Recent efforts have shown that neural text processing models are vulnerable to adversarial examples, but the nature of these examples is poorly understood. In this work, we show that adversarial attacks against CNN, LSTM and Transformer-based classification models perform word substitutions that are identifiable through frequency differences between replaced words and their corresponding substitutions. Based on these findings, we propose frequency-guided word substitutions (FGWS), a simple algorithm exploiting the frequency properties of adversarial word substitutions for the detection of adversarial examples. FGWS achieves strong performance by accurately detecting adversarial examples on the SST-2 and IMDb sentiment datasets, with F 1 detection scores of up to 91.4% against RoBERTa-based classification models. We compare our approach against a recently proposed perturbation discrimination framework and show that we outperform it by up to 13.0% F 1 .

There is an increasing demand for automated verbal deception detection systems. We propose named entity recognition (NER; i.e., the automatic identification and extraction of information from text) to model three established theoretical principles: (i) truth tellers provide accounts that are richer in detail, (ii) contain more contextual references (specific persons, locations, and times), and (iii) deceivers tend to withhold potentially checkable information. We test whether NER captures these theoretical concepts and can automatically identify truthful versus deceptive hotel reviews. We extracted the proportion of named entities with two NER tools (spaCy and Stanford's NER) and compared the discriminative ability to a lexicon word count approach (LIWC) and a measure of sentence specificity (speciteller). Named entities discriminated truthful from deceptive hotel reviews above chance level, and outperformed the lexicon approach and sentence specificity. This investigation suggests that named entities may be a useful addition to existing automated verbal deception detection approaches.

No abstract

This paper introduces the Grievance Dictionary, a psycholinguistic dictionary that can be used to automatically understand language use in the context of grievance-fueled violence threat assessment. We describe the development of the dictionary, which was informed by suggestions from experienced threat assessment practitioners. These suggestions and subsequent human and computational word list generation resulted in a dictionary of 20,502 words annotated by 2318 participants. The dictionary was validated by applying it to texts written by violent and non-violent individuals, showing strong evidence for a difference between populations in several dictionary categories. Further classification tasks showed promising performance, but future improvements are still needed. Finally, we provide instructions and suggestions for the use of the Grievance Dictionary by security professionals and (violence) researchers.