Learning Trading Negotiations Using Manually and Automatically Labelled Data

摘要: Strategic conversational agents often need to trade resources with their opponent conversants -- and trading strategically can lead better results. While rule-based or supervised be used for such a purpose, here we explore learning approach based on automatically labelled examples from human players automatic in the game of Settlers Catan. Our experiments are data collected text-based natural language. We compare performance Bayes Nets, Conditional Random Fields, Forests task ranking offers, trained both manually data. experimental results show that our best agent labels outperformed its counterpart manual (with moderate annotator agreement) terms (a) predicting negotiations better, (b) winning more games.