Multilingual Neural Semantic Parsing for Low-Resourced Languages

Proceedings of the 10th Conference on Lexical and Computational Semantics, pages 185–194 August 5–6, 2021, Bangkok, Thailand (online) ©2021 Association for Computational Linguistics 185 Multilingual Neural Semantic Parsing for Low-Resourced Languages Menglin Xia Amazon Research Cambridge ximengli@amazon.com Emilio Monti Amazon Research Cambridge monti@amazon.com Abstract Multilingual semantic parsing is a cost- effective method that allows a single model to understand different languages. However, re- searchers face a great imbalance of availability of training data, with English being resource rich, and other languages having much less data. To tackle the data limitation problem, we propose using machine translation to bootstrap multilingual training data from the more abun- dant English data. To compensate for the data quality of machine translated training data, we utilize transfer learning from pretrained multi- lingual encoders to further improve the model. To evaluate our multilingual models on human- written sentences as opposed to machine trans- lated ones, we introduce a new multilingual se- mantic parsing dataset in English, Italian and Japanese based on the Facebook Task Oriented Parsing (TOP) dataset. We show that joint multilingual training with pretrained encoders substantially outperforms our baselines on the TOP dataset and outperforms the state-of-the- art model on the public NLMaps dataset. We also establish a new baseline for zero-shot learning on the TOP dataset. We find that a semantic parser trained only on English data achieves a zero-shot performance of 44.9% exact-match accuracy on Italian sentences. 1 Introduction Semantic parsing is defined as the task of parsing a natural language sentence into a logical form that represents its meaning. The logical form, or sometimes called the meaning representation lan- guage (MRL) expression, can be executed against a knowledge base to extract information; there- fore, semantic parsing often finds its application in question answering, code generation, information retrieval, etc. Due to its wide range of applications, semantic parsing has drawn a lot of research inter- est. Among them, neural semantic parsing meth- ods have gained popularity in recent years due to their good results (Dong and Lapata, 2018). Neural semantic parsing often formulates the task as a ma- chine translation problem and uses neural networks to translate the sentences into MRL expressions. Multilingual neural semantic parsing is a cost- effective method that allows a single model to un- derstand different languages. However, similar to other machine-learning based approaches, neural semantic parsing requires large amounts of training data. To understand texts in different languages, se- mantic parsing models need training data for each target language. Unfortunately, researchers face a great imbalance of availability of training data for semantic parsing: while we have lots of data in English, the data in non-English languages is of- ten scarce. Although there is a growing number of datasets published for semantic parsing in English, very few datasets are available in other languages. Moreover, manually annotating data for semantic parsing is difficult and time-consuming, as it re- quires a lot of training and effort for annotators to write MRLs. Instead of manually annotating semantic pars- ing data in …

Content truncated to 500 words