Entity Linking for Spoken Language

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 225–230, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Entity Linking for Spoken Language Adrian Benton Human Language Technology Center of Excellence Johns Hopkins University Baltimore, MD, USA adrian@jhu.edu Mark Dredze Human Language Technology Center of Excellence Johns Hopkins University Baltimore, MD, USA mdredze@jhu.edu Abstract Research on entity linking has considered a broad range of text, including newswire, blogs and web documents in multiple languages. However, the problem of entity linking for spoken language remains unexplored. Spo- ken language obtained from automatic speech recognition systems poses different types of challenges for entity linking; transcription er- rors can distort the context, and named entities tend to have high error rates. We propose fea- tures to mitigate these errors and evaluate the impact of ASR errors on entity linking using a new corpus of entity linked broadcast news transcripts. 1 Introduction Entity linking identifies for each textual mention of an entity a corresponding entry contained in a knowledge base, or indicates when no such entry exits (NIL). Numerous studies have explored en- tity linking in a wide range of domains, including newswire (Milne and Witten, 2008; Mcnamee et al., 2009; McNamee and Dang, 2009; Dredze et al., 2010), blog posts (Ji et al., 2010), web pages (De- martini et al., 2012; Lin et al., 2012), social media (Cassidy et al., 2012; Guo et al., 2013a; Shen et al., 2013; Liu et al., 2013), email (Gao et al., 2014) and multi-lingual documents (Mayfield et al., 2011; Mc- Namee et al., 2011; Wang et al., 2012). A common theme across all these settings requires addressing two difficulties in linking decisions: matching the textual name mention to the form contained in the knowledge base, and using contextual clues to dis- ambiguate similar entities. However, all of these studies have focused on written language, while linking of spoken language remains untested. Yet many intended applications of entity linking, such as supporting search (Hachey et al., 2013) and iden- tifying relevant sources for reports (He et al., 2010; He et al., 2011), linking of spoken language is crit- ical. Search results regularly include audio content (e.g. YouTube) and numerous information sources are audio recordings (e.g. media reports.) An evalu- ation of entity linking for spoken language can help clarify issues and challenges in this domain. In addition to the two main challenges discussed above, audio entity linking presents two parallel difficulties that arise from automatic transcription (ASR) of speech. First, the context can be both shorter (than newswire formats) and contain ASR errors, which can make the context of the mention less like supporting material in the knowledge base. Second, named entities are often more difficult to recognize (Huang, 2005; Horlock and King, 2003); they are often out-of-vocabulary and less common overall in training data. This can mislead the name matching techniques on which most entity linking systems depend. In this paper we consider the task …

Content truncated to 500 words