Spoken Language Processing for Clinical Conversations

Prescriptions and medication for diabetes costs the NHS approximately £1bn per year, but it is estimated that the cost of (predominantly preventable) complications of diabetes is closer to £10bn per year. Self-management of diet and medication is an important part of preventing those complications. Motivational Interviews are a core method by which patients are guided through the change in mindset necessary to manage their own condition. The effectiveness of those interviews is largely dependent on the quality of the guidance by the clinician and the post-interview assessment by the clinician; this assessment is time consuming and therefore expensive. This assessment is broken into two parts: assessment of the linguistic content of the interview (an analysis of the words spoken by the patient and clinician); and the paralinguistic content of the interview (an analysis of the empathy and engagement between the patient and clinician). Research is already being conducted by this group regarding the linguistic features. This project will focus on the paralinguistic features.

This project aims to develop tools to help automatically assess the qualities of these interviews, by measuring levels of empathy and engagement in the patient, as well as assessing the quality of the guidance provided by the clinician, to provide insight as to how they might be more effective.

The key objectives of research for this project are as follows: Automatic sentiment and emotion analysis of conversational speech; automatic empathy detection in 2-party conversations; and overcoming problems of dealing with small amount of annotated data.

The methodology to be employed will focus on a data-driven approach, considering that only little amount, and possibly of a gross-level, annotations are available with the training data. The project work is expected to base on techniques used in state-of-the art spoken language processing research, such as i-vectors, unsupervised clustering, latent Dirichlet allocation (LDA), Deep Neural Networks (DNNs), and Hidden Markov Models (HMMs). The novelty is expected in adoption of these techniques and their modifications for the specific needs of the project application.

While this project is being designed around a healthcare application, it should be applicable as a general case solution for measuring sentimental features in any 2 person conversation.