Hot Topic Detection Algorithms and Article Deduplication Evaluation without a Test Set

Q: How can hot topics and related articles be discovered, and which algorithms are involved?

A: Hot topics evolve through three stages: emergence, heated discussion, and decline. In the emergence stage, models rely on prior features like media authority and historical popularity; during the heated discussion stage, simple metrics such as article count and ranking lists become effective; the decline stage uses similar metrics as the discussion stage.

Methods for finding related articles include:

Keyword extraction (e.g., TextRank) followed by keyword‑based matching.

Embedding‑based approaches (e.g., BERT) with vector similarity to measure article relevance.

Search‑based techniques using article keywords as queries, employing relevance measures such as BM25.

Q: When lacking a test set for article deduplication, how can the quality of different deduplication methods be assessed?

A: Accuracy can be measured directly, while recall is harder to gauge. A step‑by‑step approach includes:

Randomly sample articles, apply each deduplication algorithm, manually label the results, and compute accuracy; treat unlabeled pairs as non‑similar to estimate recall.

Expand the labeled set by using heuristic algorithms (e.g., keyword‑based deduplication), label their outputs, and evaluate both accuracy and recall, balancing labeling effort and coverage.

If resources permit, fully label a random sample of articles to obtain the most reliable accuracy and recall metrics, though this is costly.