Improved Information Structure Analysis of Scientific Documents Through Discourse and Lexical Constraints

Proceedings of NAACL-HLT 2013, pages 928–937, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Improved Information Structure Analysis of Scientific Documents Through Discourse and Lexical Constraints Yufan Guo University of Cambridge, UK yg244@cam.ac.uk Roi Reichart University of Cambridge, UK rr439@cam.ac.uk Anna Korhonen University of Cambridge, UK alk23@cam.ac.uk Abstract Inferring the information structure of scien- tific documents is useful for many down- stream applications. Existing feature-based machine learning approaches to this task re- quire substantial training data and suffer from limited performance. Our idea is to guide feature-based models with declarative domain knowledge encoded as posterior distribution constraints. We explore a rich set of discourse and lexical constraints which we incorporate through the Generalized Expectation (GE) cri- terion. Our constrained model improves the performance of existing fully and lightly su- pervised models. Even a fully unsupervised version of this model outperforms lightly su- pervised feature-based models, showing that our approach can be useful even when no la- beled data is available. 1 Introduction Techniques that enable automatic analysis of the in- formation structure of scientific articles can help sci- entists identify information of interest in the grow- ing volume of scientific literature. For example, classification of sentences according to argumenta- tive zones (AZ) – an information structure scheme that is applicable across scientific domains (Teufel et al., 2009) – can support information retrieval, in- formation extraction and summarization (Teufel and Moens, 2002; Tbahriti et al., 2006; Ruch et al., 2007; Liakata et al., 2012; Contractor et al., 2012). Previous work on sentence-based classification of scientific literature according to categories of infor- mation structure has mostly used feature-based ma- chine learning, such as Support Vector Machines (SVM) and Conditional Random Fields (CRF) (e.g. (Teufel and Moens, 2002; Lin et al., 2006; Hiro- hata et al., 2008; Shatkay et al., 2008; Guo et al., 2010; Liakata et al., 2012)). Unfortunately, the per- formance of these methods is rather limited, as indi- cated e.g. by the relatively low numbers reported by Liakata et al. (2012) in biochemistry and chemistry with per-class F-scores ranging from .18 to .76. We propose a novel approach to this task in which traditional feature-based models are augmented with explicit declarative expert and domain knowledge, and apply it to sentence-based AZ. We explore two sources of declarative knowledge for our task - dis- course and lexical. One way to utilize discourse knowledge is to guide the model predictions by en- coding a desired predicted class (i.e. information category) distribution in a given position in the doc- ument. Consider, for example, sentence (1) from the first paragraph of the Discussion section in a paper: (1) In time, this will prove to be most suitable for detailed analysis of the role of these hormones in mammary cancer development. Although the future tense and cue phrases such as “in time” can indicate that authors are discussing fu- ture work (i.e. the “Future work” class in the AZ scheme), in this case they refer to their own contri- bution (i.e. the “Conclusion” …

Content truncated to 500 words