Collecting and Exploring Everyday Language for Predicting Psycholinguistic Properties of Words

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1669–1679, Osaka, Japan, December 11-17 2016. Collecting and Exploring Everyday Language for Predicting Psycholinguistic Properties of Words Gustavo Henrique Paetzold and Lucia Specia Department of Computer Science University of Sheffield, UK {g.h.paetzold,l.specia}@sheffield.ac.uk Abstract Exploring language usage through frequency analysis in large corpora is a defining feature in most recent work in corpus and computational linguistics. From a psycholinguistic perspective, however, the corpora used in these contributions are often not representative of language usage: they are either domain-specific, limited in size, or extracted from unreliable sources. In an effort to address this limitation, we introduce SubIMDB, a corpus of everyday language spoken text we created which contains over 225 million words. The corpus was extracted from 38,102 subtitles of family, comedy and children movies and series, and is the first sizeable structured corpus of subtitles made available. Our experiments show that word frequency norms extracted from this corpus are more effective than those from well-known norms such as Kucera-Francis, HAL and SUBTLEXus in predicting various psycholinguistic properties of words, such as lexical decision times, familiarity, age of acquisition and simplicity. We also provide evidence that contradict the long-standing assumption that the ideal size for a corpus can be determined solely based on how well its word frequencies correlate with lexical decision times. 1 Introduction Large corpora of text are certainly one of the most fundamental resources in the field of Computational Linguistics. In Psycholinguistics, it has been long established that word frequencies from corpora play a very important role in cognitive processes. Brysbaert and New (2009) points out that frequently oc- curring words are often much more easily perceived, recalled and associated than rare words (Balota and Chumbley, 1984; Rayner and Duffy, 1986). In Text Simplification, researchers have found a strong relationship between frequencies and word simplicity (Devlin and Tait, 1998). An inherent limitation of work based on word frequency analysis is that the type of resource used as a corpus is often built for a specific communication purpose, such as news (Burgess and Livesay, 1998). This is however not representative of everyday language usage, particularly from a psycholinguistic per- spective. The other extreme of the spectrum features resources compiled from user-generated content, such as micro-blogs. However, these resources often suffer from grammar errors and misspellings, ex- cessive use of acronyms and shortenings, partly due to the constrains of the publication means (e.g. limited number of characters) (Pak and Paroubek, 2010). This is particularly concerning given that previous research has shown that the source from which a corpus was extracted is one of its most important defining traits. For example, the experiments of Brysbaert and New (2009) and Shardlow (2013) reveal that frequencies from spoken text have a much stronger correlation with psycholinguistic word properties than those from other sources. Their findings greatly highlight the potential of spoken language text, but there are very few examples of resources of this kind available for English. SUBTLEXus is a …

Content truncated to 500 words