Relation Extraction with Multi-instance Multi-label Convolutional Neural Networks

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1471–1480, Osaka, Japan, December 11-17 2016. Relation Extraction with Multi-instance Multi-label Convolutional Neural Networks Xiaotian Jiang†‡, Quan Wang†‡∗, Peng Li†‡, Bin Wang†‡ †Institute of Information Engineering, Chinese Academy of Sciences No.89A Minzhuang Road, Beijing 100093, China ‡University of Chinese Academy of Sciences No.19A Yuquan Road, Beijing 100049, China {jiangxiaotian,wangquan,lipeng,wangbin}@iie.ac.cn Abstract Distant supervision is an efficient approach that automatically generates labeled data for relation extraction (RE). Traditional distantly supervised RE systems rely heavily on handcrafted fea- tures, and hence suffer from error propagation. Recently, a neural network architecture has been proposed to automatically extract features for relation classification. However, this approach follows the traditional expressed-at-least-once assumption, and fails to make full use of informa- tion across different sentences. Moreover, it ignores the fact that there can be multiple relations holding between the same entity pair. In this paper, we propose a multi-instance multi-label con- volutional neural network for distantly supervised RE. It first relaxes the expressed-at-least-once assumption, and employs cross-sentence max-pooling so as to enable information sharing across different sentences. Then it handles overlapping relations by multi-label learning with a neu- ral network classifier. Experimental results show that our approach performs significantly and consistently better than state-of-the-art methods. 1 Introduction Relation extraction (RE), defined as the task of extracting binary relations from plain text, has long been a crucial task in natural language processing. Supervised methods are widely used for this task due to their relatively high performance (Zhou et al., 2005; Surdeanu and Ciaramita, 2007). Such methods, however, usually require intensive human annotation and can be time-consuming. To address this issue, distant supervision is proposed to generate labeled data automatically, by aligning facts in a knowledge base (KB) with sentences mentioning these facts (Mintz et al., 2009; Riedel et al., 2010; Riedel et al., 2013). Traditional (distantly) supervised RE methods use as input numerous lexical and syntactic features, e.g., POS tags, dependency paths, and named entity tags (Mintz et al., 2009; Riedel et al., 2010; Hoff- mann et al., 2011; Surdeanu et al., 2012). These features are extracted from sentences using various NLP algorithms, thus inevitably have errors. The induced errors become more serious for long sentences (McDonald and Nivre, 2007), which is unfortunately very common in real-world relation extraction corpus (Zeng et al., 2015). Building distant supervision methods on faulty features inevitably leads to error propagation, the main culprit responsible for performance degradation. Recent studies have shown promising results on using deep neural networks for automatic feature extraction (Zeng et al., 2014; Liu et al., 2015; Xu et al., 2015). Particularly, Zeng et al. (2015) proposed a piecewise convolutional neural network (PCNN) architecture, which can build an extractor based on distant supervision. PCNN auto- matically extracts features with convolutional neural networks, and introduces piecewise max-pooling to better fit the RE scenario. Although PCNN achieves substantial improvements in distantly supervised relation extraction, it still has the following deficiencies. First, PCNN uses the expressed-at-least-once assumption (Riedel et …

Content truncated to 500 words