven personal assistants. The Conversational KG-QA task involves diffi- culties of different nature: entity disambiguation, long tails of predicates (Saha et al., 2018), conver- sational nature of the interaction. The topology of the underlying graph is also problematic. Not only can KGs be huge (up to several billion entities), they also exhibit hub entities with a large number of neighbors. A recent prominent approach has been to cast the problem as neural semantic parsing (Jia and Liang, 2016; Dong and Lapata, 2016, 2018; Shen et al., 2019). In this setting, a semantic parsing model learns to map a natural language question to a logical form (LF), i.e. a tree of operators over the KG. These operators belong to some grammar, ei- ther standard like SPARQL or ad-hoc. The logical form is then evaluated over the KG to produce the candidate answer. In the weak supervision training setup, the true logical form is not available, but only the answer utterance is (as well as annotated entities in some cases, see Section 5.4). Hence the training data is not given but it is instead generated, in the format of (question, logical form) pairs. We refer to this data …

Content truncated to 500 words