We present discussion mining as a preliminary study of knowledge discovery from discussion content of offline meetings. Our system generates minutes for such meetings semi-automatically and links them with audio-visual data of discussion scenes. Then, not only retrieval of the discussion content, but also we are pursuing the method of searching for a similar discussion to an ongoing discussion from the past ones, and the method of generation of an answer to a certain question based on the accumulated discussion content. In terms of mailing lists and online discussion systems such as bulletin board systems, various studies have been done. However, what we think is greatly different from the previous works is that ours includes face-to-face offline meetings. We analyze meetings from diversified perspectives using audio and visual information. We also developed a tool for semantic annotation on discussion content. We consider this research not just data mining but a kind of real-world human activity mining.