How ICBC Leverages Text Mining to Transform Customer Service

ICBC Text Mining Technology Exploration

ICBC, traditionally seen as a massive but steady bank, is responding to big‑data challenges by transforming its customer service through text mining.

Listening to customer voices—both internal call‑center records and external media—provides rich textual data. This article shares the full workflow from design to implementation, covering:

Practical cases of text mining in customer service analysis;

Introducing text data into banking decision‑making;

Unlocking commercial value of massive textual data.

Traditional Customer Service Analysis Process

After a 95588 call, structured data (complaint count, satisfaction scores) is analyzed automatically, while unstructured text is reviewed manually, limiting coverage.

Integrated Text Mining Process

With text mining, hot opinions are extracted from unstructured feedback and linked to structured data, expanding analysis scope, improving efficiency, and enabling automatic reading of texts.

Business Value of Opinion Mining

Identifying common issues allows proactive measures, improving satisfaction and loyalty while reducing complaint volume and service costs.

Case Studies

Two cases are presented: (1) extracting opinions from 95588 tickets, and (2) listening to online customer voices.

Opinion Mining – Objectives

Classify feedback into categories (e.g., self‑service, card, branch) to track trends and perform targeted handling.

Hot‑word analysis visualizes frequent terms, though ambiguity remains (e.g., “staff” vs. “teller”).

Precise Opinion Representation

Use a triple object‑attribute‑evaluation (e.g., “teller attitude poor”) to summarize sentiments.

Model Building

Naïve Bayes was tried but suffered from lack of training data and independence assumptions. LDA topics were too coarse. An ontology‑based model separating business elements (objects, attributes) from language concepts was adopted.

Word2Vec embeddings assist semi‑automatic ontology construction, revealing semantic relations such as “ATM” ↔ “cash machine”.

Implementation Effects

Linking individual opinions with customer star ratings shows differing concerns across segments, enabling differentiated service strategies. Association analysis uncovers product‑issue links (e.g., U‑shield brands). Repeated‑complaint analysis highlights intolerable problems and cost‑inefficient areas.

Listening to Internet Voices

An automated system collects news and social‑media mentions of ICBC, filters spam using keywords and Naïve Bayes, and clusters events using a “Chinese restaurant process” to handle dynamic topic numbers.

Important events are identified with a logistic‑regression model based on factors such as institution relevance, sentiment, repost count, source type, and business relevance, providing early warnings and prioritized daily reports.

Key Takeaways

Choosing business‑aligned algorithms, building domain‑specific semantic resources, and starting with small, solvable problems are essential for successful text‑mining projects in traditional industries.

Overall, the discussion covered opinion extraction with object‑attribute‑evaluation, ontology‑driven modeling aided by word vectors, spam filtering, cross‑channel event clustering, and automatic importance detection.