The rapidly expanding capacity of local storage such as hard disk recorders is expected to create the need for a mechanism which enables selective TV watching based on individual viewer's preference. This paper proposes a method for creating the "Attention Graph" which depicts the amount of viewers? attention generated by TV drama. The Attention Graph, generated from the dialogues described in Internet communities concerning TV drama, is structured data mapped along the time line coincident with the progress of the drama's scenario. Thus, the Attention Graph assists in specifying noteworthy zones from complete TV programs and provides viewers with hints to watch selective scenes from their favorite TV drama. In general, dialogue from Internet communities is often expressed in poor grammatical manner, therefore natural language processing is difficult to apply. In an attempt to create the Attention Graph, we propose a statistical analysis of Symbolic Words to overcome this issue. The experimental results show that the Attention Graph successfully depicts the viewers' attention in TV drama, and indexes the zones of their attention.