The Study Of Models And Features For Non-factoid Question Answering

It is an important and challenge task to study the question answering models and features for non-factoid questions. Currently, many studies try to address this task by using traditional models with textual features and non-textual features (e.g, web log, etc.) Although these studies have made some progresses, there are still many problems to be solved in these studies. For example, features are too complicated to be extracted and generated. Besides, the traditional models can combine neither the textual features nor the non-textual features into it smoothly. Thus, the results and performances of previous studies are far from satisfactory. Moreover, our preliminary study showed the discourse relation as well as the user profile information can help on this task. To the best of our knowledge, few studies make use of the importance of the discourse relation and the user profile information of the online community question answering to improve the performance of non-factoid QA.In order to improve the features and model, this thesis firstly incorporates the user profile information and discourse relation with textual and non-textual features into this task. After that, all the features are classified and chosen by logistic regression model in order to minimize the complexity and redundancy among them. Then, we replace the whole feature set with the current chosen one. At the same time, we employ the learning to rank model to predict the best answers of non-factoid questions with the chosen features.Results on the Yahoo! Answers Manner Question show that the chosen features achieve the comparable results with the whole feature set. The results also indicate that the user profile information performs better than the discourse relation on predicting best answers. Specifically, the listwise learning to rank model is more suitable than the pairwise learning to rank model on this task.