Update README.md
Browse files
README.md
CHANGED
@@ -18,4 +18,60 @@ configs:
|
|
18 |
data_files:
|
19 |
- split: test
|
20 |
path: queries/test*
|
21 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
data_files:
|
19 |
- split: test
|
20 |
path: queries/test*
|
21 |
+
---
|
22 |
+
|
23 |
+
Dataset **CapRetrieval** introduced in [Dense Retrievers Can Fail on Simple Queries: Revealing The Granularity Dilemma of Embeddings](https://arxiv.org/abs/2506.08592).
|
24 |
+
|
25 |
+
CapRetrieval evaluates the fine-grained embedding matching (dense passage retrieval) in Chinese, tailored towards a practical image search scenario:
|
26 |
+
- Candidate passages are image captions, and queries are short phrases of entities or events reflected in captions.
|
27 |
+
- Overall, the dataset comprises seemingly simple queries and captions; however, text encoders are shown limitations resolving these cases.
|
28 |
+
- Evaluation results call for attention on embedding training strategies with different **granularity**.
|
29 |
+
|
30 |
+
### Format
|
31 |
+
|
32 |
+
CapRetrieval follows the same retrieval task format as in MTEB, with relevance labels in [0,1,2] for each pair.
|
33 |
+
Note that unlike prior datasets, we annotate full labels for each query-passage pair (1.3 million pairs), minimizing false negatives for more accurate evaluation.
|
34 |
+
|
35 |
+
A small amount of queries do not have any relevant captions; they are excluded in computation of retrieval metrics (e.g. nDCG), but can be useful for other analysis, e.g. in classification setting.
|
36 |
+
|
37 |
+
### Evaluation
|
38 |
+
|
39 |
+
Please see the evaluation script and results at https://github.com/lxucs/CapRetrieval.
|
40 |
+
|
41 |
+
|
42 |
+
| Type | Model | nDCG@10 |
|
43 |
+
|----------|-------------------------|-----------|
|
44 |
+
| **BM25** | Basic BM25 | 66.54 |
|
45 |
+
| **0.1B** | bge-base-zh-v1.5 | 78.86 |
|
46 |
+
| | gte-multilingual-base | 79.67 |
|
47 |
+
| | multilingual-e5-base | 76.33 |
|
48 |
+
| **0.3B** | bge-large-zh-v1.5 | 79.15 |
|
49 |
+
| | multilingual-e5-large | 81.01 |
|
50 |
+
| | Conan-embedding-v1 | 77.04 |
|
51 |
+
| **0.6B** | Qwen3-Embedding-0.6B | 81.04 |
|
52 |
+
| **>1B** | gte-Qwen2-1.5B-instruct | 77.35 |
|
53 |
+
| | gte-Qwen2-7B-instruct | **86.55** |
|
54 |
+
| | e5-mistral-7b-instruct | 76.40 |
|
55 |
+
| | Qwen3-Embedding-8B | 84.61 |
|
56 |
+
| | | |
|
57 |
+
| Trained | Out-of-Domain | 87.23 |
|
58 |
+
| | In-Domain | 91.83 |
|
59 |
+
|
60 |
+
|
61 |
+
The trained models (based on `bge-base-zh-v1.5`) are trained with queries by our data generation strategies described in the paper. The in-domain model can be downloaded from [Google Drive](https://drive.google.com/drive/folders/1l2pvELMQPKjhAasNGaY7d14jMK0iCRhj).
|
62 |
+
|
63 |
+
|
64 |
+
### Citation
|
65 |
+
|
66 |
+
```bibtex
|
67 |
+
@misc{xu2025denseretrieversfailsimple,
|
68 |
+
title={Dense Retrievers Can Fail on Simple Queries: Revealing The Granularity Dilemma of Embeddings},
|
69 |
+
author={Liyan Xu and Zhenlin Su and Mo Yu and Jiangnan Li and Fandong Meng and Jie Zhou},
|
70 |
+
year={2025},
|
71 |
+
eprint={2506.08592},
|
72 |
+
archivePrefix={arXiv},
|
73 |
+
primaryClass={cs.CL},
|
74 |
+
url={https://arxiv.org/abs/2506.08592},
|
75 |
+
}
|
76 |
+
```
|
77 |
+
|