← All questions
Retrieval & EmbeddingsHard· Query Expansion· Recall· nDCG

You shipped WordNet synonym expansion. Recall@100 rose...

Lexical expansion via WordNet increases the size of the retrieved document set. Because the retrieved set is now a superset of the original, recall@100 will naturally rise as a mathematical consequence of the expansion, rather than as evidence of improved retrieval relevance. This is a common pitfall in evaluating expansion techniques.

The drop in nDCG@10 is the more critical signal. It indicates that the expansion introduced ‘noise’—irrelevant documents that were ranked highly enough to displace relevant documents from the top 10 results. To improve, you should ship a more selective expansion strategy, such as filtering synonyms based on context or confidence scores, to ensure that the expansion improves recall without degrading precision.

Code playground (Python)