Who Talks to Whom: Modeling Latent Structures in Dialogue Documents

Bailu Ding, Jiang-Ming Yang, Chong Wang, Rui Cai, Zhiwei Li, Lei Zhang Fudan University, [email protected] Microsoft Research Asia, {jmyang, ruicai, zli, leizhang}@microsoft.com Princeton University, [email protected] 1 Latent Structures of Dialogue Documents Various forms of data that consist of sequential messages abound in social networks, such as citations, mail lists, chats, and forum discussions. A sequence of messages can be seen as a dialogue. We call the correlation among messages, namely the ’who-talks-to-whom’ relationship, the dialogue structure. Discovering dialogue structure is important for further studies on user behaviors. In a dialogue, we call the first message the root message, and the member who posts the root message the dialogue starter. People interested in the root message ’talk’ to the dialogue starter by posting new messages. Figure 1 (left) shows a dialogue taken from a discussion forum Slashdot(http://slashdot.org). Once we discover the tree-structured reply relationship in the dialogue, as shown in the upper-right part of Figure 1, we can construct the ’who-talks-to-whom’ network as shown in the lower-right part of Figure 1. This network can be used in various applications, such as community identification, expert ranking and friend recommendation.