A Neural Approach to KGQA via SPARQL Silhouette Generation

Anonymous

Abstract: Semantic parsing is a predominant approach to solve the Knowledge Graph Question Answering (KGQA) task where, natural language question is translated into a logic form such as SPARQL. Semantic parsing based solutions are mostly modular/pipelined where, noise introduced by the upstream modules for entity/relation linking makes it hard to solve the complex questions. Recently, Neural Machine Translation (NMT) based approaches have emerged that are capable of handling complex questions. However, NMT-based approaches struggle with handling the large number of test entities and relations that are unseen during training. In this work,we propose a modular two-stage neural approach which combines best of both the worlds - NMT and semantic parsing pipeline. Stage-I of our approach comprises an NMT-based seq2seq module that translates a question into a sketch of the desired SPARQL, called as SPARQL silhouette. This stage also contains a noise simulator which combines the masking scheme with an entity/relation linker in a novel manner so as to take care of unseen entities/relation without blowing up the vocabulary of seq2seq module.Stage-II of our approach comprises a Neural Graph Search (NGS) module which aims to distil the SPARQL silhouette in order to reduce the entity/relation linking noise. Experimental results show that, the quality of generated SPARQL silhouette is impressive for an ideal scenario where entity/relation linker is noise-free. For the realistic scenario (i.e. noisy linker), the quality of the SPARQL silhouette drops but our NGS module recovers it considerably. We show that, our proposed approach improves state-of-the-art on LC-QuAD-1 dataset by an absolute margin of $3.72 \%$ $F_1$.