Sentence Transformers are evaluated primarily through standardized benchmarks, intrinsic metrics, and downstream tasks to measure how well their embeddings capture semantic similarity. The most common approach uses datasets specifically designed for semantic textual similarity (STS), where sentence pairs are annotated with similarity scores by humans. Models generate embeddings for each sentence, and the cosine similarity between embeddings is compared to human ratings using correlation metrics like Pearson or Spearman. For example, the STS Benchmark (STS-B) dataset contains sentence pairs (e.g., “A man is playing a guitar” vs. “A musician is performing”) rated on a 0–5 scale. A strong correlation between the model’s similarity scores and human judgments indicates better performance. Other datasets like SICK-R or MRPC are also used to test robustness across diverse sentence structures and domains.
Another key evaluation method involves retrieval and classification tasks. In retrieval scenarios, models are tested on their ability to find semantically similar sentences in large collections. For instance, the MS MARCO dataset evaluates how well embeddings retrieve relevant passages for a query. Metrics like recall@k (how often the correct result appears in the top-k retrieved items) or mean average precision (MAP) quantify effectiveness. For classification, embeddings are used as input features for tasks like paraphrase detection (e.g., Quora Question Pairs) or intent recognition. High accuracy here suggests the embeddings preserve semantic meaning. Clustering tasks, evaluated using metrics like adjusted Rand index, also test if embeddings group sentences with similar meanings (e.g., clustering news articles by topic).
Finally, cross-domain and cross-lingual generalization are tested to ensure models aren’t overfitting to specific datasets. Models trained on English STS data might be evaluated on non-English datasets like XNLI (cross-lingual natural language inference) to assess multilingual capability. Ablation studies, where components like pooling strategies or loss functions are removed, help identify what drives performance. For example, replacing mean pooling with max pooling might reduce performance, highlighting the importance of that design choice. These evaluations ensure the model’s effectiveness isn’t limited to narrow scenarios and can generalize across languages, domains, and applications.