A fused forensic text comparison system using lexical features, word and character N-grams

Abstract

This study investigates the degree that the performance of a likelihood ratio (LR)-based forensic text comparison (FTC) system improves by using logistic-regression fusion on LRs that were separately estimated by three different procedures, involving lexical features, word-based N-grams and character-based N-grams. This study uses predatory chatlog messages. The number of words used for modelling each group of messages is 500 words. The performance of the FTC system is assessed in terms of its validity (= accuracy) and reliability (= precision) using the log-likelihood-ratio cost (C<inf>llr</inf>) and 95% credible intervals (CI), respectively. It is demonstrated that 1) out of the three procedures, the lexical features procedure performed best in terms of C<inf>llr</inf>; and that 2) the fused system outperformed all three of the single procedures. The C<inf>llr</inf> value of the fused system is better than that of the procedure with lexical features by a value of 0.14. It is also reported that the validity and reliability of a system is negatively correlated; the fused system that yielded the best result in terms of C<inf>llr</inf> has the worst CI value.