Mining Text Conversations - Methods for Mining and Summarizing Text Conversations - page 52

u - 1

c u

T u

D

z w

) j K

N u T

U

u + 1

Figure 3.5: Graphical model for a variation of LDA that also performs topic modeling. Additions and
changes to the standard LDA model are highlighted in black.

Once the model is learned (in an unsupervised fashion) and it is applied to a conversation,
the values predicted for all the c u variables will provide the topic segmentation for the conversation,
with a topic shift every time c u = 1. Since the model is an extensions of LDA, it will also provide
labels for the topics. When tested on the ICSI corpus, this extension of LDA performs similarly
to LCSeg for topic segmentation, with the additional advantage of providing topic labels that were
found intuitively informative by the author and semantically coherent by seven independent human
judges.
Other variations of LDA have been recently explored for modeling topics in meeting
transcripts. For more information, the interested readers can refer to Huang and Renals [ 2008 ]
and Georgescul et al. [ 2008 ].

Topic Modeling for Microblogs and Email Although most of the approaches to topic model-
ing of text conversations have been so far developed and tested on meeting transcripts, a few re-
searchers have begun to investigate their application to other conversational modalities. For example,
Ramage et al. [ 2010 ] study topic modeling for microblogs such as Twitter. In particular, they show

microblogs

how a semi-supervised variation of LDA, called labeled LDA, can be effectively applied to Twitter Labeled

LDA

conversations, in spite of the very limited length of Twitter posts - 140 characters or less. Labeled
LDA [ Ramage et al. , 2009 ] is a generalization of LDA that can be effectively applied in domains
in which there is prior knowledge on the topical structure of the documents. In essence, if there are
some topics that one cares about, for instance if the documents are already annotated with some
topic labels, then it is possible for the LDA to only use those when it learns the topic model.