Adversarial Training for Relation Extraction

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1778–1783 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Adversarial Training for Relation Extraction Yi Wu Computer Science Division UC Berkeley jxwuyi@gmail.com David Bamman School of Information UC Berkeley dbamman@berkeley.edu Stuart Russell Computer Science Division UC Berkeley russell@berkeley.edu Abstract Adversarial training is a mean of regu- larizing classification algorithms by gen- erating adversarial noise to the training data. We apply adversarial training in re- lation extraction within the multi-instance multi-label learning framework. We eval- uate various neural network architectures on two different datasets. Experimental re- sults demonstrate that adversarial training is generally effective for both CNN and RNN models and significantly improves the precision of predicted relations. 1 Introduction Despite the recent successes of deep neural net- works on various applications, neural network models tend to be overconfident about the noise in input signals. Adversarial examples (Szegedy et al., 2013) are examples generated by adding noise in the form of small perturbations to the original data, which are often indistinguishable for humans but drastically increase the loss incurred in a deep model. Adversarial training (Goodfel- low et al., 2014) is a technique for regularizing deep models by encouraging the neural network to correctly classify both unmodified examples and perturbed ones, which in practice not only en- hances the robustness of the neural network but also improves its generalizability. Previous work has largely applied adversarial training on straight- forward classification tasks, including image clas- sification (Goodfellow et al., 2014) and text clas- sification (Miyato et al., 2016), where the goal is simply predicting a single label for every exam- ple and the training examples are able to provide strong supervision. It remains unclear whether ad- versarial training could be still effective for tasks with much weaker supervision, e.g., distant super- vision (Mintz et al., 2009), or a different evalu- ation metric other than prediction accuracy (e.g., F1 score). This paper focuses on the task of relation ex- traction, where the goal is to predict the relation that exists between a particular entity pair given several text mentions. One popular way to han- dle this problem is the multi-instance multi-label learning framework (MIML) (Hoffmann et al., 2011; Surdeanu et al., 2012) with distant super- vision (Mintz et al., 2009), where the mentions for an entity pair are aligned with the relations in Freebase (Bollacker et al., 2008). In this setting, relation extraction is much harder than the canon- ical classification problem in two respects: (1) although distant supervision can provide a large amount of data, the training labels are very noisy, and due to the multi-instance framework, the su- pervision is much weaker; (2) the evaluation met- ric of relation extraction is often the precision- recall curve or F1 score, which cannot be repre- sented (and thereby optimized) directly in the loss function. In order to evaluate the effectiveness of adver- sarial training for relation extraction, we apply it to two different architectures (a convoluational neu- ral …

Content truncated to 500 words