Predatory chatlog messages as forensic evidence in court: A comparison of two different procedures for estimating the weight of evidence

Abstract

The performances of two different procedures for calculating the likelihood ratio (LR) for forensic text comparison (FTC) are empirically compared. One is the multivariate kernel density (MVKD) procedure with so-called lexical features. The MVKD procedure has been successfully applied to various types of evidence, including texts. The other is the procedure based on character N-grams. N-gram is a widely-used, robust probabilistic language model. The effectiveness of character N-grams has been reported in authorship analysis, however, to the best of my knowledge, it has not yet been applied to LR-based FTC. In this study, the log-likelihood-ratio-cost (Cllr), which is an appropriate assessment metric for LR-based systems, is used to assess the performance of the two procedures. It will be reported that the MVKD procedure outperforms the character N-gram procedure. Through the comparison of the two procedures, this study also demonstrates how the weight of evidence (= LR) can be estimated from text messages.