Speaker Identification Function Words and Beyond说话人识别的功能的话，超越.pptVIP

SpeakerIdentification: Function Words and Beyond Peter W.H. Smith and Gea de Jong City University, London, UK2 July 2005Presentation at IAFL2005, Cardiff, UK. Rationale for the Experiment Large police operations using audio surveillance. Recent developments in the UK: transcripts of both reference and disputed material more readily available. Complementary approach to forensic phonetics. Stylometric Analysis A set of methods for author identification. Largely used on volume data. Employs statistical techniques on features within the texts. Used in cases of disputed or unknown authorship. Stylometric Analysis - Methods Methods include: Simple methods – word length distributions, sentence lengths, letter collocations. Markov Models based on letter collocations (Khmelev and Tweedie 2001). Non-linear Neural Network models (Tweedie, Singh and Holmes 1996) Function word methods, using multivariate statistics, e.g. principal components analysis or discriminant analysis (W. Smith 1986, Binongo 1997, Smith and Smart 2001) Function Word Approach Typical Method Select texts Break texts into blocks Count specific function words within each block Perform a multivariate analysis on the resultant data. Present results Function Word Approach to Stylometric Analysis Some very successful results, e.g. Civil War Letters (Holmes, Gordon and Wilson 2001) De Doctrina Christiana and Milton (Tweedie et al. 1998) Consolatio – falsely attributed to Cicero (Forsyth et al.1999) The function word approach appears to work, i.e. it can distinguish authors with a claim to some reliability. Recent Study using Function Words 7 Texts by Author A, 4 Texts by author B and one disputed text. A and B are 18 century American political writers – large volume of available written material. Compare A with B A1,A2,B1 A1,A2,B2 …A1,A3,B1 (separates A from B in 69/84 cases) 40/42 cases. Compare B with A A1,B1,B2 A2,B1,B2 … (separates A from B in 40/42 cases) Function Words Appear