Update README.md
Browse files
README.md
CHANGED
|
@@ -1,86 +1,86 @@
|
|
| 1 |
-
---
|
| 2 |
-
tags:
|
| 3 |
-
- feature-extraction
|
| 4 |
-
language: en
|
| 5 |
-
datasets:
|
| 6 |
-
- SciDocs
|
| 7 |
-
- s2orc
|
| 8 |
-
metrics:
|
| 9 |
-
- F1
|
| 10 |
-
- accuracy
|
| 11 |
-
- map
|
| 12 |
-
- ndcg
|
| 13 |
-
license: mit
|
| 14 |
-
---
|
| 15 |
-
|
| 16 |
-
## SciNCL
|
| 17 |
-
|
| 18 |
-
SciNCL is a pre-trained BERT language model to generate document-level embeddings of research papers.
|
| 19 |
-
It uses the citation graph neighborhood to generate samples for contrastive learning.
|
| 20 |
-
Prior to the contrastive training, the model is initialized with weights from [scibert-scivocab-uncased](https://huggingface.co/allenai/scibert_scivocab_uncased).
|
| 21 |
-
The underlying citation embeddings are trained on the [S2ORC citation graph](https://github.com/allenai/s2orc).
|
| 22 |
-
|
| 23 |
-
Paper: [Neighborhood Contrastive Learning for Scientific Document Representations with Citation Embeddings (
|
| 24 |
-
|
| 25 |
-
Code: https://github.com/malteos/scincl
|
| 26 |
-
|
| 27 |
-
## How to use the pretrained model
|
| 28 |
-
|
| 29 |
-
```python
|
| 30 |
-
from transformers import AutoTokenizer, AutoModel
|
| 31 |
-
|
| 32 |
-
# load model and tokenizer
|
| 33 |
-
tokenizer = AutoTokenizer.from_pretrained('malteos/scincl')
|
| 34 |
-
model = AutoModel.from_pretrained('malteos/scincl')
|
| 35 |
-
|
| 36 |
-
papers = [{'title': 'BERT', 'abstract': 'We introduce a new language representation model called BERT'},
|
| 37 |
-
{'title': 'Attention is all you need', 'abstract': ' The dominant sequence transduction models are based on complex recurrent or convolutional neural networks'}]
|
| 38 |
-
|
| 39 |
-
# concatenate title and abstract with [SEP] token
|
| 40 |
-
title_abs = [d['title'] + tokenizer.sep_token + (d.get('abstract') or '') for d in papers]
|
| 41 |
-
|
| 42 |
-
# preprocess the input
|
| 43 |
-
inputs = tokenizer(title_abs, padding=True, truncation=True, return_tensors="pt", max_length=512)
|
| 44 |
-
|
| 45 |
-
# inference
|
| 46 |
-
result = model(**inputs)
|
| 47 |
-
|
| 48 |
-
# take the first token ([CLS] token) in the batch as the embedding
|
| 49 |
-
embeddings = result.last_hidden_state[:, 0, :]
|
| 50 |
-
```
|
| 51 |
-
|
| 52 |
-
## Triplet Mining Parameters
|
| 53 |
-
|
| 54 |
-
| **Setting** | **Value** |
|
| 55 |
-
|-------------------------|--------------------|
|
| 56 |
-
| seed | 4 |
|
| 57 |
-
| triples_per_query | 5 |
|
| 58 |
-
| easy_positives_count | 5 |
|
| 59 |
-
| easy_positives_strategy | 5 |
|
| 60 |
-
| easy_positives_k | 20-25 |
|
| 61 |
-
| easy_negatives_count | 3 |
|
| 62 |
-
| easy_negatives_strategy | random_without_knn |
|
| 63 |
-
| hard_negatives_count | 2 |
|
| 64 |
-
| hard_negatives_strategy | knn |
|
| 65 |
-
| hard_negatives_k | 3998-4000 |
|
| 66 |
-
|
| 67 |
-
## SciDocs Results
|
| 68 |
-
|
| 69 |
-
These model weights are the ones that yielded the best results on SciDocs (`seed=4`).
|
| 70 |
-
In the paper we report the SciDocs results as mean over ten seeds.
|
| 71 |
-
|
| 72 |
-
| **model** | **mag-f1** | **mesh-f1** | **co-view-map** | **co-view-ndcg** | **co-read-map** | **co-read-ndcg** | **cite-map** | **cite-ndcg** | **cocite-map** | **cocite-ndcg** | **recomm-ndcg** | **recomm-P@1** | **Avg** |
|
| 73 |
-
|-------------------|-----------:|------------:|----------------:|-----------------:|----------------:|-----------------:|-------------:|--------------:|---------------:|----------------:|----------------:|---------------:|--------:|
|
| 74 |
-
| Doc2Vec | 66.2 | 69.2 | 67.8 | 82.9 | 64.9 | 81.6 | 65.3 | 82.2 | 67.1 | 83.4 | 51.7 | 16.9 | 66.6 |
|
| 75 |
-
| fasttext-sum | 78.1 | 84.1 | 76.5 | 87.9 | 75.3 | 87.4 | 74.6 | 88.1 | 77.8 | 89.6 | 52.5 | 18 | 74.1 |
|
| 76 |
-
| SGC | 76.8 | 82.7 | 77.2 | 88 | 75.7 | 87.5 | 91.6 | 96.2 | 84.1 | 92.5 | 52.7 | 18.2 | 76.9 |
|
| 77 |
-
| SciBERT | 79.7 | 80.7 | 50.7 | 73.1 | 47.7 | 71.1 | 48.3 | 71.7 | 49.7 | 72.6 | 52.1 | 17.9 | 59.6 |
|
| 78 |
-
| SPECTER | 82 | 86.4 | 83.6 | 91.5 | 84.5 | 92.4 | 88.3 | 94.9 | 88.1 | 94.8 | 53.9 | 20 | 80 |
|
| 79 |
-
| SciNCL (10 seeds) | 81.4 | 88.7 | 85.3 | 92.3 | 87.5 | 93.9 | 93.6 | 97.3 | 91.6 | 96.4 | 53.9 | 19.3 | 81.8 |
|
| 80 |
-
| **SciNCL (seed=4)** | 81.2 | 89.0 | 85.3 | 92.2 | 87.7 | 94.0 | 93.6 | 97.4 | 91.7 | 96.5 | 54.3 | 19.6 | 81.9 |
|
| 81 |
-
|
| 82 |
-
Additional evaluations are available in the paper.
|
| 83 |
-
|
| 84 |
-
## License
|
| 85 |
-
|
| 86 |
-
MIT
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- feature-extraction
|
| 4 |
+
language: en
|
| 5 |
+
datasets:
|
| 6 |
+
- SciDocs
|
| 7 |
+
- s2orc
|
| 8 |
+
metrics:
|
| 9 |
+
- F1
|
| 10 |
+
- accuracy
|
| 11 |
+
- map
|
| 12 |
+
- ndcg
|
| 13 |
+
license: mit
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
## SciNCL
|
| 17 |
+
|
| 18 |
+
SciNCL is a pre-trained BERT language model to generate document-level embeddings of research papers.
|
| 19 |
+
It uses the citation graph neighborhood to generate samples for contrastive learning.
|
| 20 |
+
Prior to the contrastive training, the model is initialized with weights from [scibert-scivocab-uncased](https://huggingface.co/allenai/scibert_scivocab_uncased).
|
| 21 |
+
The underlying citation embeddings are trained on the [S2ORC citation graph](https://github.com/allenai/s2orc).
|
| 22 |
+
|
| 23 |
+
Paper: [Neighborhood Contrastive Learning for Scientific Document Representations with Citation Embeddings (EMNLP 2022 paper)](https://arxiv.org/abs/2202.06671).
|
| 24 |
+
|
| 25 |
+
Code: https://github.com/malteos/scincl
|
| 26 |
+
|
| 27 |
+
## How to use the pretrained model
|
| 28 |
+
|
| 29 |
+
```python
|
| 30 |
+
from transformers import AutoTokenizer, AutoModel
|
| 31 |
+
|
| 32 |
+
# load model and tokenizer
|
| 33 |
+
tokenizer = AutoTokenizer.from_pretrained('malteos/scincl')
|
| 34 |
+
model = AutoModel.from_pretrained('malteos/scincl')
|
| 35 |
+
|
| 36 |
+
papers = [{'title': 'BERT', 'abstract': 'We introduce a new language representation model called BERT'},
|
| 37 |
+
{'title': 'Attention is all you need', 'abstract': ' The dominant sequence transduction models are based on complex recurrent or convolutional neural networks'}]
|
| 38 |
+
|
| 39 |
+
# concatenate title and abstract with [SEP] token
|
| 40 |
+
title_abs = [d['title'] + tokenizer.sep_token + (d.get('abstract') or '') for d in papers]
|
| 41 |
+
|
| 42 |
+
# preprocess the input
|
| 43 |
+
inputs = tokenizer(title_abs, padding=True, truncation=True, return_tensors="pt", max_length=512)
|
| 44 |
+
|
| 45 |
+
# inference
|
| 46 |
+
result = model(**inputs)
|
| 47 |
+
|
| 48 |
+
# take the first token ([CLS] token) in the batch as the embedding
|
| 49 |
+
embeddings = result.last_hidden_state[:, 0, :]
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
## Triplet Mining Parameters
|
| 53 |
+
|
| 54 |
+
| **Setting** | **Value** |
|
| 55 |
+
|-------------------------|--------------------|
|
| 56 |
+
| seed | 4 |
|
| 57 |
+
| triples_per_query | 5 |
|
| 58 |
+
| easy_positives_count | 5 |
|
| 59 |
+
| easy_positives_strategy | 5 |
|
| 60 |
+
| easy_positives_k | 20-25 |
|
| 61 |
+
| easy_negatives_count | 3 |
|
| 62 |
+
| easy_negatives_strategy | random_without_knn |
|
| 63 |
+
| hard_negatives_count | 2 |
|
| 64 |
+
| hard_negatives_strategy | knn |
|
| 65 |
+
| hard_negatives_k | 3998-4000 |
|
| 66 |
+
|
| 67 |
+
## SciDocs Results
|
| 68 |
+
|
| 69 |
+
These model weights are the ones that yielded the best results on SciDocs (`seed=4`).
|
| 70 |
+
In the paper we report the SciDocs results as mean over ten seeds.
|
| 71 |
+
|
| 72 |
+
| **model** | **mag-f1** | **mesh-f1** | **co-view-map** | **co-view-ndcg** | **co-read-map** | **co-read-ndcg** | **cite-map** | **cite-ndcg** | **cocite-map** | **cocite-ndcg** | **recomm-ndcg** | **recomm-P@1** | **Avg** |
|
| 73 |
+
|-------------------|-----------:|------------:|----------------:|-----------------:|----------------:|-----------------:|-------------:|--------------:|---------------:|----------------:|----------------:|---------------:|--------:|
|
| 74 |
+
| Doc2Vec | 66.2 | 69.2 | 67.8 | 82.9 | 64.9 | 81.6 | 65.3 | 82.2 | 67.1 | 83.4 | 51.7 | 16.9 | 66.6 |
|
| 75 |
+
| fasttext-sum | 78.1 | 84.1 | 76.5 | 87.9 | 75.3 | 87.4 | 74.6 | 88.1 | 77.8 | 89.6 | 52.5 | 18 | 74.1 |
|
| 76 |
+
| SGC | 76.8 | 82.7 | 77.2 | 88 | 75.7 | 87.5 | 91.6 | 96.2 | 84.1 | 92.5 | 52.7 | 18.2 | 76.9 |
|
| 77 |
+
| SciBERT | 79.7 | 80.7 | 50.7 | 73.1 | 47.7 | 71.1 | 48.3 | 71.7 | 49.7 | 72.6 | 52.1 | 17.9 | 59.6 |
|
| 78 |
+
| SPECTER | 82 | 86.4 | 83.6 | 91.5 | 84.5 | 92.4 | 88.3 | 94.9 | 88.1 | 94.8 | 53.9 | 20 | 80 |
|
| 79 |
+
| SciNCL (10 seeds) | 81.4 | 88.7 | 85.3 | 92.3 | 87.5 | 93.9 | 93.6 | 97.3 | 91.6 | 96.4 | 53.9 | 19.3 | 81.8 |
|
| 80 |
+
| **SciNCL (seed=4)** | 81.2 | 89.0 | 85.3 | 92.2 | 87.7 | 94.0 | 93.6 | 97.4 | 91.7 | 96.5 | 54.3 | 19.6 | 81.9 |
|
| 81 |
+
|
| 82 |
+
Additional evaluations are available in the paper.
|
| 83 |
+
|
| 84 |
+
## License
|
| 85 |
+
|
| 86 |
+
MIT
|