Intelligent Customer Service System for Airline Ticket Business: Architecture, Data Analysis, and AI Techniques

Introduction

The article introduces the need for intelligent customer service in the airline ticket domain, highlighting the limitations of traditional support and positioning AI‑driven solutions as a key improvement for user satisfaction and cost reduction.

Problem Analysis

Analysis of chat logs shows that user issues are roughly 30% business‑related and 70% conversational, with an average of 13 dialogue turns per ticket and some cases exceeding 100 turns. Complex scenarios such as ticket changes, refunds, and cancellations require extensive knowledge.

Intelligent Customer Service Scenarios

Two main scenarios are described: a "customer‑assistant" that suggests answers to human agents, and an autonomous chatbot that directly interacts with users. The assistant achieves over 90% answer push rate, 60% recommendation coverage, 90% intent‑recognition accuracy, and 70% answer accuracy.

Basic QA

Knowledge is extracted from historical logs using language‑model quality filtering, TextRank for domain term extraction, and sentence‑vector similarity. The resulting QA pairs feed a retrieval system that combines inverted indexes, word2vec similarity, and an ES+BERT re‑ranking layer.

Intent Recognition

Initially fastText was used, but as intent categories grew to over 200, BERT was adopted for few‑shot fine‑tuning. A hybrid approach combines BERT predictions with SVM classifiers built on BERT embeddings to quickly iterate on error cases, achieving stable 91% accuracy online.

Technical Architecture

The system consists of four modules:

Basic Service Module : query preprocessing, entity extraction, intent detection, and other NLP services.

Dialogue Control Module : manages dialogue state, slot filling, and orchestrates downstream services.

Dialogue Generation Module : generates business‑specific answers based on order status and identified intent.

Data Management Module : updates real‑time recommendation feedback and refines dialogue strategies.

The overall flow starts with query cleaning, named‑entity recognition, query rewriting, and vectorization, followed by either QA retrieval for chit‑chat or dialogue management for business intents, as illustrated in the architecture diagram.

Future Plans

Enhance intent‑recognition and build a ticket‑domain knowledge graph.

Improve entity and time extraction, and refine slot‑filling in multi‑turn dialogues.

Extend NLG capabilities by leveraging historical logs and order status.

Modularize components to form an independent NLP service platform.

References

Devlin et al., BERT: Pre‑training of Deep Bidirectional Transformers for Language Understanding, 2018.

McCandless & Hatcher, Lucene in Action, 2010.

Mikolov et al., Efficient Estimation of Word Representations in Vector Space, 2013.

Jay Alammar, The Illustrated BERT, ELMo, and Co., 2018.

Joulin et al., Bag of Tricks for Efficient Text Classification, 2016.