eclair-transformer

Installation

Python 3.7 or higher is recommended:

pip install .

Dataset Arrangement

train

To train the model, arrange the data in the following structure:

{
	"train": \
	[
		CV1,
		CV2,
		CV3,
		...
	],
	"eval": \
	[
		CV1,
		CV2,
		CV3,
		...
	]
}

where each CV is the structure below (must contain the specified fields):

{
	"Qualification":"CONTENT",
	"Certification":"CONTENT",
	"Experience":"CONTENT",
	"JobProfile":"CONTENT",
	"t1_label":"T1_LABEL", // For T1 training
	"t2_label":"T2_LABEL", // For T2 training
	"job_description":"JOB_DESCRIPTION_CONTENT" // For T2 training
}

The recommended way of generating data in such structure is to generate each CV first in the following format with their resume ID as there name (12345.json for example):

{
	"Qualification":"CONTENT",
	"Certification":"CONTENT",
	"Experience":"CONTENT",
	"JobProfile":"CONTENT",
}

and then create a splits file in the following json format:

[
	{
		"resume_id":12345,
		"t1_label":"T1_LABEL", // If available
		"t2_label": "T2_LABEL", // If available
		"job_description": "JOB_DESCRIPTION_CONTENT" // If available
	}
	...
]

and merge the two fields to create resumes. For detailed implementation of data generation process as such please refer to demo_transformer.py

decode

To decode, arrange a single CV in as following:

{
	"Qualification":"CONTENT",
	"Certification":"CONTENT",
	"Experience":"CONTENT",
	"JobProfile":"CONTENT",
	"job_description": "JOB_DESCRIPTION" // Necessary for T2 Training
}

Initialization

Eclair-transformer need either a specified (task & encoder) or a specified model to work. For training purpose, both task and encoder have to be specified. For decoding purpose, the model to be used for decoding has to be specified.

• task: the task to run on the object. T1, or "t1" for T1 tasks, or multi-labeled competence level classification; T2, or "t2" for T2 tasks, or binary-labeled acceptance prediction.
• encoder: the type of tokenizer to use for training (or the type of the model for training or decoding). put a BERT structure tokenizer ("bert-large-cased", for example) if using a BERT model or a RoBERTa structure tokenizer ("roberta-large") if using a RoBERTa model.
• model_path: the path to the model file loaded upon initialization as the model for the object.
• output_dir: the directory to store the checkpoint files generated during the training process.

Train

The train() method takes in the training data and the development data and trains the model.

To train the model, initialize the eclair-transformer with specified task and encoder

task = T1
encoder = 'roberta-large'
transformer = ECLAIRTransformer(task, encoder)

and then call on the train() method

trn_data = "path/to/training/data"
dev_data = "path/to/development/data"
model_path = os.path.join('{}-model.torch'.format(task))
transformer.train(trn_data, dev_data)

Decode

To generate the prediction of a resume using the tool, call the decode() method.

To decode a resume, initialize the eclair-transformer with specified model

model_path = "path/to/model"
transformer = ECLAIRTransformer(model_path=model_path)

The method takes in one parameter:

• resume: the resume, in dictionary format, to be parsed.

resume = json.load(open("path/to/resume"))
res = transformer.decode(resume)
print(res)

It is expected that for T1 prediction the output should be something like:

[
  {
    "best_prediction": [
      [
        "CRCI",
        0.2707805931568146
      ]
    ],
    "predictions": {
      "NQ": 0.14438876509666443,
      "CRCI": 0.2707805931568146,
      "CRCII": 0.10479626059532166,
      "CRCIII": 0.2588427662849426,
      "CRCIV": 0.22119157016277313
    }
  }
]

And for T2 prediction:

[
  {
    "best_prediction": [
      [
        "YES",
        0.8009408712387085
      ]
    ],
    "predictions": {
      "NO": 0.1990591585636139,
      "YES": 0.8009408712387085
    }
  }
]