LIPN at SemEval-2017 Task 10: Filtering Candidate Keyphrases from Scientific Publications with Part-of-Speech Tag Sequences to Train a Sequence Labeling Model

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 995–999, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics LIPN at SemEval-2017 Task 10: Filtering Candidate Keyphrases from Scientific Publications with Part-of-Speech Tag Sequences to Train a Sequence Labeling Model Simon David Hernandez, Davide Buscaldi, Thierry Charnois Laboratoire d’Informatique de Paris Nord, CNRS (UMR 7030) Universit´e Paris 13, Sorbonne Paris Cit´e, Villetaneuse, France {hernandez-perez,buscaldi,thierry.charnois}@lipn.univ-paris13.fr Abstract This paper describes the system used by the team LIPN in SemEval 2017 Task 10: Extracting Keyphrases and Rela- tions from Scientific Publications. The team participated in Scenario 1, that in- cludes three subtasks, Identification of keyphrases (Subtask A), Classification of identified keyphrases (Subtask B) and Ex- traction of relationships between two iden- tified keyphrases (Subtask C). The pre- sented system was mainly focused on the use of part-of-speech tag sequences to fil- ter candidate keyphrases for Subtask A. Subtasks A and B were addressed as a sequence labeling problem using Condi- tional Random Fields (CRFs) and even though Subtask C was out of the scope of this approach, one rule was included to identify synonyms. 1 Introduction Identifying candidate keyphrases in texts is com- monly a first step in systems for keyphrase extrac- tion (Hasan and Ng, 2014; Haddoud et al., 2015), this can be done by filtering words or phrases from documents using heuristics to determine which can be candidate keyphrases. The system uses sequences of part-of-speech tags (PoS sequences) as patterns to filter candi- date keyphrases. These candidates are used to train two Conditional Random Field (CRF) mod- els, one for keyphrase identification and other for keyphrase classification. CRF was trained with or- thographic features, additionally to features from WordNet and titles from academic papers. The PoS sequences were extracted from the annotated keyphrases in the corpus provided for the task (Augenstein et al., 2017). The PoS sequences used in this system are de- scribed in Section 2, there is an explanation of how they were used and how they were extracted from the training data. In Section 3 is detailed how CRF was trained with the candidate keyphrases and in Section 4 the features are described. In Section 5 there is an explanation of how CRF was applied to identify and classify keyphrases. Section 6 shows the post-processing steps and Section 7 introduces some experiments. 2 PoS sequences In this paper, we use the term PoS sequences to refer to sequences of part-of-speech tags. PoS sequences are used in automatic keyphrase ex- traction as features (Kim and Kan, 2009; Hasan and Ng, 2014) or to filter candidate keyphrases (Kim and Kan, 2009; Haddoud et al., 2015; Hasan and Ng, 2014), for example, with small sets of patterns matching all noun phrases and preposi- tional phrases, avoiding patterns that increase er- ror, like sequences containing adverbs (Kim and Kan, 2009). In this system, PoS sequences are used only to filter candidate keyphrases. From the annotated keyphrases in the training data, were extracted 1445 different PoS sequences 1, …

Content truncated to 500 words