DATA ANALYSIS: LINGUISTICS


or defi nitional issues, frequently still corpus based. Kestemont, for instance, describes experiments using memory-based machine learning (MBML) to tackle the lemmatisation (that is, base form categorisation of component words) of all surviving Middle Dutch literary manuscripts dated before 1300 CE. As with all mediaeval documents, spelling varies greatly which makes this a challenging area to shift from human to machine mediation. Attempts to generate new spelling variants and to refi ne string distance ranking result in ‘a language-independent system that can “learn” intra-lemma spelling variation... of interest to other research domains’. A very different approach and application


can be seen in the medicophysical study of early speech recognition by, for example, Telkemeyer[3]


et al (see ‘Locating the building


blocks’). This puts aside study of language itself in favour of the mechanisms by which it is acquired, and has profound implications for everything from social communication to management of learning. Acquisition presupposes ability to acquire.


In language, this is assumed to be hard-wired into us, but the exact mix and interaction of nature and nurture are, as in so many areas, a matter of vigorous debate. An intensively analytic study[4]


run by Dan Dediu at the Max


Planck Institute for Linguistics earlier this year, however, illustrates the methods which can be brought to bear in teasing out at least some threads of that classic tangle. Using a variety of open source tools, including MrBayes and his own ‘written from scratch’ BayesLang, across millions of runs, Dedui investigates the likelihood of proposals that two genes associated with development and growth of the brain also appear to infl uence the cultural transmission of a particular language component: linguistic tone. The words, syllable and parts of speech


from my Hamlet days are only a small part (and a crude one, at that) of the diverse structural means by which natural languages signify content. Linguistic tone, in which voice pitch conveys semiotic or grammatic information, is one of the more subtle, and genetic bias would predict greater stability. Applied uses for computerised linguistic data analysis are many and divergent, with more being proposed all the time, but broad areas which receive considerable attention, however, are psychomedical, educational, social and forensic.


www.scientifi c-computing.com


The stability of typological features, from Dediu[4]


Forensic possibilities are particularly in the spotlight at the moment with the parallel growth in mass digital communication and fear of terrorism. While documented and quotable specifi cs are thin on the ground, there are numerous anecdotal examples of heavily computer analytic fi lters being used to profi le conversations within monitored telecommunications traffi c. Concentrating on form, composition and structure rather than content, linguistic analysis of such


data streams is a conceptual, but more sophisticated, descendant of keyword recognition approaches. Forensic approaches to organised and


corporate crime increasingly resemble those to terrorism. Nor is the focus of linguistic attention unidirectional. One study[5]


of language used in


police interrogation of suspects, for instance, examines ‘the way power relations are created and maintained by... recurring discourse markers’. The use of corpus analyses can (see ➤


Sing a song of statistics


As I write this, the British press are airing allegations that police offi cers inserted song titles into their evidence at an inquest into a fatal shooting. Such a game would cause considerable distress to relatives and, if substantiated, constitute contempt of court, but such allegations are very hard to prove. As Language Log and former Scientifi c Computing World contributor Ray Girvan[10] comments: ‘Song titles frequently use commonplace words and idiom... and the language of most speakers will contain words and phrases that match song titles.’ A fi rm of solicitors with a forensic statistician working on a different but analogous case, while unwilling to discuss specifi cs on the record, were happy to discuss the principles involved. Broadly speaking, the method in this case


would involve comparing text corpora. One would be a body of stored court evidence transcripts, selected to match as closely as possible to the profi le (background, profession, type of case, and so on) of the witness under review. Another would be a database subset of song titles within a time and genre frame matching those alleged to have been planted.


A third, in cases where it was available, might be transcripts of other recorded speech by the same witness to provide a reference control against which the specifi c testimony could be tested. In the case of a police offi cer, communication logs and recorded suspect interviews might be suitable sources allowing cross reference with similar control. Material from other offi cers would also provide triangulation. Though these differ in many ways from language in the more formal court setting, and from everyday usage[5]


, those


differences are likely to be classifi able and comparable across individuals.


Degrees of match fuzziness would also be


defi ned to refl ect those in the allegations – in the specifi c case mentioned, for example, the song title I’m Kicking Myself is alleged to match the words ‘I’m sort of kicking myself’ in testimony. Distribution of various statistical descriptors in the selected corpora would be compared along conventional signifi cance testing lines. The particular statistician working for this fi rm also uses various pseudospatial comparisons, including Voronoi tessellation using Systat.


SCIENTIFIC COMPUTING WORLD DECEMBER 2010/JANUARY 2011 9


Page 1  |  Page 2  |  Page 3  |  Page 4  |  Page 5  |  Page 6  |  Page 7  |  Page 8  |  Page 9  |  Page 10  |  Page 11  |  Page 12  |  Page 13  |  Page 14  |  Page 15  |  Page 16  |  Page 17  |  Page 18  |  Page 19  |  Page 20  |  Page 21  |  Page 22  |  Page 23  |  Page 24  |  Page 25  |  Page 26  |  Page 27  |  Page 28  |  Page 29  |  Page 30  |  Page 31  |  Page 32  |  Page 33  |  Page 34  |  Page 35  |  Page 36