Anzeige
Erschienen in:
2019 | OriginalPaper | Buchkapitel
A Transformer-Based Semantic Parser for NLPCC-2019 Shared Task 2
verfasst von : Donglai Ge, Junhui Li, Muhua Zhu
Erschienen in: Natural Language Processing and Chinese Computing
Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by
Abstract
Sequence-to-Sequence (seq2seq) approaches formalize semantic parsing as a translation task from a source sentence to its corresponding logical form. However, in the absence of large-scale annotated dataset, even the state-of-the-art seq2seq model, i.e., the Transformer may suffer from the data sparsity issue. In order to address this issue, this paper explores three techniques which are widely used in neural machine translation to better adapt seq2seq models for semantic parsing. First, we use byte pair encoding (BPE) to segment words into subwords to transfer rare words into frequent subwords. Second, we share word vocabulary on both the source and the target sides. Finally, we define heuristic rules to generate synthetic instances to increase the coverage of training dataset. Experimental results on the NLPCC 2019 shared task 2 show that our approach achieves state-of-the-art performance and gets the first place in the task from the current rankings.