Varying Vector Representations and Integrating Meaning Shifts into a PageRank Model for Automatic Term Extraction

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4388–4394 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4388 Varying Vector Representations and Integrating Meaning Shifts into a PageRank Model for Automatic Term Extraction Anurag Nigam1, Anna Hätty1,2, Sabine Schulte im Walde1 1Institute for Natural Language Processing, University of Stuttgart, Germany 2Robert Bosch GmbH, Corporate Research, Renningen, Germany {anurag.nigam, schulte}@ims.uni-stuttgart.de, anna.haetty@de.bosch.com Abstract We perform a comparative study for automatic term extraction from domain-specific language using a PageRank model with different edge-weighting methods. We vary vector space representations within the PageRank graph algorithm, and we go beyond standard co-occurrence and investigate the influence of measures of association strength and first- vs. second-order co-occurrence. In addition, we incorporate meaning shifts from general to domain-specific language as personalized vectors, in order to distinguish between termhood strengths of ambiguous words across word senses. Our study is performed for two domain-specific English corpora: ACL and do-it-yourself (DIY); and a domain-specific German corpus: cooking. The models are assessed by applying average precision and the roc score as evaluation metrices. Keywords: Term Extraction, PageRank Algorithm, Vector Representations, Meaning Shifts 1. Introduction Terms are linguistic expressions which characterize a do- main, i.e., words and phrases which are typical of a docu- ment or a corpus from a specific domain (in contrast to gen- eral language usage). The automatic recognition of terms represents an important basis for further Natural Language Processing (NLP) tasks, such as thesaurus creation, auto- matic translation, and, in general, for domain knowledge acquisition and comprehension. Approaches for automatic term extraction can broadly be classified into four cate- gories: linguistic (Justeson and Katz, 1995; Basili et al., 1997), statistical (Schäfer et al., 2015), hybrid (Frantzi et al., 1998; Maynard and Ananiadou, 1999) and machine- learning approaches (da Silva Conrado et al., 2013). Re- cently, word vector and deep learning approaches have emerged (Zadeh and Handschuh, 2014b; Amjadian et al., 2016; Wang et al., 2016). A niche of approaches is represented by graph-based sta- tistical methods. In particular, we only find a handful of ideas for how to exploit the PageRank algorithm for auto- matic term extraction. The underlying motivation is that graph-based ranking algorithms –relying on graphs whose vertices represent the vocabulary of a corpus and whose edges represent some relationship between the words in the vocabulary– have the potential to inform about the relative importance of vertices in a graph, and that important ver- tices are likely to represent terms. For example, Mihal- cea and Tarau (2004) used co-occurrence counts as edge weights in a graph; Khan et al. (2016) used embedding similarity for the same purpose. Zhanga et al. (2017) incor- porated semantic relatedness via a personalized PageRank algorithm for extracting terms from the corpus. In this study, we present a comparison of vector space representations within a PageRank graph algorithm for automatic term extraction. We go beyond standard co- occurrence and investigate the influence of measures of as- sociation strength (Evert, 2005) and first- …

Content truncated to 500 words