Problem
Alfa-Bank call-center operators answer clients who have already been through a chatbot, so the questions are atypical. Three things slow them down: finding the right one of 15 internal interfaces, finding a reply template stored in Word files (search needs the exact wording), and searching a knowledge base of tens of thousands of articles with weak built-in search. The team had to build recommenders for all three with very little labeled data.
Example from the article: the knowledge-base model had about 50 thousand articles to choose from and only 300 labeled examples.
Solution in brief
Show the operator three automatic recommendation blocks driven by the chat context. Train each model with self-supervised pretraining on chat logs first. For templates and articles, add contrastive training on mined context-to-template pairs, and mix in unsupervised steps (chat context vs. operator reply) to stop overfitting. The article model also uses title augmentation and weights from the template model.
Questions this case answers
- How to recommend a short list of templates, interfaces and documents from the context of a conversation when labeled context-to-item pairs are scarce and the item set is large
- How to stop a contrastive retrieval model from overfitting when it is trained on a small supervised set, using unlabeled context-reply pairs as a second training stream
- How to recommend interfaces, reply templates and knowledge-base articles to call-center operators from the chat context
- How to train a text-matching model when almost no labeled data exists
Tools and methods
- Multilabel classification (TinyBERT, fastText with LSTM) — picks the interfaces the operator needs now.
- BERT pretrained with masked language modeling — the base of the template and article models.
- Contrastive loss on in-batch pairs — pulls a chat context toward its template and pushes other pairs apart.
- Semi-supervised training — alternates supervised and unsupervised contrastive steps.
- Weight transfer — the article model starts from the template model.
Google or a one-time email code. No password.