In this paper, we combine the discourse coherence principles of Elementary Discourse Unit segmentation and Rhetorical Structure Theory parsing to construct meaningful graph-based text representations. We then evaluate a Graph Convolutional Network and a Graph Attention Network on these representations. Our results establish a new benchmark in F1-score assessment for discourse coherence modelling while also showing that Graph Convolutional Network models are generally more computationally efficient and provide superior accuracy.

This paper exploits band-limited cepstral coefficients (BLCCs) in forensic voice comparison (FVC), with the primary aim of locating speaker-sensitive spectral regions. BLCCs are sub-band cepstral coefficients (CCs) which are easily obtained by a linear transformation of full-band CCs. The transformation gives the flexibility of selecting any sub-band region without the recurrent cost of spectral analyses. Using multi-band BLCCs obtained by sliding a 600-Hz sub-band every 400 Hz across the full [0-5kHz] range, FVC experiments were attempted using citation recordings of the 5 Japanese vowels from 297 adult-male, native speakers. The FVC results give locations and ranges for the most speaker-sensitive sub-bands, and show that combining 3-4 of these yields comparable FVC performance with full-band CCs. Owing to their ability to easily extract locally-encoded speaker information from full-band CCs, it can be conjectured that BLCCs have a significant role to play in the search for meaningful interpretations of the numerical outcome of forensic analyses.

This study is the first likelihood ratio (LR)-based forensic text comparison study in which each text is mapped onto an embedding vector using RoBERTa as the pre-trained model. The scores obtained with Cosine distance and probabilistic linear discriminant analysis (PLDA) were calibrated to LRs with logistic regression; the quality of the LRs was assessed by log LR cost (Cllr). Although the documents in the experiments were very short (maximum 100 words), the systems reached the Cllr values of 0.55595 and 0.71591 for the Cosine and PLDA systems, respectively. The effectiveness of deep-learning-based text representation is discussed by comparing the results of the current study to those of the previous studies of systems based on conventional feature engineering tested with longer documents.

This study investigates the robustness and stability of a likelihood ratio–based (LR-based) forensic text comparison (FTC) system against the size of background population data. Focus is centred on a score-based approach for estimating authorship LRs. Each document is represented with a bag-of-words model, and the Cosine distance is used as the score-generating function. A set of population data that differed in the number of scores was synthesised 20 times using the Monte-Carol simulation technique. The FTC system’s performance with different population sizes was evaluated by a gradient metric of the log–LR cost (Cllr). The experimental results revealed two outcomes: 1) that the score-based approach is rather robust against a small population size—in that, with the scores obtained from the 40 60 authors in the database, the stability and the performance of the system become fairly comparable to the system with a maximum number of authors (720); and 2) that poor performance in terms of Cllr, which occurred because of limited background population data, is largely due to poor calibration. The results also indicated that the score-based approach is more robust against data scarcity than the feature-based approach; however, this finding obliges further study.

Score- and feature-based methods are the two main ones for estimating a forensic likelihood ratio (LR) quantifying the strength of evidence. In this forensic text comparison (FTC) study, a score-based method using the Cosine distance is compared with a feature-based method built on a Poisson model with texts collected from 2,157 authors. Distance measures (e.g. Burrows’s Delta, Cosine distance) are a standard tool in authorship attribution studies. Thus, the implementation of a score-based method using a distance measure is naturally the first step for estimating LRs for textual evidence. However, textual data often violates the statistical assumptions underlying distance-based models. Furthermore, such models only assess the similarity, not the typicality, of the objects (i.e. documents) under comparison. A Poisson model is theoretically more appropriate than distance-based measures for authorship attribution, but it has never been tested with linguistic text evidence within the LR framework. The log-LR cost (Cllr) was used to assess the performance of the two methods. This study demonstrates that: (1) the feature-based method outperforms the score-based method by a Cllr value of ca. 0.09 under the best-performing settings and; (2) the performance of the feature-based method can be further improved by feature selection.

Among the more typical forensic voice comparison (FVC) approaches, the acoustic-phonetic statistical approach is suitable for text-dependent FVC, but it does not fully exploit available time-varying information of speech in its modelling. The automatic approach, on the other hand, essentially deals with text-independent cases, which means temporal information is not explicitly incorporated in the modelling. Text-dependent likelihood ratio (LR)-based FVC studies, in particular those that adopt the automatic approach, are few. This preliminary LR-based FVC study compares two statistical models, the Hidden Markov Model (HMM) and the Gaussian Mixture Model (GMM), for the calculation of forensic LRs using the same speech data. FVC experiments were carried out using different lengths of Japanese short words under a forensically realistic, but challenging condition: only two speech tokens for model training and LR estimation. Log-likelihood-ratio cost (Cllr) was used as the assessment metric. The study demonstrates that the HMM system constantly outperforms the GMM system in terms of average Cllr values. However, words longer than three mora are needed if the advantage of the HMM is to become evident. With a seven-mora word, for example, the HMM outperformed the GMM by a Cllr value of 0.073.