File size: 5,732 Bytes
1c9c7eb
 
 
 
 
 
 
 
 
 
 
 
034082e
1c9c7eb
594f96a
1c9c7eb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
594f96a
1c9c7eb
 
 
 
 
 
 
 
 
 
 
 
594f96a
1c9c7eb
 
 
 
594f96a
1c9c7eb
 
 
594f96a
1c9c7eb
 
 
 
 
 
 
 
 
594f96a
1c9c7eb
 
 
 
 
594f96a
 
 
 
 
 
 
 
 
 
 
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
---
pipeline_tag: sentence-similarity
language: 
- ja
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 : oshizo/sbert-jsnli-luke-japanese-base-lite. It is pretrained for the language : - ja.


This model was trained on a dataset consisting of company aliases from wiki data using the LinkTransformer framework. 
                               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 562 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": 281,
    "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": 56200,
    "weight_decay": 0.01
}
```




LinkTransformer(
  (0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: LukeModel 
  (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}
                }

```