Topic Segmentation Model Focusing on Local Context

Topic segmentation is important in understanding scientific documents since it can not only provide better readability but also facilitate downstream tasks such as information retrieval and question answering by creating appropriate sections or paragraphs. In the topic segmentation task, topic coherence is critical in predicting segmentation boundaries. Most of the existing models have tried to exploit as many contexts as possible to extract useful topic-related information. However, additional context does not always bring promising results, because the local context between sentences becomes incoherent despite more sentences being supplemented. To alleviate this issue, we propose siamese sentence embedding layers which process two input sentences independently to get appropriate amount of information without being hampered by excessive information. Also, we adopt multi-task learning techniques including Same Topic Prediction (STP), Topic Classification (TC) and Next Sentence Prediction (NSP). When these three classification layers are combined in a multi-task manner, they can make up for each other's limitations, improving performance in all three tasks. We experiment different combinations of the three layers and report how each layer affects other layers in the same combination as well as the overall segmentation performance. The model we proposed achieves the state-of-the-art result in the WikiSection dataset.