Automatic Coding of Therapist and Client Language in Motivational Interviewing to Predict Reductions in Alcohol Use and Problems Using Machine-based Dyadic Multimodal Representation Learning

尽管动机性面试（MI）被广泛使用，但其成功的潜在机制 仍然知之甚少[14]，特别是客户改变谈话和后续行为之间的联系 [14，69]。先前的研究已经确定了MI疗效的两种可能的活性成分： 涉及治疗师-来访者二元体的元素的关系成分，包括以下表达： 同理心，以及专注于客户行为的差异唤起的技术组件，例如 改变谈话或客户对他们改变承诺的说法[57]。目前对这些问题的分析 组件仅限于与语言有关的调查，并受到昂贵和 艰苦的手动编码，尽管花费了时间和精力来实现可靠性， 足够敏感或具体，以充分测试所支持的复杂理论命题， MI理论家 我们的项目将通过引入一种新的方法来解决当前MI编码系统的缺点。 利用我们最近在自动语言和非语言行为方面的进展的计算框架 分析以及多模态机器学习。我们的框架旨在共同分析口头（即， 正在说的话），非语言的（即，如何说），和二元（即，在什么样的人际环境中 说了些什么）行为，以更好地识别会话中的改变谈话并维持预测 会后饮酒我们将利用已经收集和注释的音频数据从两个NIAAA- 资助的单次MI随机临床试验，以改善饮酒行为（N=91; N=158）。我们 我们将通过广泛收集客户和二元行为来传播我们的发现， 客户端和二元行为数据库。此外，我们将验证我们的普遍性 计算框架，使用另外七个NIAAA和联邦资助的RCT，使用不同的 针对不同目标人群的MI方案。

Despite the widespread use of Motivational Interviewing (MI), the underlying mechanisms of its success are still poorly understood [14], especially the link between client change talk and subsequent behavior change [14,69]. Previous research has identified two possible active components underlying MI efficacy: a relational component involving elements of the therapist-client dyad including the expression of empathy, and a technical component focused on the differential evocation of client behaviors such as change talk or what a client says about their commitment to change [57]. Current analyses of these components are limited to investigations pertaining to language only and restricted by expensive and arduous manual coding which, despite the time and efforts expended to achieve reliability, may still not be sufficiently sensitive or specific to adequately test the complex theoretical propositions espoused by MI theorists. Our project will address shortcomings of current MI coding systems by introducing a novel computational framework that leverages our recent advances in automatic verbal and nonverbal behavior analyses as well as multimodal machine learning. Our framework aims to jointly analyze verbal (i.e., what is being said), nonverbal (i.e., how something is said), and dyadic (i.e., in what interpersonal context something is said) behavior to better identify in-session change talk and sustain talk that is predictive of post-session alcohol use. We will leverage already collected and annotated audio data from two NIAAA- funded single-session MI randomized clinical trials to improve drinking behavior (N=91; N=158). We will disseminate our findings through an extensive collection of client and dyadic behaviors through our proposed Client and Dyadic Behavior Databases. In addition, we will validate the generalizability of our computational framework using seven additional NIAAA- and federally funded RCTs that used different MI protocols for different target populations.