Learning Soft Linear Constraints with Application to Citation Field Extraction

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 593–602, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Learning Soft Linear Constraints with Application to Citation Field Extraction Sam Anzaroot Alexandre Passos David Belanger Andrew McCallum Department of Computer Science University of Massachusetts, Amherst {anzaroot, apassos, belanger, mccallum}@cs.umass.edu Abstract Accurately segmenting a citation string into fields for authors, titles, etc. is a chal- lenging task because the output typically obeys various global constraints. Previous work has shown that modeling soft con- straints, where the model is encouraged, but not require to obey the constraints, can substantially improve segmentation per- formance. On the other hand, for impos- ing hard constraints, dual decomposition is a popular technique for efficient predic- tion given existing algorithms for uncon- strained inference. We extend dual decom- position to perform prediction subject to soft constraints. Moreover, with a tech- nique for performing inference given soft constraints, it is easy to automatically gen- erate large families of constraints and learn their costs with a simple convex optimiza- tion problem during training. This allows us to obtain substantial gains in accuracy on a new, challenging citation extraction dataset. 1 Introduction Citation field extraction, an instance of informa- tion extraction, is the task of segmenting and la- beling research paper citation strings into their constituent parts, including authors, editors, year, journal, volume, conference venue, etc. This task is important because citation data is often pro- vided only in plain text; however, having an ac- curate structured database of bibliographic infor- mation is necessary for many scientometric tasks, such as mapping scientific sub-communities, dis- covering research trends, and analyzing networks of researchers. Automated citation field extrac- tion needs further research because it has not yet reached a level of accuracy at which it can be prac- tically deployed in real-world systems. Hidden Markov models and linear-chain condi- tional random fields (CRFs) have previously been applied to citation extraction (Hetzner, 2008; Peng and McCallum, 2004) . These models support ef- ficient dynamic-programming inference, but only model local dependencies in the output label se- quence. However citations have strong global reg- ularities not captured by these models. For exam- ple many book citations contain both an author section and an editor section, but none have two disjoint author sections. Since linear-chain mod- els are unable to capture more than Markov depen- dencies, the models sometimes mislabel the editor as a second author. If we could enforce the global constraint that there should be only one author section, accuracy could be improved. One framework for adding such global con- straints into tractable models is constrained infer- ence, in which at inference time the original model is augmented with restrictions on the outputs such that they obey certain global regularities. When hard constraints can be encoded as linear equa- tions on the output variables, and the underlying model’s inference task can be posed as linear opti- mization, one can formulate this constrained infer- ence problem as an integer …

Content truncated to 500 words