Multilingual Entity and Relation Extraction from Unified to Language-specific Training

  • 2023-01-11 12:26:53
  • Zixiang Wang, Jian Yang, Tongliang Li, Jiaheng Liu, Ying Mo, Jiaqi Bai, Longtao He, Zhoujun Li
  • 1

Abstract

Entity and relation extraction is a key task in information extraction, wherethe output can be used for downstream NLP tasks. Existing approaches for entityand relation extraction tasks mainly focus on the English corpora and ignoreother languages. Thus, it is critical to improving performance in amultilingual setting. Meanwhile, multilingual training is usually used to boostcross-lingual performance by transferring knowledge from languages (e.g.,high-resource) to other (e.g., low-resource) languages. However, languageinterference usually exists in multilingual tasks as the model parameters areshared among all languages. In this paper, we propose a two-stage multilingualtraining method and a joint model called Multilingual Entity and RelationExtraction framework (mERE) to mitigate language interference across languages.Specifically, we randomly concatenate sentences in different languages to traina Language-universal Aggregator (LA), which narrows the distance of embeddingrepresentations by obtaining the unified language representation. Then, weseparate parameters to mitigate interference via tuning a Language-specificSwitcher (LS), which includes several independent sub-modules to refine thelanguage-specific feature representation. After that, to enhance the relationaltriple extraction, the sentence representations concatenated with the relationfeature are used to recognize the entities. Extensive experimental results showthat our method outperforms both the monolingual and multilingual baselinemethods. Besides, we also perform detailed analysis to show that mERE islightweight but effective on relational triple extraction and mERE{} is easy totransfer to other backbone models of multi-field tasks, which furtherdemonstrates the effectiveness of our method.

 

Quick Read (beta)

loading the full paper ...