Long Document Summarization — NAVER CLOVA AI RUSH 2022
1st place in the NAVER CLOVA AI RUSH 2022 Long Document Summarization track. Domain-specific branched summarization system for book, dialogue, document, and journal inputs.
2022.07 — 2022.09 NAVER CLOVA Participant (1st Place)
NLP Summarization PLM ELECTRA T5 BART Competition
Overview
Won 1st place in the Long Document Summarization track of the NAVER CLOVA AI RUSH 2022 competition. Designed and implemented a domain-specific branched summarization system that routes book, dialogue, document, and journal inputs to domain-optimized PLMs. Evaluated on the average of ROUGE-1/2/L F-scores.
Key Results
- Built a branched summarization pipeline using an ELECTRA-base classifier to detect input document domain and route to the optimal per-domain PLM
- Per-domain PLM selection and fine-tuning:
- Book: paust/pko-t5-base + finetuning (512 tokens)
- Dialogue: alaggung/bart-rl (512 tokens)
- Document: ainize/kobart-news (1024 tokens)
- Journal: paust/pko-t5-base + finetuning & extractive summarization (512 tokens)
- Journal domain leveraged format keywords from structured official documents for extractive summarization, achieving high ROUGE with fast inference
- ELECTRA classifier detected OOD (Out-of-Distribution) documents and routed them to the most similar domain
Design Constraints and Decisions
- Domain-specific branching outperformed a single unified model on total ROUGE: document length, style, and structure varied too much across domains for a single model to generalize
- Adjusted max token count per domain based on average input length differences to save inference time
- Added ELECTRA-based OOD routing to handle unknown domains in the test set, since training data covered only 4 domains