Evaluating anaphora and coreference resolution to improve automatic keyphrase extraction

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 804–814, Osaka, Japan, December 11-17 2016. Evaluating anaphora and coreference resolution to improve automatic keyphrase extraction Marco Basaldella, Giorgia Chiaradia and Carlo Tasso {basaldella.marco.1, chiaradia.giorgia}@spes.uniud.it carlo.tasso@uniud.it Artificial Intelligence Laboratory Universit`a degli Studi di Udine Via delle Scienze 208, Udine, Italy Abstract In this paper we analyze the effectiveness of using linguistic knowledge from coreference and anaphora resolution for improving the performance for supervised keyphrase extraction. In order to verify the impact of these features, we define a baseline keyphrase extraction system and evaluate its performance on a standard dataset using different machine learning algorithms. Then, we consider new sets of features by adding combinations of the linguistic features we propose and we evaluate the new performance of the system. We also use anaphora and coreference resolution to transform the documents, trying to simulate the cohesion process performed by the human mind. We found that our approach has a slightly positive impact on the performance of automatic keyphrase extraction, in particular when considering the ranking of the results. 1 Introduction Automatic Keyphrase Extraction (henceforth AKE), i.e. the task of extracting a list of phrases of one or more words “that capture the main topics discussed in a given document” (Turney, 2000) is a natural language processing (herein NLP) task which received widespread attention in the last years, with appli- cations, e.g., in the fields of digital libraries (Gutwin et al., 1999) or community modelling (De Nart et al., 2015). Many AKE algorithms have been developed, which can be roughly divided into two categories (Hasan and Ng, 2014): • Supervised algorithms: after the generation of candidate keyphrases (henceforth KPs) by means of linguistic knowledge, these candidates are associated to a set of features such as TF-IDF, position in the text, and so on; then, a supervised machine learning (herein ML) algorithm learns over a training set how to decide if a candidate is a suitable KP or not. • Unsupervised algorithms: for example, the document is represented using a graph structure, whose nodes are candidate KPs. Then, the popularity of each candidate is evaluated using graph algorithms usually derived from the PageRank algorithm (Mihalcea and Tarau, 2004; Wan and Xiao, 2008). Other approaches include for example clustering-based algorithms, such as the one presented in (Liu et al., 2009), or techniques which rely on building a statistical language model to rank KPs, like the one presented in (Tomokiyo and Hurst, 2003). However, the performance of the state of the art systems is still much lower than many other NLP tasks. An idea of the current top performing systems can be obtained by looking at the results of “SE- MEVAL 2010 Task 5: Automatic Keyphrase Extraction from Scientific Articles” (Kim et al., 2010), where systems were ranked by F-score on the top 15 extracted keyphrases. The best system, presented by (Lopez and Romary, 2010), achieved a score of 27.5%. After SEMEVAL 2010, many systems tried to improve this level of …

Content truncated to 500 words