No abstract

No abstract

Word embeddings carry stereotypical connotations from the text they are trained on, which can lead to invalid inferences in downstream models that rely on them. We use this observation to design a mechanism for measuring stereotypes using the task of natural language inference. We demonstrate a reduction in invalid inferences via bias mitigation strategies on static word embeddings (GloVe). Further, we show that for gender bias, these techniques extend to contextualized embeddings when applied selectively only to the static components of contextualized embeddings (ELMo, BERT).

Machine reading calls for programs that read and understand text, but most current work only attempts to extract facts from redundant web-scale corpora. In this paper, we focus on a new reading comprehension task that requires complex reasoning over a single document. The input is a paragraph describing a biological process, and the goal is to answer questions that require an understanding of the relations between entities and events in the process. To answer the questions, we first predict a rich structure representing the process in the paragraph. Then, we map the question to a formal query, which is executed against the predicted structure. We demonstrate that answering questions via predicted structures substantially improves accuracy over baselines that use shallower representations.

Motivational Interviewing (MI) is an efficacious treatment for substance use disorders and other problem behaviors. Studies on MI fidelity and mechanisms of change typically use human raters to code therapy sessions, which requires considerable time, training, and financial costs. Natural language processing techniques have recently been utilized for coding MI sessions using machine learning techniques, rather than human coders, and preliminary results have suggested these methods hold promise. The current study extends this previous work by introducing two natural language processing models for automatically coding MI sessions via computer. The two models differ in the way they semantically represent session content, utilizing either 1) simple discrete sentence features (DSF model) and 2) more complex recursive neural networks (RNN model). Utterance- and session-level predictions from these models were compared to ratings provided by human coders using a large sample of MI sessions (N = 341 sessions; 78,977 clinician and client talk turns) from 6 MI studies. Results show that the DSF model generally had slightly better performance compared to the RNN model. The DSF model had “good” or higher utterance-level agreement with human coders (Cohen’s kappa > 0.60) for open and closed questions, affirm, giving information, and follow/neutral (all therapist codes); considerably higher agreement was obtained for session-level indices, and many estimates were competitive with human-to-human agreement. However, there was poor agreement for client change talk, client sustain talk, and therapist MI-inconsistent behaviors. Natural language processing methods provide accurate representations of human derived behavioral codes and could offer substantial improvements to the efficiency and scale in which MI mechanisms of change research and fidelity monitoring are conducted.