Joint Mention Extraction and Classification with Mention Hypergraphs

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 857–867, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Joint Mention Extraction and Classification with Mention Hypergraphs Wei Lu Singapore University of Technology and Design luwei@sutd.edu.sg Dan Roth University of Illinois at Urbana-Champaign danr@illinois.edu Abstract We present a novel model for the task of joint mention extraction and classifi- cation. Unlike existing approaches, our model is able to effectively capture over- lapping mentions with unbounded lengths. The model is highly scalable, with a time complexity that is linear in the number of words in the input sentence and linear in the number of possible mention classes. Our model can be extended to additionally capture mention heads explicitly in a joint manner under the same time complexity. We demonstrate the effectiveness of our model through extensive experiments on standard datasets. 1 Introduction One of the essential goals in natural language pro- cessing (NLP) is the development of effective sys- tems that can capture the underlying semantics conveyed by human languages. An important step towards such a goal is the development of practi- cal systems that can efficiently extract useful shal- low semantic information such as entities and at the same time identify their semantic classes (e.g., person, organization, etc). Such a task is often known as named entity recognition and classification (NERC), one of the standard tasks in information extraction (IE). While such a task focuses on the extraction and classification of entities in the texts which are named, recently researchers also showed inter- est in a closely related task – mention extraction and classification/typing. Unlike a named entity, a mention is typically defined as a reference to an entity in natural language text that can be ei- ther named, nominal or pronominal (Florian et al., 2004). The task of mention detection and track- ing has received substantial attention, largely due to its important role in conducting several down- stream tasks, such as relation extraction (Mintz et al., 2009), entity linking (Guo et al., 2013), and coreference resolution (Chang et al., 2013). While most existing work on named entity recognition and mention extraction and classifi- cation have been effective, there remain several key limitations associated with existing models. In fact, one can view these problems as instances of the more general problem of semantic tagging – the task of assigning appropriate semantic tags to certain text spans for a given input sentence. Un- like part-of-speech (POS) tagging, which has been extensively studied in the past few decades by the community, such a semantic tagging task presents several additional new challenges. First, a men- tion can consist of multiple words, so its length can be arbitrarily long. Second, the mentions can overlap with one another. Popular models used for POS tagging, such as linear-chain conditional ran- dom fields (Lafferty et al., 2001) or semi-Markov conditional random fields (Sarawagi and Cohen, 2004) have difficulties coping with these issues. While approaches on addressing these issues ex- ist, current algorithms …

Content truncated to 500 words