Recognizing stances in Mandarin social ideological debates with text and acoustic features

Recognizing stances is of great importance to understand intention of human beings. While previous related researches mainly focused on text modality, in this paper, we aim to combine textual and acoustic features to automatically recognize stances in social debates. For acoustic features, we find that speaking rate is an indispensable feature to distinguish whether speaker is taking a stance or not. In addition, we also demonstrate that emphasis information is helpful for recognizing stances. For textual modality, we present a novel Support Topic Feature (STF) and use it to recognize which stance the speaker is taking (positive, negative or neutral). Experiments on four debate datasets confirm that the performance of STF is much better than that of n-gram features. When using STF only, F1-measure can be improved by 3% ~ 8% as compared to the baseline. Combining acoustic features with STF leads to even better performance by improving F1-measure with 2% ~ 8% further.