A Generative Joint, Additive, Sequential Model of Topics and Speech Acts in Patient-Doctor Communication

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1765–1775, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics A Generative Joint, Additive, Sequential Model of Topics and Speech Acts in Patient-Doctor Communication Byron C. Wallace†, Thomas A. Trikalinos†, M. Barton Laws†, Ira B. Wilson† and Eugene Charniak‡ †Dept. of Health Services, Policy & Practice, Brown University, Providence, RI ‡Dept. of Computer Science, Brown University, Providence, RI {byron wallace, thomas trikalinos, michael barton laws ira wilson, eugene charniak}@brown.edu Abstract We develop a novel generative model of con- versation that jointly captures both the top- ical content and the speech act type asso- ciated with each utterance. Our model ex- presses both token emission and state tran- sition probabilities as log-linear functions of separate components corresponding to topics and speech acts (and their interactions). We apply this model to a dataset comprising anno- tated patient-physician visits and show that the proposed joint approach outperforms a base- line univariate model. 1 Introduction Communication involves at least two aspects: the words one says and the acts one performs in saying them. Examples of the latter include asking ques- tions, issuing commands, and so on. These are re- ferred to as speech acts under the sociolinguistic the- ory of Austin (1955), which was further developed by Searle (1969; 1985). Recognizing speech acts is crucial to understanding communication because a speaker’s meaning is only partially captured by the words they use; much of their intent is expressed im- plicitly via speech acts (Searle, 1969). On this view, conversational utterances can be as- signed both a topic and a speech act. The former describes the subject matter of what was said and the latter captures the “social act” (e.g., promising) performed by saying it. For example, the utterance “Obama won the election” is topically political and is an example of an information giving speech act. “Did Obama win the election?”, meanwhile, belongs Role Utterance Topic Speech act D Let me just write down some of these issues here so I get them straight in my mind. Logistics Commissive P Doctor you ain’t got to tell me nuttin’. Socializing Directive P I’m in very good hands when I’m around you. Socializing Give Info. P If push comes to a shove, you open the window and throw me out. Socializing Humor/Levity D I wanted to ask you, too - Biomedical Conv. Mgmt. D you know you had that colonic polyp - Biomedical Ask Q. D - is it two years from now that they’re going to be doing the repeat? Biomedical Ask Q. P Yeah. Biomedical Conv. Mgmt. D We’ll do the repeat coloscopy in about two years. Biomedical Give Info. Table 1: An excerpt from a patient-doctor interaction, annotated with topic and speech act codes. The D and P roles denote doctor and patient, respectively. Conv. Mgmt. abbreviates conversation management; Ask Q. ab- breviates ask question. to the same topic but is a question. Both aspects are necessary to understand conversation. …

Content truncated to 500 words