TL;DR
Researchers have introduced K-SportsSum, a large-scale, manually cleaned dataset designed to bridge the "knowledge gap" in automated sports reporting. By combining a new knowledge-enhanced summarizer (KES) with a massive corpus of player and team metadata, the system can transform messy, colloquial live commentaries into polished, fact-rich news articles that outperform previous state-of-the-art models.
What Problem Does This Paper Address?
Generating a professional news article from a live sports commentary is significantly harder than standard text summarization. The authors identify three primary hurdles:
- The Length Challenge: Live commentaries are massive, often exceeding the token limits of standard transformer models like BERT.
- The Style Gap: Commentaries are informal and colloquial ("Goal!!!"), while news articles are formal and structured.
- The Knowledge Gap: This is the paper's core focus. Professional sports news often contains background information—such as a player’s height, a team’s nickname (e.g., "Purple Lily" for ACF Fiorentina), or historical context—that simply isn't present in the live play-by-play.
Existing datasets were either too small or filled with "noise" (ads, HTML tags, and irrelevant hyperlinks) because they relied on automated cleaning.
Methods
The authors propose a two-step Knowledge Enhanced Summarizer (KES):
- Step 1: Key Commentary Selection: Since most play-by-play lines are filler, the model uses a RoBERTa-based selector to identify the most critical events (goals, substitutions, etc.) in a sliding window.
- Step 2: Knowledge Retrieval: For every selected sentence, the system performs Named Entity Recognition (using a model called FLAT) to identify players and teams. It then links these entities to a custom Knowledge Corpus using Levenshtein distance to pull in relevant background facts.
- Step 3: Knowledge-Aware Rewriting: The final news sentence is generated using mT5 (a multilingual T5 model). Crucially, the authors use Segment Embeddings to tell the model which tokens are players, teams, or timestamps, and Knowledge Embeddings to inject the external facts retrieved in Step 2.
Datasets
The paper introduces K-SportsSum, which is currently the largest and highest-quality dataset in this niche:
- 7,854 commentary-news pairs: Manually cleaned by 9 annotators to remove ads and irrelevant "noise" found in previous datasets like SportsSum.
- Knowledge Corpus: A supplementary database containing structured info for 14,724 players and Wikipedia-aligned articles for 523 teams.
- Evaluation: The model was tested against SportsSum and K-SportsSum using ROUGE scores and human evaluation.
Key Insights and Ideas
- Summarization is actually "Enrichment": The authors argue that sports summarization isn't just about shortening text; it’s about fusing real-time events with historical/biographical data.
- Manual Cleaning is Essential: The researchers found that over 15% of news articles in previous datasets contained noise that confused models. Their manual "denoising" process led to a significant jump in model performance.
- Contextual Selection: Unlike previous models that looked at commentary lines in isolation, KES uses RoBERTa to understand the context surrounding a play, leading to much better "key event" detection.
Key Results
- Quantitative: KES achieved a ROUGE-L score of 47.17 on the K-SportsSum dataset, significantly outperforming the previous state-of-the-art SportsSUM model (44.16).
- Qualitative: In human trials, KES was rated higher in informativeness and fluency. For example, while a baseline model might just say a player was substituted, KES could add, "...who is 1.7 meters height," providing the kind of detail expected in professional journalism.
Limitations and Future Work
Despite the improvements, the authors note that the model still occasionally suffers from factual hallucinations. In one instance, the model correctly identified that it should add a player's height but stated "1.7 meters" when the actual height was 1.8 meters.
Future work involves:
- Exploring models that explicitly utilize knowledge graphs rather than just text passages.
- Leveraging "long-context" transformers (like Longformer or ETC) to process the entire game commentary at once rather than using a two-step selection process.
Summary updated