CRII: RI: Alignment in Web-Forum Discourse: Computational Models of Adaptation and Language Change

Language use in real-world dialogue happens in context. Linguistic choices depend on previous ones: for example, the chosen words and sentence structures tend to mirror what was used previously by a conversation partner. This subtle adaptation process has been called 'alignment'. Alignment appears to help people understand each other in dialogue, and it seems to extend to human-computer interfaces, too. The concrete functions of alignment in dialogue are, however, unclear. Is it merely a useful epiphenomenon of how human memory works? Does it serve as a social or communicative signal? Is it indicative of a person's empathy? Does it help communities find a common language over long periods of time? Recent work has established that one of the consequences of alignment is persistent language change in the individual. There also is preliminary evidence that over time, groups of people talking to one another will converge in their choice of words and sentence structure. In other words, they find a common language. The project will devise computational models that describe and quantify these processes. With these, one can detect them in actual language use, such as in web-forums. In fact, the project will use big datasets from decades of web-forum messages to produce those models. The computational models will explain and predict processes in a way that makes them exploitable in modern social networks as well as for data science. Consider the example of a web-forum that connects those suffering from a disease so they can lend each other emotional and informational support. The models can detect and predict which messages in this web-forum are most supportive on the intended level, and whether they align to the person asking a question. A possible application of this may improve web-forum discourse by prioritizing search results and by making reading suggestions. Alignment models may also improve analysis techniques for large datasets by spotting networks of mutual supporters.

Models will be created in order to describe and explain alignment and language change in natural-language dialogue. The models will be computational and statistical to allow for exploitation of interactive alignment in natural-language dialogue as a feature in social network applications. Statistical alignment models describe language change over time as a function of variables that characterize the individual's behavior, memory, and of network information. These models will be fitted to longitudinal datasets derived from web-based, topic-oriented conversation threads. At the individual level, they will help refine cognitive-computational models of memory function in language production, which will be constrained by the well-validated ACT-R framework. The viability of the approach is supported by preliminary work on corpus-based syntactic priming and ACT-R models of language production, and pilot experiments showing alignment in the corpus. The outcomes of the project may point to novel methods of prioritizing and filtering the most helpful content and can address quality of life and well-being of patients such as those of the peer-support community whose conversations were studied in the investigator's work motivating the proposal.