sentence similarity passage retrieval

Dense Passage Retrieval-DPR is a set of tools for performing State of the Art open-domain question answering. It was initially developed by Facebook and there is an official repository. DPR is intended to retrieve the relevant documents to answer a given question, and is composed of 2 models, one for encoding passages and other for encoding questions. This concrete model is the one used for encoding passages.

With this and the question encoder model we introduce the best passage retrievers in Spanish up to date (to the best of our knowledge), improving over the previous model we developed, by training it for longer and with more data.

Regarding its use, this model should be used to vectorize a question that enters in a Question Answering system, and then we compare that encoding with the encodings of the database (encoded with the passage encoder) to find the most similar documents , which then should be used for either extracting the answer or generating it.

For training the model, we used a collection of Question Answering datasets in Spanish:

With this complete dataset we created positive and negative examples for the model (For more information look at the paper to understand the training process for DPR). We trained for 25 epochs with the same configuration as the paper. The previous DPR model was trained for only 3 epochs with about 60% of the data.

Example of use:

from transformers import DPRContextEncoder, DPRContextEncoderTokenizer

model_str = "IIC/dpr-spanish-passage_encoder-allqa-base"
tokenizer = DPRContextEncoderTokenizer.from_pretrained(model_str)
model = DPRContextEncoder.from_pretrained(model_str)

input_ids = tokenizer("Usain Bolt ganó varias medallas de oro en las Olimpiadas del año 2012", return_tensors="pt")["input_ids"]
embeddings = model(input_ids).pooler_output

The full metrics of this model on the evaluation split of SQUADES are:

eval_loss: 0.010779764448327261
eval_acc: 0.9982682224158297
eval_f1: 0.9446059155411182
eval_acc_and_f1: 0.9714370689784739
eval_average_rank: 0.11728500598392888

And the classification report:

                precision    recall  f1-score   support

hard_negative     0.9991    0.9991    0.9991   1104999
     positive     0.9446    0.9446    0.9446     17547

     accuracy                         0.9983   1122546
    macro avg     0.9719    0.9719    0.9719   1122546
 weighted avg     0.9983    0.9983    0.9983   1122546

Contributions

Thanks to @avacaondata, @alborotis, @albarji, @Dabs, @GuillemGSubies for adding this model.