96abhishekarora's picture
Updated model with better training and evaluation. Test and val data included as pickle files. Older Legacy files were removed to avoid confusion.
b369926
|
raw
history blame
5.97 kB
---
pipeline_tag: sentence-similarity
language:
- en
- fr
- es
tags:
- linktransformer
- sentence-transformers
- sentence-similarity
- tabular-classification
---
# {MODEL_NAME}
This is a [LinkTransformer](https://linktransformer.github.io/) model. At its core this model this is a sentence transformer model [sentence-transformers](https://www.SBERT.net) model- it just wraps around the class.
It is designed for quick and easy record linkage (entity-matching) through the LinkTransformer package. The tasks include clustering, deduplication, linking, aggregation and more.
Notwithstanding that, it can be used for any sentence similarity task within the sentence-transformers framework as well.
It maps sentences & paragraphs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search.
Take a look at the documentation of [sentence-transformers](https://www.sbert.net/index.html) if you want to use this model for more than what we support in our applications.
This model has been fine-tuned on the model : sentence-transformers/paraphrase-multilingual-mpnet-base-v2. It is pretrained for the language : - en
- fr
- es.
This model was trained on a dataset prepared by linking product classifications from [UN stats](https://unstats.un.org/unsd/classifications/Econ).
This model is designed to link different products to their industrial classification (ISIC) - trained on variation brought on by product level correspondance. It was trained for 100 epochs using other defaults that can be found in the repo's LinkTransformer config file - LT_training_config.json
## Usage (LinkTransformer)
Using this model becomes easy when you have [LinkTransformer](https://github.com/dell-research-harvard/linktransformer) installed:
```
pip install -U linktransformer
```
Then you can use the model like this:
```python
import linktransformer as lt
import pandas as pd
##Load the two dataframes that you want to link. For example, 2 dataframes with company names that are written differently
df1=pd.read_csv("data/df1.csv") ###This is the left dataframe with key CompanyName for instance
df2=pd.read_csv("data/df2.csv") ###This is the right dataframe with key CompanyName for instance
###Merge the two dataframes on the key column!
df_merged = lt.merge(df1, df2, on="CompanyName", how="inner")
##Done! The merged dataframe has a column called "score" that contains the similarity score between the two company names
```
## Training your own LinkTransformer model
Any Sentence Transformers can be used as a backbone by simply adding a pooling layer. Any other transformer on HuggingFace can also be used by specifying the option add_pooling_layer==True
The model was trained using SupCon loss.
Usage can be found in the package docs.
The training config can be found in the repo with the name LT_training_config.json
To replicate the training, you can download the file and specify the path in the config_path argument of the training function. You can also override the config by specifying the training_args argument.
Here is an example.
```python
##Consider the example in the paper that has a dataset of Mexican products and their tariff codes from 1947 and 1948 and we want train a model to link the two tariff codes.
saved_model_path = train_model(
model_path="hiiamsid/sentence_similarity_spanish_es",
dataset_path=dataset_path,
left_col_names=["description47"],
right_col_names=['description48'],
left_id_name=['tariffcode47'],
right_id_name=['tariffcode48'],
log_wandb=False,
config_path=LINKAGE_CONFIG_PATH,
training_args={"num_epochs": 1}
)
```
You can also use this package for deduplication (clusters a df on the supplied key column). Merging a fine class (like product) to a coarse class (like HS code) is also possible.
Read our paper and the documentation for more!
## Evaluation Results
<!--- Describe how your model was evaluated -->
You can evaluate the model using the [LinkTransformer](https://github.com/dell-research-harvard/linktransformer) package's inference functions.
We have provided a few datasets in the package for you to try out. We plan to host more datasets on Huggingface and our website (Coming soon) that you can take a look at.
## Training
The model was trained with the parameters:
**DataLoader**:
`torch.utils.data.dataloader.DataLoader` of length 281 with parameters:
```
{'batch_size': 64, 'sampler': 'torch.utils.data.dataloader._InfiniteConstantSampler', 'batch_sampler': 'torch.utils.data.sampler.BatchSampler'}
```
**Loss**:
`linktransformer.modified_sbert.losses.SupConLoss_wandb`
Parameters of the fit()-Method:
```
{
"epochs": 100,
"evaluation_steps": 141,
"evaluator": "sentence_transformers.evaluation.SequentialEvaluator.SequentialEvaluator",
"max_grad_norm": 1,
"optimizer_class": "<class 'torch.optim.adamw.AdamW'>",
"optimizer_params": {
"lr": 2e-05
},
"scheduler": "WarmupLinear",
"steps_per_epoch": null,
"warmup_steps": 28100,
"weight_decay": 0.01
}
```
LinkTransformer(
(0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: XLMRobertaModel
(1): Pooling({'word_embedding_dimension': 768, '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
```
@misc{arora2023linktransformer,
title={LinkTransformer: A Unified Package for Record Linkage with Transformer Language Models},
author={Abhishek Arora and Melissa Dell},
year={2023},
eprint={2309.00789},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
```