Persuasive dialogue understanding: The baselines and negative results

Persuasion aims at forming one’s opinion and action via a series of persuasive messages containing persuader’s strategies. Due to its potential application in dialogue systems, the task strategy recognition has gained much attention lately. Previous methods on user intent systems adopt recurrent neural network (RNN) or convolutional (CNN) model context conversational history, neglecting tactic history intra-speaker relation. In this paper, we demonstrate limitations Transformer-based approach coupled with Conditional Random Field (CRF) for recognition. model, leverage inter- contextual semantic features, as well label dependencies improve Despite extensive hyper-parameter optimizations, architecture fails outperform baseline methods. We observe two negative results. Firstly, CRF cannot capture dependencies, possibly strategies dialogues do not follow any strict grammar rules cases Named Entity Recognition (NER) part-of-speech (POS) tagging. Secondly, Transformer encoder trained from scratch is less capable capturing sequential information than Long Short-Term Memory (LSTM). attribute reason that vanilla does efficiently consider relative position sequence elements.