Skip to main content

NeuralQA: Question Answering on Large Datasets

Project description

NeuralQA: A Usable Library for (Extractive) Question Answering on Large Datasets with BERT

License: MIT docs

Still in alpha, lots of changes anticipated. View demo on neuralqa.fastforwardlabs.com.

NeuralQA provides an easy to use api and visual interface for Extractive Question Answering (QA), on large datasets. The QA process is comprised of two main stages - Passage retrieval (Retriever) is implemented using ElasticSearch and Document Reading (Reader) is implemented using pretrained BERT models via the Huggingface Transformers api.

Usage

pip3 install neuralqa

Create (or navigate to) a folder you would like to use with NeuralQA. Run the following command line instruction within that folder.

neuralqa ui --port 4000

navigate to http://localhost:4000/#/ to view the NeuralQA interface. Learn about other command line options in the documentation here or how to configure NeuralQA to use your own reader models or retriever instances.

Note: To use NeuralQA with a retriever such as ElasticSearch, follow the instructions here to download, install, and launch a local elasticsearch instance and add it to your config.yaml file.

How Does it Work?

NeuralQA is comprised of several high level modules:

  • Retriever: For each search query (question), scan an index (elasticsearch), and retrieve a list of candidate matched passages.

  • Reader: For each retrieved passage, a BERT based model predicts a span that contains the answer to the question. In practice, retrieved passages may be lengthy and BERT based models can process a maximum of 512 tokens at a time. NeuralQA handles this in two ways. Lengthy passages are chunked into smaller sections with a configurable stride. Secondly, NeuralQA offers the option of extracting a subset of relevant snippets (RelSnip) which a BERT reader can then scan to find answers. Relevant snippets are portions of the retrieved document that contain exact match results for the search query.

  • Expander: Methods for generating additional (relevant) query terms to improve recall. Currently, we implement Contextual Query Expansion using finetuned Masked Language Models. This is implemented via a user in the loop flow where the user can choose to include any suggested expansion terms.

  • User Interface: NeuralQA provides a visual user interface for performing queries (manual queries where question and context are provided as well as queries over a search index), viewing results and also sensemaking of results (reranking of passages based on answer scores, highlighting keyword match, model explanations).

Configuration

Properties of modules within NeuralQA (ui, retriever, reader, expander) can be specified via a yaml configuration file. When you launch the ui, you can specify the path to your config file --config-path. If this is not provided, NeuralQA will search for a config.yaml in the current folder or create a default copy) in the current folder. Sample configuration shown below:

ui:
  queryview:
    intro:
      title: "NeuralQA: Question Answering on Large Datasets"
      subtitle: "Subtitle of your choice"
    views: # select sections of the ui to hide or show
      intro: True
      advanced: True
      samples: False
      passages: True
      explanations: True
      allanswers: True
    options: # values for advanced options
      stride: ..
      maxpassages: ..
      highlightspan: ..

  header: # header tile for ui
    appname: NeuralQA
    appdescription: Question Answering on Large Datasets

reader:
  title: Reader
  selected: twmkn9/distilbert-base-uncased-squad2
  options:
    - name: DistilBERT SQUAD2
      value: twmkn9/distilbert-base-uncased-squad2
      type: distilbert
    - name: BERT SQUAD2
      value: deepset/bert-base-cased-squad2
      type: bert

Documentation

An attempt is being made to better document NeuralQA here - https://victordibia.github.io/neuralqa/.

Citation

A paper introducing NeuralQA and its components can be found here.

@article{dibia2020neuralqa,
    title={NeuralQA: A Usable Library for Question Answering (Contextual Query Expansion + BERT) on Large Datasets},
    author={Victor Dibia},
    year={2020},
    journal={arXiv preprint arXiv:2007.15211}
}

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

neuralqa-0.0.31a0.tar.gz (629.9 kB view details)

Uploaded Source

Built Distribution

neuralqa-0.0.31a0-py3-none-any.whl (640.7 kB view details)

Uploaded Python 3

File details

Details for the file neuralqa-0.0.31a0.tar.gz.

File metadata

  • Download URL: neuralqa-0.0.31a0.tar.gz
  • Upload date:
  • Size: 629.9 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/3.2.0 pkginfo/1.5.0.1 requests/2.24.0 setuptools/49.2.0 requests-toolbelt/0.9.1 tqdm/4.48.0 CPython/3.8.5

File hashes

Hashes for neuralqa-0.0.31a0.tar.gz
Algorithm Hash digest
SHA256 d35fe87f930c1bbbdb61a413bf7309dcd7214dd767c846f7c8a78a39c9bdaf4c
MD5 8436709d8bd8ced85a316e2643c17c9f
BLAKE2b-256 4a6554098aa8030198cab09666ed26ce0a7745a454a0c44490a3c619b4084db9

See more details on using hashes here.

File details

Details for the file neuralqa-0.0.31a0-py3-none-any.whl.

File metadata

  • Download URL: neuralqa-0.0.31a0-py3-none-any.whl
  • Upload date:
  • Size: 640.7 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/3.2.0 pkginfo/1.5.0.1 requests/2.24.0 setuptools/49.2.0 requests-toolbelt/0.9.1 tqdm/4.48.0 CPython/3.8.5

File hashes

Hashes for neuralqa-0.0.31a0-py3-none-any.whl
Algorithm Hash digest
SHA256 c16661353228e40bae73bd7876d37216935684dc283203c4350c281cee15158a
MD5 ed381b5cd3d5d452901307552f5e1dd7
BLAKE2b-256 a384c1fb8742427faeae59098c1729a4d28565e4bf5e240e1f6a336c0be1c46d

See more details on using hashes here.

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page