distilbert-depression-base

This model is a fine-tuned version of distilbert-base-uncased trained on CLPsych 2015 and evaluated on a scraped dataset from Twitter to detect potential users in Twitter for depression. It achieves the following results on the evaluation set:

• Evaluation Loss: 0.64
• Accuracy: 0.65
• F1: 0.70
• Precision: 0.61
• Recall: 0.83
• AUC: 0.65

Intended uses & limitations

Feed a corpus of tweets to the model to generate label if input is indicative of a depressed user or not. Label 1 is depressed, Label 0 is not depressed.

Limitation: All token sequences longer than 512 are automatically truncated. Also, training and test data may be contaminated with mislabeled users.

How to use

You can use this model directly with a pipeline for sentiment analysis:

>>> from transformers import DistilBertTokenizerFast, AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained('distilbert-base-uncased')
>>> from transformers import DistilBertForSequenceClassification
>>> model = DistilBertForSequenceClassification.from_pretrained(r"distilbert-depression-base")
>>> from transformers import pipeline
>>> classifier = pipeline("sentiment-analysis", model=model, tokenizer=tokenizer)
>>> tokenizer_kwargs = {'padding':True,'truncation':True,'max_length':512}
>>> result=classifier('pain peko',**tokenizer_kwargs) #For truncation to apply in the pipeline. 
>>> #Should note that the string passed as the input can be a corpus of tweets concatenated together into one document.

[{'label': 'LABEL_1', 'score': 0.5048992037773132}]

Otherwise, download the files and specify within the pipeline the path to the folder that contains the config.json, pytorch_model.bin, and training_args.bin

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 3.39e-05
• train_batch_size: 16
• eval_batch_size: 16
• weight_decay: 0.13
• num_epochs: 3.0

Training results