Abstract: The problem of keyword extraction from conversations is with the intention to find the potentially relevant documents using the retrieval of keywords for each short conversational fragment then which are going to be recommended to participants. Short conversational fragment contains variety of words which are related to several topics; it is difficult to find out the information needed by the participants. To solve this, an algorithm to extract keywords from output of an ASR system is introduced which uses topic modeling techniques and sub modular reward function which favors diversity in keyword sets to reduce ASR noise. This method derives multiple topically separated queries from the keyword set which helps in maximizing the chances of, at least one relevant recommendation while searching these queries over English Wikipedia.

Keywords: Document recommendation, keyword extraction, topic modeling.