LTR_CrossEncoder: Legal Text Retrieval Zalo AI Challenge 2021
We propose a cross encoder model (LTR_CrossEncoder) for information retrieval, re-retrieval text relevant base on result of elasticsearch
- Model achieved 0.747 F2 score in public test (Legal Text Retrieval Zalo AI Challenge 2021)
- If using elasticsearch only, our F2 score is 0.54
Algorithm design
Our algorithm includes two key components:
- Elasticsearch
- Cross Encoder Model
Elasticsearch
Elasticsearch is used for filtering top-k most relevant articles based on BM25 score.
Cross Encoder Model
Our model accepts query, article text (passage) and article title as inputs and outputs a relevant score of that query and that article. Higher score, more relavant. We use pretrained vinai/phobert-base and CrossEntropyLoss or BCELoss as loss function
Train dataset
Non-relevant samples in dataset are obtained by top-10 result of elasticsearch, the training data (train_data_model.json) has format as follow:
[
{
"question_id": "..."
"question": "..."
"relevant_articles":[
{
"law_id": "..."
"article_id": "..."
"title": "..."
"text": "..."
},
...
]
"non_relevant_articles":[
{
"law_id": "..."
"article_id": "..."
"title": "..."
"text": "..."
},
...
]
},
...
]
Test dataset
First we use elasticsearch to obtain k relevant candidates (k=top-50 result of elasticsearch), then LTR_CrossEncoder classify which actual relevant article. The test data (test_data_model.json) has format as follow:
[
{
"question_id": "..."
"question": "..."
"articles":[
{
"law_id": "..."
"article_id": "..."
"title": "..."
"text": "..."
},
...
]
},
...
]
Training
Run the following bash file to train model:
bash run_phobert.sh
Inference
We also provide model checkpoints. Please download these checkpoints if you want to make inference on a new text file without training the models from scratch. Create new checkpoint folder, unzip model file and push it in checkpoint folder. https://drive.google.com/file/d/1oT8nlDIAatx3XONN1n5eOgYTT6Lx_h_C/view?usp=sharing
Run the following bash file to infer test dataset:
bash run_predict.sh
![[arXiv22] Disentangled Representation Learning for Text-Video Retrieval](https://github.com/foolwood/DRL/raw/main/demo/pipeline.png)
49 Dec 18, 2022
25 Oct 30, 2022
38 Oct 18, 2022
99 Dec 27, 2022
24 May 30, 2022
![[2021 MultiMedia] CONQUER: Contextual Query-aware Ranking for Video Corpus Moment Retrieval](https://github.com/houzhijian/CONQUER/raw/master/figures/problem_definition.png)
23 Dec 26, 2022
42 Nov 17, 2022
5 Nov 25, 2022
![Joint Versus Independent Multiview Hashing for Cross-View Retrieval[J] (IEEE TCYB 2021, PyTorch Code)](https://github.com/penghu-cs/DCHN/raw/main/paper/DCHN.jpg)
4 Nov 19, 2022
105 Dec 16, 2022
128 Dec 27, 2022
82 Dec 29, 2022
47 Dec 16, 2022
39 Jan 7, 2023
95 Dec 8, 2022
48 Dec 30, 2022
61 Nov 14, 2022
54 Dec 8, 2022
8 Oct 7, 2022