SK Telecom · May 2022 – Dec 2023
Cupid: Answer Recommendation for Similar Questions
16→31%AI auto-answer rateduplicate questions, Aug 1 → Aug 21, 2022
38→57%positive feedback"helpful" ratio on AI answers
01Background & Goals
- Cupid is a community service where nearby users ask and answer each other's questions.
- We wanted an AI feature that finds answers already given to similar questions.
02Key Challenges
- C1 Question–answer matching accuracy across topics and regions
- C2 Automated evaluation of the main models
03Contributions
Content-based recommendation pipelineC1
- Classification and extraction: extracts location and category from the question (BERT-based classifiers and NER)
- Filtering: narrows candidates by location, category and stop words
- Encoding: masks place names and converts the question to a vector with the embedding model
- Retrieval: finds similar questions in Elasticsearch with BM25 + cosine similarity
- Re-ranking: scores question containment and question–answer consistency and returns the top N
Models behind the pipelineC1
- Embedding model: trained a BERT-based model with a Siamese network so similar sentences sit close together
- Hierarchical multi-class classifier: built hierarchical topic data and trained a BERT-based classifier
- NER model: trained with BIO tagging to extract places and keywords
Model evaluation pipelineC2
- Sent items whose category changed plus N sampled items to an external vendor for ground-truth labeling
- Evaluated with F1 and accuracy and extracted items with low prediction accuracy
- Gated deployment on a golden-set evaluation after retraining
Batch processing of new dataC1
- Automated an Airflow batch that encodes new question–answer pairs and re-indexes the vector DB so they become recommendation candidates
04Tech Stack
- Framework / Platform
- PyTorch, Transformers, Elasticsearch, Airflow
- Methodology
- Siamese network, Hierarchical multi-class classification, BIO NER
05Results
- AI auto-answer rate for duplicate questions rose from 16% to 31% (Aug 1 → Aug 21, 2022)
- Positive ("helpful") feedback on AI answers rose from 38% to 57%