Abstract
Despite the rapid growth of user-generated unstructured text from online group discussions, business decision-makers are facing the challenge of understanding its highly incoherent content. Coherence analysis attempts to reconstruct the order of discussion messages. However, existing methods only focus on system and cohesion features. While they work with asynchronous discussions, they fail with synchronous discussions because these features rarely appear. We believe that discussion logic features play an important role in coherence analysis. Therefore, we propose a TCA method for coherence analysis, which is composed of a novel message similarity measure algorithm, a subtopic segmentation algorithm and a TBL-based classification algorithm. System, cohesion and discussion logic features are all incorporated into our TCA method. Results from experiments showed that the TCA method achieved significantly better performance than existing methods. Furthermore, we illustrate that the DATree generated by the TCA method can enhance decision-makers’ content analysis capability.
Recommended Citation
Deng, Shasha; Zhang, Pengzhu; and Zhou, Yilu, "Turning Unstructured and Incoherent Group Discussion into DATree: A TBL Coherence Analysis Approach" (2011). ICIS 2011 Proceedings. 22.
https://aisel.aisnet.org/icis2011/proceedings/knowledge/22
Turning Unstructured and Incoherent Group Discussion into DATree: A TBL Coherence Analysis Approach
Despite the rapid growth of user-generated unstructured text from online group discussions, business decision-makers are facing the challenge of understanding its highly incoherent content. Coherence analysis attempts to reconstruct the order of discussion messages. However, existing methods only focus on system and cohesion features. While they work with asynchronous discussions, they fail with synchronous discussions because these features rarely appear. We believe that discussion logic features play an important role in coherence analysis. Therefore, we propose a TCA method for coherence analysis, which is composed of a novel message similarity measure algorithm, a subtopic segmentation algorithm and a TBL-based classification algorithm. System, cohesion and discussion logic features are all incorporated into our TCA method. Results from experiments showed that the TCA method achieved significantly better performance than existing methods. Furthermore, we illustrate that the DATree generated by the TCA method can enhance decision-makers’ content analysis capability.