Adversarial Feature Adaptation for Cross-lingual Relation Classification

Proceedings of the 27th International Conference on Computational Linguistics, pages 437–448 Santa Fe, New Mexico, USA, August 20-26, 2018. 437 Adversarial Feature Adaptation for Cross-lingual Relation Classification Bowei Zou, Zengzhuang Xu, Yu Hong, Guodong Zhou∗ School of Computer Science and Technology, Soochow University, Suzhou, 215006, China zoubowei@suda.edu.cn, nedxuwork@gmail.com, {yhong, gdzou}@suda.edu.cn Abstract Relation Classification aims to classify the semantic relationship between two marked entities in a given sentence. It plays a vital role in a variety of natural language processing applications. Most existing methods focus on exploiting mono-lingual data, e.g., in English, due to the lack of annotated data in other languages. In this paper, we come up with a feature adaptation approach for cross-lingual relation classification, which employs a generative adversarial network (GAN) to transfer feature representations from one language with rich annotated data to another language with scarce annotated data. Such a feature adaptation approach enables feature imitation via the competition between a relation classification network and a rival discriminator. Experimental results on the ACE 2005 multilingual training corpus, treating English as the source language and Chinese the target, demonstrate the effectiveness of our proposed approach, yielding an improvement of 5.7% over the state-of-the-art. 1 Introduction Relation classification aims to identify the semantic relationship between two nominals labeled in a given sentence. It is critical to many natural language processing (NLP) applications, such as question answering and knowledge base population. For example, the following sentence contains an instance of the Content-Container(e2,e1) relation between two labeled entity mentions “e1=cartridge” and “e2=ink”. The [cartridge]e1 was marked as empty, even with [ink]e2 in both chambers. An open challenge is how to train a model which is suitable for languages with insufficient available data of relation classification, since manual annotation is time-consuming and human-intensive. This makes it difficult to transferrably use the existing well-trained classification models in other languages. To tackle this problem, we propose an adversarial feature adaptation approach to transfer latent feature representations from the source language with rich labeled data to the target language with only unlabeled data. Such an approach are both discriminative for relation classification and invariant across languages. This is largely motivated by the adversarial mechanism which has been effectively applied to measure the similarity between distributions in a variety of scenarios, such as domain adaptation (Bousmalis et al., 2016) and multi-modal representation learning (Park and Im, 2016). In particular, we build two counterpart networks, using convolutional neural networks (CNNs), for source language and target language, to generate the latent feature representations respectively. Then, we use a rival discriminator to identify the correct source of feature representations. At the training step, the network of the source language is trained to maximize the performance on the annotated dataset, while the network of the target language is trained to imitate the feature representations of the source language by rival confusing the discriminator. We perform the proposed approach on ACE 2005 multilingual training corpus by regarding English as the richly-labeled language (source) and Chinese as the poorly-labeled language (target). Our …

Content truncated to 500 words