Sungjoon Park
November 5(Thu) - November 5(Thu), 2020
7PM
ZOOM (ID: 728-142-6028 )
CNIR Seminar
Date: 7PM, Thursday, November 5th
ZOOM 회의 참가 ID: 728-142-6028
Speaker: Sungjoon Park
(Department of Computer Science, Ph.D. Candidate, KAIST, Republic of Korea)
Title: Detecting Client's States and Emotion from Text via Pre-trained Language Models
Abstract:
The recent surge of text-based online counseling applications enables us to collect and analyze interactions between counselors and clients. A dataset of those interactions can be used to learn to automatically classify the client utterances into categories that help counselors in diagnosing client status and predicting counseling outcomes. With proper anonymization, we collect counselor-client dialogues, define meaningful categories of client utterances with professional counselors, and develop a novel neural network model for classifying the client utterances. The central idea of our model, ConvMFiT, is a pre-trained conversation model that consists of a general language model built from an out-of-domain corpus and two role-specific language models built from unlabeled in-domain dialogues. The classification result shows that ConvMFiT outperforms state-of-the-art comparison models. Further, the attention weights in the learned model confirm that the model finds expected linguistic patterns for each category.
Second, we propose a framework that makes a model predict fine-grained dimensional emotions (valence-arousal-dominance, VAD) trained on corpus annotated with coarse-grained categorical emotions. We train a model by minimizing EMD distances between predicted VAD score distribution and sorted categorical emotion distributions in terms of VAD, as a proxy of target VAD score distributions. With our model, we can simultaneously classify a given sentence to categorical emotions as well as predict VAD scores. We use pre-trained BERT-Large and fine-tune on the SemEval dataset (11 categorical emotions) and evaluate on EmoBank (VAD dimensional emotions), in order to show our approach reaches comparable performance to that of the state-of-the-art classifiers in categorical emotion classification task and significant positive correlations with ground truth VAD scores. Also, if one continues training our model with the supervision of VAD labels, it outperforms state-of-the-art VAD regression models. We further present examples showing our model can annotate emotional words suitable for a given text even those words are not seen as categorical labels during training.