Extra-iNet:an extractive framework for explaining text-driven predictions
A self-explanatory sentence extraction framework(Extra-iNet)was proposed,which relied solely on coarse-grained task-level labels to extract certain sentences from input documents to guide document label prediction and provide human-readable explanations for the prediction results.This framework utilized a convolutional neural network to encode input sentences,and integrated the extraction and representation of predictive explanatory sentences into a single module.Reinforcement learning and gating mechanisms were adopted as two strategies to extract the explanatory subset,resulting in two variant models,namely Hard Extra-iNet and Soft Extra-iNet.The predictive and explanatory capabilities of the models were validated on sentiment analysis task and cumulative abnormal return prediction task.The results show that,compared to the baselines,the model's F1(the harmonic mean of precision and recall)in sentiment analysis task increases by an average of 15.5%,and its accuracy in cumulative abnormal return prediction task increases by an average of 3.7%.The results extracted by the model are highly consistent with the results annotated by humans.Extra-iNet framework can effectively extract explanatory sentences from documents and can be applied to the prediction in other fields.