Skip to main content

Question Answering - transfomers

In short​

  • The Question and Answer (Q&A) module is a Weaviate module for answer extraction from data.
  • The module depends on a text vectorization module that should be running with Weaviate.
  • The module adds an ask {} operator to the GraphQL Get {} queries
  • The module returns a max. of 1 answer in the GraphQL _additional {} field.
  • The answer with the highest certainty (confidence level) will be returned.

Introduction​

The Question and Answer (Q&A) module is a Weaviate module for answer extraction from data. It uses BERT-related models for finding and extracting answers. This module can be used in GraphQL Get{...} queries, as a search operator. The qna-transformers module tries to find an answer in the data objects of the specified class. If an answer is found within the given certainty range, it will be returned in the GraphQL _additional { answer { ... } } field. There will be a maximum of 1 answer returned, if this is above the optionally set certainty. The answer with the highest certainty (confidence level) will be returned.

There are currently five different Question Answering models available (source: Hugging Face Model Hub): distilbert-base-uncased-distilled-squad (uncased), bert-large-uncased-whole-word-masking-finetuned-squad (uncased), distilbert-base-cased-distilled-squad (cased), deepset/roberta-base-squad2, and deepset/bert-large-uncased-whole-word-masking-squad2 (uncased). Note that not all models perform well on every dataset and use case. We recommend to use bert-large-uncased-whole-word-masking-finetuned-squad (uncased), which performs best on most datasets (although it's quite heavyweighted).

Starting with v1.10.0, the answer score can be used as a reranking factor for the search results.

How to enable (module configuration)​

Docker Compose​

The Q&A module can be added as a service to the Docker Compose file. You must have a text vectorizer like text2vec-contextionary or text2vec-transformers running. An example Docker Compose file for using the qna-transformers module (bert-large-uncased-whole-word-masking-finetuned-squad (uncased)) in combination with the text2vec-transformersis as follows:

---
version: '3.4'
services:
  weaviate:
    command:
    - --host
    - 0.0.0.0
    - --port
    - '8080'
    - --scheme
    - http
    image: semitechnologies/weaviate:1.22.5
    ports:
    - 8080:8080
    restart: on-failure:0
    environment:
      TRANSFORMERS_INFERENCE_API: 'http://t2v-transformers:8080'
      QNA_INFERENCE_API: "http://qna-transformers:8080"
      QUERY_DEFAULTS_LIMIT: 25
      AUTHENTICATION_ANONYMOUS_ACCESS_ENABLED: 'true'
      PERSISTENCE_DATA_PATH: '/var/lib/weaviate'
      DEFAULT_VECTORIZER_MODULE: 'text2vec-transformers'
      ENABLE_MODULES: 'text2vec-transformers,qna-transformers'
      CLUSTER_HOSTNAME: 'node1'
  t2v-transformers:
    image: semitechnologies/transformers-inference:sentence-transformers-msmarco-distilbert-base-v2
    environment:
      ENABLE_CUDA: '1'
      NVIDIA_VISIBLE_DEVICES: all
    deploy:
      resources:
        reservations:
          devices:
          - capabilities: [gpu]
  qna-transformers:
    image: semitechnologies/qna-transformers:bert-large-uncased-whole-word-masking-finetuned-squad
    environment:
      ENABLE_CUDA: '1'
      NVIDIA_VISIBLE_DEVICES: all
    deploy:
      resources:
        reservations:
          devices:
          - capabilities: [gpu]
...

Variable explanations:

  • QNA_INFERENCE_API: where the qna module is running
  • ENABLE_CUDA: if set to 1 it uses GPU (if available on the host machine)

Note: at the moment, text vectorization modules cannot be combined in a single setup. This means that you can either enable the text2vec-contextionary, the text2vec-transformers or no text vectorization module.

How to use (GraphQL)​

This module adds a search operator to GraphQL Get{...} queries: ask{}. This new operator takes the following arguments:

FieldData TypeRequiredExample valueDescription
questionstringyes"What is the name of the Dutch king?"The question to be answered.
certaintyfloatno0.75Desired minimal certainty or confidence of answer to the question. The higher the value, the stricter the search becomes. The lower the value, the fuzzier the search becomes. If no certainty is set, any answer that could be extracted will be returned
propertieslist of stringsno["summary"]The properties of the queries Class which contains text. If no properties are set, all are considered.
rerankboolnotrueIf enabled, the qna module will rerank the result based on the answer score. For example, if the 3rd result - as determined by the previous (semantic) search contained the most likely answer, result 3 will be pushed to position 1, etc. Not supported prior to v1.10.0

Notes:

  • The GraphQL Explore { } function does support the ask searcher, but the result is only a beacon to the object containing the answer. It is thus not any different from performing a nearText semantic search with the question. No extraction is happening.
  • You cannot use the 'ask' operator along with a 'nearXXX' operator!

Example query​

import weaviate
import os

client = weaviate.Client(
    "https://edu-demo.weaviate.network",
    auth_client_secret=weaviate.AuthApiKey("learn-weaviate"),
    additional_headers={
        "X-OpenAI-Api-Key": os.environ["OPENAI_API_KEY"]  # Replace w/ your OPENAI API key
    }
)

ask = {
  "question": "Who is the king of the Netherlands?",
  "properties": ["summary"]
}

result = (
  client.query
  .get("Article", ["title", "_additional {answer {hasAnswer property result startPosition endPosition} }"])
  .with_ask(ask)
  .with_limit(1)
  .do()
)

print(result)

GraphQL response​

The answer is contained in a new GraphQL _additional property called answer. It contains the following fields:

  • hasAnswer (boolean): could an answer be found?
  • result (nullable string): An answer if one could be found. null if hasAnswer==false
  • certainty (nullable float): The certainty of the answer returned. null if hasAnswer==false
  • property (nullable string): The property which contains the answer. null if hasAnswer==false
  • startPosition (int): The character offset where the answer starts. 0 if hasAnswer==false
  • endPosition (int): The character offset where the answer ends 0 if hasAnswer==false

Note: startPosition, endPosition and property in the response are not guaranteed to be present. They are calculated by a case-insensitive string matching function against the input text. If the transformer model formats the output differently (e.g. by introducing spaces between tokens which were not present in the original input), the calculation of the position and determining the property fails.

Example response​

{
  "data": {
    "Get": {
      "Article": [
        {
          "_additional": {
            "answer": {
              "certainty": 0.73,
              "endPosition": 26,
              "hasAnswer": true,
              "property": "summary",
              "result": "king willem - alexander",
              "startPosition": 48
            }
          },
          "title": "Bruised Oranges - The Dutch royals are botching covid-19 etiquette"
        }
      ]
    }
  },
  "errors": null
}

Custom Q&A Transformer module​

You can use the same approach as for text2vec-transformers, see here, i.e. either pick one of the pre-built containers or build your own container from your own model using the semitechnologies/qna-transformers:custom base image. Make sure that your model is compatible with Hugging Face's transformers.AutoModelForQuestionAnswering.

How it works (under the hood)​

Under the hood, the model uses a two-step approach. First it performs a semantic search to find the documents (e.g. a Sentence, Paragraph, Article, etc.) most likely to contain the answer. In a second step, a BERT-style answer extraction is performed on all text and string properties of the document. There are now three possible outcomes:

  1. No answer was found because the question can not be answered,
  2. An answer was found, but did not meet the user-specified minimum certainty, so it was discarded (typically the case when the document is on topic, but does not contain an actual answer to the question), and
  3. An answer was found that matches the desired certainty. It is returned to the user.

The module performs a semantic search under the hood, so a text2vec-... module is required. It does not need to be of the same type as the qna-... module. For example, you can use a text2vec-contextionary module to perform the semantic search, and a qna-transformers module to extract the answer.

Automatic sliding window for long documents​

If a text value in a data object is longer than 512 tokens, the Q&A Transformer module automatically splits the text into smaller texts. The module uses a sliding window, i.e. overlapping pieces of text, to avoid a scenario that an answer cannot be found if it lies on a boundary. If an answer lies on the boundary, the Q&A module returns the result (answer) with the highest score (as the sliding mechanism could lead to duplicates).

Model license(s)​

The qna-transformers module is compatible with various models, each with their own license. For detailed information, please review the license of the model you are using in the Hugging Face Hub.

It is your responsibility to evaluate whether the terms of its license(s), if any, are appropriate for your intended use.