Predicting correctness of eyewitness statements using the Semantic Evaluation Method (SEM)

Evaluating the correctness of eyewitness statements is one of the biggest challenges for the legal system, and this task is currently typically performed by human evaluations. Here we study whether a computational method could be applied to discriminate between correct and incorrect statements. The Semantic Evaluation Method (SEM) is based on Latent Semantic Analysis (LSA, Landauer & Dumais, 1997), - a method for automatically generating high dimensional semantic representations of words and sentences. The verbal data was extracted from the recorded narratives from a prior eyewitness study investigating the role of repeated retellings on subsequent recall accuracy and confidence (Sarwar, Allwood, & Innes-Ker, 2011). Participants watched a film of a kidnapping and then either retold the events to a single listener, or discussed the content with a confederate at five separate times over a 20-day period. Their subsequent written recall was analyzed using the Semantic Evaluation Method (SEM). The results show that accuracy can be predicted from quantification of the semantic content of eyewitness memory reports using SEM. This result also held true when data was separated into three distinct categories and the SEM was trained and tested on different categories of data.