PM-AI commited on
Commit
bb99ef8
1 Parent(s): 858f925

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +7 -2
README.md CHANGED
@@ -186,7 +186,7 @@ The following table shows the evaluation results for different approaches and mo
186
  **model**|**NDCG@1**|**NDCG@10**|**NDCG@100**|**comment**
187
  :-----:|:-----:|:-----:|:-----:|:-----:
188
  bi-encoder_msmarco_bert-base_german (new) | 0.5300 <br /> 🏆 | 0.7196 <br /> 🏆 | 0.7360 <br /> 🏆 | "OUR model"
189
- [deepset/gbert-base-germandpr-X](https://huggingface.co/deepset/gbert-base-germandpr-ctx_encoder) | 0.4828 | 0.6970 | 0.7147 | "has two encoder models (one for queries and one for corpus), is SOTA approach"
190
  [distiluse-base-multilingual-cased-v1](https://huggingface.co/sentence-transformers/distiluse-base-multilingual-cased-v1) | 0.4561 | 0.6347 | 0.6613 | "trained on 15 languages"
191
  [paraphrase-multilingual-mpnet-base-v2](https://huggingface.co/sentence-transformers/paraphrase-multilingual-mpnet-base-v2) | 0.4511 | 0.6328 | 0.6592 | "trained on huge corpus, support for 50+ languages"
192
  [distiluse-base-multilingual-cased-v2](https://huggingface.co/sentence-transformers/distiluse-base-multilingual-cased-v2) | 0.4350 | 0.6103 | 0.6411 | "trained on 50+ languages"
@@ -195,10 +195,15 @@ bi-encoder_msmarco_bert-base_german (new) | 0.5300 <br /> 🏆 | 0.7196 <br />
195
  [BM25](https://www.elastic.co/guide/en/elasticsearch/reference/current/index-modules-similarity.html#bm25) | 0.3196 | 0.5377 | 0.5740 | "lexical approach"
196
 
197
  **It is crucial to understand that the comparisons are also made with models based on other transformer approaches.**
198
- For example, in particular [deepset/gbert-base-germandpr-X](https://huggingface.co/deepset/gbert-base-germandpr-ctx_encoder) is theoretically a more up-to-date approach that is nevertheless beaten.
199
  A direct comparison based on the same approach can be made with [svalabs/bi-electra-ms-marco-german-uncased](svalabs/bi-electra-ms-marco-german-uncased).
200
  In this case, the model presented here outperforms its predecessor by up to 14 percentage points.
201
 
 
 
 
 
 
 
202
  Note:
203
  - Texts used for evaluation are sometimes very long. All models, except for BM25 approach, truncate the incoming texts some point. This can decrease performance.
204
  - Evaluation of deepset's gbert-base-germandpr model might give an incorrect impression. The model was originally trained on the data we used for evaluation (not 1:1 but almost).
 
186
  **model**|**NDCG@1**|**NDCG@10**|**NDCG@100**|**comment**
187
  :-----:|:-----:|:-----:|:-----:|:-----:
188
  bi-encoder_msmarco_bert-base_german (new) | 0.5300 <br /> 🏆 | 0.7196 <br /> 🏆 | 0.7360 <br /> 🏆 | "OUR model"
189
+ [deepset/gbert-base-germandpr-X_encoder](https://huggingface.co/deepset/gbert-base-germandpr-ctx_encoder) | 0.4828 | 0.6970 | 0.7147 | "has two encoder models (one for queries and one for corpus), is SOTA approach"
190
  [distiluse-base-multilingual-cased-v1](https://huggingface.co/sentence-transformers/distiluse-base-multilingual-cased-v1) | 0.4561 | 0.6347 | 0.6613 | "trained on 15 languages"
191
  [paraphrase-multilingual-mpnet-base-v2](https://huggingface.co/sentence-transformers/paraphrase-multilingual-mpnet-base-v2) | 0.4511 | 0.6328 | 0.6592 | "trained on huge corpus, support for 50+ languages"
192
  [distiluse-base-multilingual-cased-v2](https://huggingface.co/sentence-transformers/distiluse-base-multilingual-cased-v2) | 0.4350 | 0.6103 | 0.6411 | "trained on 50+ languages"
 
195
  [BM25](https://www.elastic.co/guide/en/elasticsearch/reference/current/index-modules-similarity.html#bm25) | 0.3196 | 0.5377 | 0.5740 | "lexical approach"
196
 
197
  **It is crucial to understand that the comparisons are also made with models based on other transformer approaches.**
 
198
  A direct comparison based on the same approach can be made with [svalabs/bi-electra-ms-marco-german-uncased](svalabs/bi-electra-ms-marco-german-uncased).
199
  In this case, the model presented here outperforms its predecessor by up to 14 percentage points.
200
 
201
+ Comparing with [deepset/gbert-base-germandpr-X_encoder](https://huggingface.co/deepset/gbert-base-germandpr-ctx_encoder) is theoretically a little unfair since deepset's approach is based on two models at the same time!
202
+ Queries and passages are encoded separately which leads to a better, more superior contextualization.
203
+ Still, our newly trained model is outperforming the other approach by around two percentage points.
204
+ In addition, using two models at the same time also increases demands on memory and CPU power which causes higher costs.
205
+ This makes the approach presented here even more valuable.
206
+
207
  Note:
208
  - Texts used for evaluation are sometimes very long. All models, except for BM25 approach, truncate the incoming texts some point. This can decrease performance.
209
  - Evaluation of deepset's gbert-base-germandpr model might give an incorrect impression. The model was originally trained on the data we used for evaluation (not 1:1 but almost).