Emotion Analysis Techniques in Alibaba's Intelligent Customer Service System

The paper introduces the growing importance of human‑machine dialogue in natural language processing and describes how Alibaba's intelligent customer service system (AliMe) integrates emotion analysis to improve both robot and human assistance.

It outlines a six‑dimensional framework covering user emotion detection, emotion soothing, generative emotional dialogue, service quality inspection, conversation satisfaction estimation, and intelligent human entry, each supported by specific models and datasets.

User Emotion Detection Model: An integrated model combines word‑level, phrase‑level, and sentence‑level semantic features using SWEM, CNN n‑gram extraction, and LEAM label‑embedding attention, achieving high accuracy on a large annotated dataset of 865,666 samples.

Emotion Smoothing and Knowledge‑Based Response: Offline modules identify seven target emotions and 35 topic categories, constructing knowledge bases for precise soothing replies; online modules employ deep text‑matching (BCNN and MatchPyramid) with Lucene pre‑retrieval to select appropriate knowledge entries.

Generative Emotional Dialogue: A Seq2Seq architecture with attention incorporates emotion embeddings (E) and topic embeddings (T) to generate context‑aware, emotionally appropriate responses, improving content quality and meeting length constraints.

Service Quality Inspection: Models detect negative or poor‑attitude service utterances using GRU‑based context encoding, multi‑level semantic features, and emotion labels, achieving better detection than baseline approaches.

Conversation Satisfaction Prediction: A hierarchical GRU model extracts semantic, emotional, and answer‑source features, with embeddings to compress sparse one‑hot vectors, and predicts satisfaction via classification and regression, significantly reducing error compared to raw user feedback.

Intelligent Human Entry Ranking: A hierarchical DeepFM model combines current conversation features and historical user statistics to rank users for handoff, improving click‑through rates and reducing dissatisfaction.

Extensive offline experiments and online A/B tests demonstrate improvements in model performance, user satisfaction, and operational efficiency across all modules.

The conclusion highlights future work on multilingual emotion analysis, sentiment‑driven recommendation, and further integration of emotional intelligence into e‑commerce dialogue systems.