Facing the most difficult case of Semantic Role Labeling: A collaboration of word embeddings and co-training

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1275–1284, Osaka, Japan, December 11-17 2016. Facing the most difficult case of Semantic Role Labeling: A collaboration of word embeddings and co-training Quynh Ngoc Thi Do1, Steven Bethard2, Marie-Francine Moens1 1Katholieke Universiteit Leuven, Belgium 2University of Arizona, United States quynhngocthi.do@cs.kuleuven.be bethard@email.arizona.edu sien.moens@cs.kuleuven.be Abstract We present a successful collaboration of word embeddings and co-training to tackle in the most difficult test case of semantic role labeling: predicting out-of-domain and unseen semantic frames. Despite the fact that co-training is a successful traditional semi-supervised method, its application in SRL is very limited. In this work, co-training is used together with word embeddings to improve the performance of a system trained on CoNLL 2009 training dataset. We also introduce a semantic role labeling system with a simple learning architecture and effective inference that is easily adaptable to semi-supervised settings with new training data and/or new features. On the out-of-domain testing set of the standard benchmark CoNLL 2009 data our simple approach achieves high performance and improves state-of-the-art results. 1 Introduction Semantic role labeling (SRL) is an essential natural language processing (NLP) task that identifies the relations between a predicate and its arguments in a given sentence. Intuitively, it aims at answering the questions of “Who did What to Whom, and How, When and Where?” in text. For example, the processing of the sentence “He bought tons of roses yesterday” should result in the identification of a “buying” event corresponding to the predicate “bought” with three arguments including “he” as the Agent (A0), “tons of roses” as the Thing being bought (A1), and “yesterday” as the Time (AM-TMP) arguments. Traditional SRL systems have concentrated on supervised learning from several manually-built semantic corpora, (e.g., FrameNet (Baker et al., 1998) and PropBank (Palmer et al., 2005)). One important limitation of supervised approaches is that they depend heavily on the accuracy, coverage and labeling scheme of the labeled corpus. When the training and the testing data are in different domains, the linguistic patterns and their distributions in the testing domain are different from the ones observed in the training data, resulting in a considerable performance drop. Developing more manually-built semantic corpora is expensive and requires huge human efforts. Thus, exploiting large unlabeled datasets by semi-supervised or unsupervised approaches is a promising solution. Our contribution in this paper is two-fold: First, we introduce a SRL system with a simple learning architecture and effective inference that is easily adaptable to new training data or new features in semi- supervised settings. Second, we present a semi-supervised approach that is a combination of using word embeddings as extra features and using a variant of a co-training algorithm to create new training data facing the most difficult cases of SRL: improving the SRL system trained on a relatively large training dataset (CoNLL 2009) when working with the “out-of-domain” and especially “unseen” semantic frames. Although co-training is a successful traditional semi-supervised method, its application to SRL is very limited. To …

Content truncated to 500 words