WING-NUS at SemEval-2017 Task 10: Keyphrase Extraction and Classification as Joint Sequence Labeling

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 973–977, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics WING-NUS at SemEval-2017 Task 10: Keyphrase Identification and Classification as Joint Sequence Labeling Animesh Prasad Min-Yen Kan School of Computing, National University of Singapore Computing 1, 13 Computing Drive, Singapore 11741 {animesh,kanmy}@comp.nus.edu.sg Abstract We describe an end-to-end pipeline pro- cessing approach for SemEval 2017’s Task 10 to extract keyphrases and their relations from scientific publications. We jointly identify and classify keyphrases by mod- eling the subtasks as sequential labeling. Our system utilizes standard, surface-level features along with the adjacent word fea- tures, and performs conditional decoding on whole text to extract keyphrases. We focus only on the identification and typing of keyphrases (Subtasks A and B, together referred as extraction), but pro- vide an end-to-end system inclusive of keyphrase relation identification (Subtask C) for completeness. Our top performing configuration achieves an F1 of 0.27 for the end-to-end keyphrase extraction and relation identification scenario on the fi- nal test data, and compares on par to other top ranked systems for keyphrase extrac- tion. Our system outperforms other tech- niques that do not employ global decod- ing and hence do not account for depen- dencies between keyphrases. We believe this is crucial for keyphrase classification in the given context of scientific document mining. 1 Introduction Keyphrases are often used for representing the salient concepts of a document. In scientific docu- ments, keyphrase extraction is an important pre- requisite task that feeds downstream tasks such as summarization, clustering and indexing, among others. As such, automatic keyphrase extraction has garnered attention and become a focal point for many researchers (Kim et al., 2010). Usually, the most common scenario of keyphrase extraction is to identify the keyphrases over the whole scien- tific document. Existing techniques in aforesaid setups use elaborate, hand-crafted features fed for selected candidate keyphrases to machine learn- ing models such as support vector machines and multilayer perceptrons, to learn keyphrases (Kim et al., 2013). The scope of features vary from sim- ple, surface-level features like character n-grams, token type, and part-of-speech tags – to features drawn from global statistics and lexical semantics, such as TF-IDF, keywordness, relation to docu- ment’s logical structure (Nguyen and Kan, 2007). However, the given task setup is inherently dif- ferent as it requires to identify all the keyphrases of certain types (or classes – Material, Process and Task) over an excerpt of a scientific document. As inferred from our primary analysis of the training data, some of the crucial challenges of keyphrase extraction in this particular task setup are: • Keyphrases occur more densely in the excerpts compared against the standard keyphrase extraction task where systems typ- ically identify 5–25 keyphrases over an entire document; • Keyphrases overlap significantly. For exam- ple “equally sized blocks” and “blocks” both need to be extracted as keyphrases of type Materials; • Determining the keyphrase type depends on the context. For example “oxidation test” and “assessment of …

Content truncated to 500 words