File size: 8,540 Bytes
2ada614
7c05269
 
 
936ce3d
 
 
7c05269
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2ada614
936ce3d
7c05269
 
 
 
 
936ce3d
7c05269
 
 
936ce3d
7c05269
936ce3d
7c05269
936ce3d
7c05269
 
936ce3d
7c05269
936ce3d
88d4aa7
7c05269
 
 
 
936ce3d
 
 
 
 
 
7c05269
936ce3d
7c05269
936ce3d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7c05269
 
936ce3d
 
 
 
 
 
 
 
 
 
 
 
 
 
7c05269
936ce3d
7c05269
 
 
 
936ce3d
7c05269
936ce3d
7c05269
 
936ce3d
7c05269
936ce3d
7c05269
0fdc362
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
936ce3d
 
0fdc362
7c05269
 
 
 
 
 
 
 
936ce3d
7c05269
 
 
 
 
 
 
936ce3d
7c05269
 
 
 
 
 
 
 
 
 
 
 
 
 
936ce3d
7c05269
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
---
language:
- pt
thumbnail: Portuguese BERT for the Legal Domain
tags:
- sentence-transformers
- transformers
- bert
- pytorch
- sentence-similarity
license: mit
pipeline_tag: sentence-similarity
datasets:
- stjiris/portuguese-legal-sentences-v0
- assin
- assin2
- stsb_multi_mt
widget:
- source_sentence: "O advogado apresentou as provas ao juíz."
  sentences:
    - "O juíz leu as provas."
    - "O juíz leu o recurso."
    - "O juíz atirou uma pedra."
model-index:
- name: BERTimbau
  results:
  - task:
      name: STS
      type: STS
    metrics:
      - name: Pearson Correlation - assin Dataset
        type: Pearson Correlation
        value: 0.77274424197425
      - name: Pearson Correlation - assin2 Dataset
        type: Pearson Correlation
        value: 0.8121493150774828
      - name: Pearson Correlation - stsb_multi_mt pt Dataset
        type: Pearson Correlation
        value: 0.8499768588704257
      - name: Pearson Correlation - IRIS sts Dataset
        type: Pearson Correlation
        value: 0.8499768588704257
---

[![INESC-ID](https://www.inesc-id.pt/wp-content/uploads/2019/06/INESC-ID-logo_01.png)](https://www.inesc-id.pt/projects/PR07005/)

[![A Semantic Search System for Supremo Tribunal de Justiça](https://rufimelo99.github.io/SemanticSearchSystemForSTJ/_static/logo.png)](https://rufimelo99.github.io/SemanticSearchSystemForSTJ/)

Work developed as part of [Project IRIS](https://www.inesc-id.pt/projects/PR07005/).

Thesis: [A Semantic Search System for Supremo Tribunal de Justiça](https://rufimelo99.github.io/SemanticSearchSystemForSTJ/)

# stjiris/bert-large-portuguese-cased-legal-mlm-mkd-nli-sts-v1 (Legal BERTimbau)
This is a [sentence-transformers](https://www.SBERT.net) model: It maps sentences & paragraphs to a 1024 dimensional dense vector space and can be used for tasks like clustering or semantic search.
stjiris/bert-large-portuguese-cased-legal-mlm-mkd-nli-sts-v1 derives from stjiris/bert-large-portuguese-cased-legal-mlm (legal variant of [BERTimbau](https://huggingface.co/neuralmind/bert-large-portuguese-cased) large).

It was trained using the MLM technique with a learning rate 1e-5 [Legal Sentences from +-30000 documents](https://huggingface.co/datasets/stjiris/portuguese-legal-sentences-v1.0) 15000 training steps (best performance for our semantic search system implementation)

This model was subjected to Multilingual Knowledge Distillation technique (mkd). For the Multilingual Knowledge Distillation process, the teacher model was 'sentence-transformers/stsb-roberta-large', the supposed supported language as English and the language to learn was portuguese
The dataset used was: TED 2020 – Parallel Sentences Corpus. TED 2020 contains around 4000 TED and TED-X transcripts from July 2020. These transcripts were translated by volunteers into more than 100 languages, adding up to a total of 10 544 174 sentences. 

The model was presented to NLI data. 16 batch size,  2e-5 lr

It was trained for Semantic Textual Similarity, being submitted to a fine tuning stage with the [assin](https://huggingface.co/datasets/assin), [assin2](https://huggingface.co/datasets/assin2), [stsb_multi_mt pt](https://huggingface.co/datasets/stsb_multi_mt)  and [IRIS STS](https://huggingface.co/datasets/stjiris/IRIS_sts) datasets. 'lr': 1e-5


## Usage (Sentence-Transformers)
Using this model becomes easy when you have [sentence-transformers](https://www.SBERT.net) installed:
```
pip install -U sentence-transformers
```
Then you can use the model like this:
```python
from sentence_transformers import SentenceTransformer
sentences = ["Isto é um exemplo", "Isto é um outro exemplo"]

model = SentenceTransformer('stjiris/bert-large-portuguese-cased-legal-mlm-mkd-nli-sts-v1')
embeddings = model.encode(sentences)
print(embeddings)
```
## Usage (HuggingFace Transformers)
```python
from transformers import AutoTokenizer, AutoModel
import torch


#Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

# Sentences we want sentence embeddings for
sentences = ['This is an example sentence', 'Each sentence is converted']

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('stjiris/bert-large-portuguese-cased-legal-mlm-mkd-nli-sts-v1')
model = AutoModel.from_pretrained('stjiris/bert-large-portuguese-cased-legal-mlm-mkd-nli-sts-v1')

# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)
# Perform pooling. In this case, mean pooling.
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
print("Sentence embeddings:")
print(sentence_embeddings)
```


## Full Model Architecture
```
SentenceTransformer(
  (0): Transformer({'max_seq_length': 514, 'do_lower_case': False}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 1028, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False})
)
```
## Citing & Authors

### Contributions
[@rufimelo99](https://github.com/rufimelo99)

If you use this work, please cite:

```bibtex
@InProceedings{MeloSemantic,
  author="Melo, Rui
  and Santos, Pedro A.
  and Dias, Jo{\~a}o",
  editor="Moniz, Nuno
  and Vale, Zita
  and Cascalho, Jos{\'e}
  and Silva, Catarina
  and Sebasti{\~a}o, Raquel",
  title="A Semantic Search System for the Supremo Tribunal de Justi{\c{c}}a",
  booktitle="Progress in Artificial Intelligence",
  year="2023",
  publisher="Springer Nature Switzerland",
  address="Cham",
  pages="142--154",
  abstract="Many information retrieval systems use lexical approaches to retrieve information. Such approaches have multiple limitations, and these constraints are exacerbated when tied to specific domains, such as the legal one. Large language models, such as BERT, deeply understand a language and may overcome the limitations of older methodologies, such as BM25. This work investigated and developed a prototype of a Semantic Search System to assist the Supremo Tribunal de Justi{\c{c}}a (Portuguese Supreme Court of Justice) in its decision-making process. We built a Semantic Search System that uses specially trained BERT models (Legal-BERTimbau variants) and a Hybrid Search System that incorporates both lexical and semantic techniques by combining the capabilities of BM25 and the potential of Legal-BERTimbau. In this context, we obtained a {\$}{\$}335{\backslash}{\%}{\$}{\$}335{\%}increase on the discovery metric when compared to BM25 for the first query result. This work also provides information on the most relevant techniques for training a Large Language Model adapted to Portuguese jurisprudence and introduces a new technique of Metadata Knowledge Distillation.",
  isbn="978-3-031-49011-8"
}


@inproceedings{souza2020bertimbau,
  author    = {F{\'a}bio Souza and
               Rodrigo Nogueira and
               Roberto Lotufo},
  title     = {{BERT}imbau: pretrained {BERT} models for {B}razilian {P}ortuguese},
  booktitle = {9th Brazilian Conference on Intelligent Systems, {BRACIS}, Rio Grande do Sul, Brazil, October 20-23 (to appear)},
  year      = {2020}
}

@inproceedings{fonseca2016assin,
  title={ASSIN: Avaliacao de similaridade semantica e inferencia textual},
  author={Fonseca, E and Santos, L and Criscuolo, Marcelo and Aluisio, S},
  booktitle={Computational Processing of the Portuguese Language-12th International Conference, Tomar, Portugal},
  pages={13--15},
  year={2016}
}

@inproceedings{real2020assin,
  title={The assin 2 shared task: a quick overview},
  author={Real, Livy and Fonseca, Erick and Oliveira, Hugo Goncalo},
  booktitle={International Conference on Computational Processing of the Portuguese Language},
  pages={406--412},
  year={2020},
  organization={Springer}
}
@InProceedings{huggingface:dataset:stsb_multi_mt,
title = {Machine translated multilingual STS benchmark dataset.},
author={Philip May},
year={2021},
url={https://github.com/PhilipMay/stsb-multi-mt}
}

```