Adversarial Multi-lingual Neural Relation Extraction

Proceedings of the 27th International Conference on Computational Linguistics, pages 1156–1166 Santa Fe, New Mexico, USA, August 20-26, 2018. 1156 Adversarial Multi-lingual Neural Relation Extraction Xiaozhi Wang1∗, Xu Han1∗, Yankai Lin1, Zhiyuan Liu1†, Maosong Sun1,2 1Department of Computer Science and Technology, State Key Lab on Intelligent Technology and Systems, Beijing National Research Center for Information Science and Technology, Tsinghua University, Beijing, China 2Beijing Advanced Innovation Center for Imaging Technology, Capital Normal University, Beijing, China Abstract Multi-lingual relation extraction aims to find unknown relational facts from text in various lan- guages. Existing models cannot well capture the consistency and diversity of relation patterns in different languages. To address these issues, we propose an adversarial multi-lingual neural rela- tion extraction (AMNRE) model, which builds both consistent and individual representations for each sentence to consider the consistency and diversity among languages. Further, we adopt an adversarial training strategy to ensure those consistent sentence representations could effectively extract the language-consistent relation patterns. The experimental results on real-world datasets demonstrate that our AMNRE model significantly outperforms the state-of-the-art models. The source code of this paper can be obtained from https://github.com/thunlp/AMNRE. 1 Introduction Relation extraction (RE) is a crucial task in NLP, which aims to extract semantic relations between entity pairs from the sentences containing them. For example, given an entity pair (Bill Gates, Microsoft) and a sentence “Bill Gates is the co-founder and CEO of Microsoft”, we want to figure out the relation Founder between the two entities. RE can potentially benefit many applications, such as knowledge base construction (Zhong et al., 2015; Han et al., 2018) and question answering (Xiang et al., 2017). Recently, neural models have shown their great abilities in RE. Zeng et al. (2014) introduce a convolu- tional neural network (CNN) to extract relational facts with automatically learning features from text. To address the issue of lack of data, Zeng et al. (2015) incorporate multi-instance learning with a piece-wise convolutional neural network (PCNN) to extract relations in distantly supervised data. Because distant supervision suffer from wrong labeling problems, Lin et al. (2016) further employ a sentence-level selec- tive attention to filter out those noisy sentences in distantly supervised data and achieve state-of-the-art performance. All these neural relation extraction (NRE) models merely focus on extracting relational facts from mono-lingual data, ignoring the rich information in multi-lingual data. Lin et al. (2017) propose a multi-lingual attention-based neural relation extraction (MNRE) model, which considers the consistency and complementarity in multi-lingual data. MNRE builds a sentence representation for each sentence in various languages and employs a multi-lingual attention to capture the pattern consistency and complementarity among languages. Although MNRE achieves great success in multi-lingual RE, it still has some problems. MNRE learns a single representation for each sentence in various languages, which cannot well capture both the con- sistency and diversity of relation patterns in different languages. Moreover, MNRE simply utilizes a multi-lingual attention mechanism and a global relation predictor to capture the consistent relation pat- terns among multiple languages. From the experimental data, we …

Content truncated to 500 words