This paper presents models for detecting agreement/disagreement in onlinediscussions. In this work we show that by using a Siamese inspired architectureto encode the discussions, we no longer need to rely on hand-crafted featuresto exploit the meta thread structure. We evaluate our model on existing onlinediscussion corpora - ABCD, IAC and AWTP. Experimental results on ABCD datasetshow that by fusing lexical and word embedding features, our model achieves thestate of the art performance of 0.804 average F1 score. We also show that themodel trained on ABCD dataset performs competitively on relatively smallerannotated datasets (IAC and AWTP).
展开▼
