Polimatrix: Conversation Analysis as a Method for Evaluating Performance in Matrix Games

This paper analyses the spoken interactions that occur during matrix games, serious role-playing exercises used in military and strategic training, and argues that linguistic performance should be incorporated into how these games are evaluated. The authors built a Russia-Ukraine negotiation game and recorded nine sessions, five face-to-face with novice master's students at a Dutch university and four online with mostly expert volunteers, producing a corpus of roughly 103,600 tokens. Analysis of keywords, n-grams, and speech functions shows that novice players tend to speak in a direct, informal register focused on immediate gameplay, with markers such as expletives, laughter, and "I don't want to," while expert players use more indirect, hesitant, and deliberate phrasing, such as "I think" constructions and first-name address, and reflect on strategy beyond the game.

The study frames negotiation exchanges through conversation analysis, using adjacency pairs such as negotiation open, probe, and response, and identifies meta-level moves including meta-critique, meta-strategising, and meta-roleplay. Two short excerpts illustrate how a negotiation cycle unfolds and how players use game-state references to help a counterpart save face. The authors note a gap between participants' recollections of conversations and what was actually said, which they cite as a reason to keep transcripts and analyse them directly.

Drawing on prior research linking discourse features to negotiation outcomes, the authors propose a three-step, scripted debriefing method: obtaining a transcript, selecting a significant extract, and applying a structured set of questions about experience, observations, conclusions, and reflectivity. The proposal has not yet been tested with participants, and the authors plan trials in two countries that separate expertise from mode of play. The paper's findings rest on a small sample, and the correlations between linguistic features and contextual variables are presented as preliminary.