Reflection #7 – [02/13] – Jiameng Pu

Niculae, V., Kumar, S., Boyd-Graber, J., & Danescu-Niculescu-Mizil, C. (2015). Linguistic harbingers of betrayal: A case study on an online strategy game. arXiv preprint arXiv:1506.04744.


The paper explores linguistic cues that indicate fickle interpersonal relations, like close friends becoming enemies. Since data that define the relationship between friends or enemies are not extensively accessible, researchers turn to a war-themed strategy game in which friendships and betrayals are orchestrated primarily through language. By studying dyadic interactions in the game and analyzing languages under cases that players form alliances and betray each other, they characterize subtle signs of imminent betrayal in players’ conversation and examine temporal patterns that foretell betrayal. From conversation scenarios in Diplomacy (the war-themed strategy game), we can actually see that betrayer would unconsciously reveal their planned treachery, meanwhile, the eventual victim can rarely be able to notice these signals.
They find that if the balance of conversational attributes like positive sentiment, politeness, and structured discourse shows sudden changes, imminent betrayal would happen in the future conversation. Then researchers provide a framework for analyzing communication patterns and explore linguistic features that are predictive of whether friendships will end in betrayal. They also discuss how to generalize methods to other domain and how automatically predicting relationships between people can help advance the study of trust and relationships using computational linguistics.


“Despite people’s best effort to hide it, the intention to betray can leak through the language one uses.” This reminds me of another idea that we may use the same strategy to detect people’s relationship pattern. Under normal circumstances, there are different relationship patterns when people getting along with each other. For example, some are balanced relationships like friends, colleagues, and relatives, while others are unbalanced relationships like leaders and subordinates, professors and students. Through the conversation content between people, we can extract the linguistic features following the same direction such as sentiment, argumentation, and discourse, politeness, and talkativeness, to predict the possible relationship between people. By analyzing the patterns of interpersonal relationships, we can have a deeper understanding of the current status of people’s life or whether the patterns will change over time, which is a more macro-sociological issue to figure out.
Apart from detecting the relationship pattern, we may use specific semantic features to study much other potential information hidden in the conversation, such as trust, familiarity, and intimacy between people.
Since the intuition of this paper is that a stable relationship should be balanced, it makes sense that all the predictions of Betrayal in the paper is based on signal an imbalance in the communication patterns of the dyad. However, my concern is whether these mentioned semantic features could provide a complete and efficient predictive analysis. Are there other available properties? e.g., Humor, straightforwardness. Or, in addition to detecting the imbalance of both sides in conversation, we may analyze the change of speech mode of the betrayer according to the timeline, i.e., imbalance of speech pattern before and after the decision to betray. This can solve difficult problems, for example, some people in nature are not as polite as the other party of the conversation.
In addition to the logistic regression which is often used for binary classification, support vector machine(SVM) is another classic algorithm that can be used for classification. As they have different advantages, we can design a control group in the experiment to choose the best classifier. Similarly, the semantic features can also be experimented in a controlled manner, so as to select the optimal combination of linguistic features that are most efficient for predicting betrayal.

Leave a Reply

Your email address will not be published.