dpr-question_encoder-single-nq-base

Table of Contents

• Model Details
• How To Get Started With the Model
• Uses
• Risks, Limitations and Biases
• Training
• Evaluation
• Environmental Impact
• Technical Specifications
• Citation Information
• Model Card Authors

Model Details

Model Description: Dense Passage Retrieval (DPR) is a set of tools and models for state-of-the-art open-domain Q&A research. dpr-question_encoder-single-nq-base is the question encoder trained using the Natural Questions (NQ) dataset (Lee et al., 2019; Kwiatkowski et al., 2019).

• Developed by: See GitHub repo for model developers
• Model Type: BERT-based encoder
• Language(s): CC-BY-NC-4.0, also see Code of Conduct
• License: English
• Related Models:
  • dpr-ctx_encoder-single-nq-base
  • dpr-reader-single-nq-base
  • dpr-ctx_encoder-multiset-base
  • dpr-question_encoder-multiset-base
  • dpr-reader-multiset-base
• Resources for more information:
  • Research Paper
  • GitHub Repo
  • Hugging Face DPR docs
  • BERT Base Uncased Model Card

How to Get Started with the Model

Use the code below to get started with the model.
from transformers import DPRQuestionEncoder, DPRQuestionEncoderTokenizer
tokenizer = DPRQuestionEncoderTokenizer.from_pretrained("facebook/dpr-question_encoder-single-nq-base")
model = DPRQuestionEncoder.from_pretrained("facebook/dpr-question_encoder-single-nq-base")
input_ids = tokenizer("Hello, is my dog cute ?", return_tensors="pt")["input_ids"]
embeddings = model(input_ids).pooler_output


Uses

Direct Use

dpr-question_encoder-single-nq-base, dpr-ctx_encoder-single-nq-base, and dpr-reader-single-nq-base can be used for the task of open-domain question answering.

Misuse and Out-of-scope Use

The model should not be used to intentionally create hostile or alienating environments for people. In addition, the set of DPR models was not trained to be factual or true representations of people or events, and therefore using the models to generate such content is out-of-scope for the abilities of this model.

Risks, Limitations and Biases

CONTENT WARNING: Readers should be aware this section may contain content that is disturbing, offensive, and can propogate historical and current stereotypes.
Significant research has explored bias and fairness issues with language models (see, e.g., Sheng et al., 2021 and Bender et al., 2021). Predictions generated by the model can include disturbing and harmful stereotypes across protected classes; identity characteristics; and sensitive, social, and occupational groups.

Training

Training Data

This model was trained using the Natural Questions (NQ) dataset (Lee et al., 2019; Kwiatkowski et al., 2019). The model authors write that:

[The dataset] was designed for end-to-end question answering. The questions were mined from real Google search queries and the answers were spans in Wikipedia articles identified by annotators.

Training Procedure

The training procedure is described in the associated paper:

Given a collection of M text passages, the goal of our dense passage retriever (DPR) is to index all the passages in a low-dimensional and continuous space, such that it can retrieve efficiently the top k passages relevant to the input question for the reader at run-time.

Our dense passage retriever (DPR) uses a dense encoder EP(·) which maps any text passage to a d- dimensional real-valued vectors and builds an index for all the M passages that we will use for retrieval. At run-time, DPR applies a different encoder EQ(·) that maps the input question to a d-dimensional vector, and retrieves k passages of which vectors are the closest to the question vector.

The authors report that for encoders, they used two independent BERT (Devlin et al., 2019) networks (base, un-cased) and use FAISS (Johnson et al., 2017) during inference time to encode and index passages. See the paper for further details on training, including encoders, inference, positive and negative passages, and in-batch negatives.

Evaluation

The following evaluation information is extracted from the associated paper.

Testing Data, Factors and Metrics

The model developers report the performance of the model on five QA datasets, using the top-k accuracy (k ∈ {20, 100}). The datasets were NQ, TriviaQA, WebQuestions (WQ), CuratedTREC (TREC), and SQuAD v1.1.

Results

Top 20	Top 100
NQ	TriviaQA	WQ	TREC	SQuAD	NQ	TriviaQA	WQ	TREC	SQuAD
78.4	79.4	73.2	79.8	63.2	85.4	85.0	81.4	89.1	77.2