Game-Theoretic Modeling of Semantic Interpretation in Cross-Linguistic Translation

Abstract

Translation is about more than just choosing a word that has a similar dictionary definition; it's about predicting what will make sense to a reader given a target expression. This paper plays out that interaction as a signaling game between an informed translator and a reader in the target language. The translator is observing a meaning he wishes to convey and selects an expression that is short and has multiple interpretations or a more explicit expression that specifies one interpretation. The reader puts together the expression and guesses the translator's decisions. If utilities give them credit for a correct inference, the translator has to expend an extra communication investment on being explicit with the word. Bayesian updating and sequential best responses are used to analyze a two-meaning model. If it is cheaper to clarify than to be successful in interpretation, then there is a separating equilibrium, where the most likely meaning is expressed in the shorter phrase and the less likely meaning is expressed in the longer phrase. Based on this convention, there is an expected semantic recovery probability of 1.00 and expected utility of the translator of 0.94 for the made-up example with prior probabilities of 0.70 and 0.30 and cost of clarification 0.20.If the two meanings are both expressed explicitly, the utility is 0.80; if the two meanings are both expressed on a short statement, the utility is 0.70, which is less. These values are based on the assumptions of payoffs and shared conventions, and not empirical translation scores. The results of the paper are related to the concept of ambiguity preservation, audience expectations, and contextual error analysis. Future studies should quantify beliefs and communication cost among readers as well as assess the learning and retention of proposed conventions across language pairs.