Sentiment Classification in Customer Service Dialogue with Topic-Aware Multi-Task Learning

Authors: Jiancheng Wang, Jingjing Wang, Changlong Sun, Shoushan Li, Xiaozhong Liu, Luo Si, Min Zhang, Guodong Zhou9177-9184

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this study, we focus on the sentiment classification task in an important type of dialogue, namely customer service dialogue, and propose a novel approach which captures overall information to enhance the classification performance. Specifically, we propose a topic-aware multi-task learning (TML) approach which learns topicenriched utterance representations in customer service dialogue by capturing various kinds of topic information. In the experiment, we propose a large-scale and high-quality annotated corpus for the sentiment classification task in customer service dialogue and empirical studies on the proposed corpus show that our approach significantly outperforms several strong baselines.
Researcher Affiliation Collaboration 1School of Computer Science and Technology, Soochow University, China 2Alibaba Group, China
Pseudocode No The paper includes architectural diagrams but no structured pseudocode or algorithm blocks.
Open Source Code Yes Furthermore, to facilitate the research, we annotate a large-scale corpus1 of high quality on a real-world dataset in E-commerce customer service. ... 1https://github.com/jc-wang/TML
Open Datasets Yes We collect a dialogue dataset from an online customer service system in a top E-commerce company in China. For the annotation, we define five sentiment labels, i.e., very negative, negative, neutral, positive, and very positive. ... Table 1 shows the detail statistics of the final dataset. ... 1https://github.com/jc-wang/TML
Dataset Splits Yes Table 1: Data distributions of the annotated data. Train 2555 49710 ... Dev. 320 6357 ... Test 621 12149
Hardware Specification No The paper does not provide any specific hardware details used for running its experiments.
Software Dependencies No The paper mentions software components like BERT, LSTM, Glove, and Adam optimizer but does not specify their version numbers.
Experiment Setup Yes The dimension of LSTM hidden state is set to be 256. The dimension M of embeddings in topic model is set to be 100 and the number of topics (K) is set to be 20 for all topic models. ... Batch size is set to be 32. In addition, other hyper-parameters are fine-tuned with the development data. Specifically, λ in Eq.(18) is set to be 0.01. The dropout rate (Srivastava et al. 2014) is 0.3. The L2 regularization weight of parameters is 10 5. Finally, we use Adam optimizer (Kingma and Ba 2014) for training our TML approach with the initial learning rate of 0.001.