cdQA: Closed Domain Question Answering
An End-To-End Closed Domain Question Answering System. Built on top of the HuggingFace transformers library.
cdQA in details
If you are interested in understanding how the system works and its implementation, we wrote an article on Medium with a high-level explanation.
We also made a presentation during the #9 NLP Breakfast organised by Feedly. You can check it out here.
Table of Contents
Installation
With pip
pip install cdqa
From source
git clone https://github.com/cdqa-suite/cdQA.git
cd cdQA
pip install -e .
Hardware Requirements
Experiments have been done with:
- CPU ? AWS EC2
t2.mediumDeep Learning AMI (Ubuntu) Version 22.0 - GPU ? AWS EC2
p3.2xlargeDeep Learning AMI (Ubuntu) Version 22.0 + a single Tesla V100 16GB.
Getting started
Preparing your data
Manual
To use cdQA you need to create a pandas dataframe with the following columns:, title, paragraphs, -----------------, ------------------------------------------------------, The Article Title, [Paragraph 1 of Article, ... , Paragraph N of Article], #### With converters
The objective of cdqa converters is to make it easy to create this dataframe from your raw documents database. For instance the pdf_converter can create a cdqa dataframe from a directory containing .pdf files:
from cdqa.utils.converters import pdf_converter
df = pdf_converter(directory_path='path_to_pdf_folder')
You will need to install Java OpenJDK to use this converter. We currently have converters for:
- markdown
We plan to improve and add more converters in the future. Stay tuned!
Downloading pre-trained models and data
You can download the models and data manually from the GitHub releases or use our download functions:
from cdqa.utils.download import download_squad, download_model, download_bnpp_data
directory = 'path-to-directory'
# Downloading data
download_squad(dir=directory)
download_bnpp_data(dir=directory)
# Downloading pre-trained BERT fine-tuned on SQuAD 1.1
download_model('bert-squad_1.1', dir=directory)
# Downloading pre-trained DistilBERT fine-tuned on SQuAD 1.1
download_model('distilbert-squad_1.1', dir=directory)
Training models
Fit the pipeline on your corpus using the pre-trained reader:
import pandas as pd
from ast import literal_eval
from cdqa.pipeline import QAPipeline
df = pd.read_csv('your-custom-corpus-here.csv', converters={'paragraphs': literal_eval})
cdqa_pipeline = QAPipeline(reader='bert_qa.joblib') # use 'distilbert_qa.joblib' for DistilBERT instead of BERT
cdqa_pipeline.fit_retriever(df=df)
If you want to fine-tune the reader on your custom SQuAD-like annotated dataset:
cdqa_pipeline = QAPipeline(reader='bert_qa.joblib') # use 'distilbert_qa.joblib' for DistilBERT instead of BERT
cdqa_pipeline.fit_reader('path-to-custom-squad-like-dataset.json')
Save the reader model after fine-tuning:
cdqa_pipeline.dump_reader('path-to-save-bert-reader.joblib')
Making predictions
To get the best prediction given an input query:
cdqa_pipeline.predict(query='your question')
To get the N best predictions:
cdqa_pipeline.predict(query='your question', n_predictions=N)
There is also the possibility to change the weight of the retriever score
versus the reader score in the computation of final ranking score (the default is 0.35, which is shown to be the best weight on the development set of SQuAD 1.1-open)
cdqa_pipeline.predict(query='your question', retriever_score_weight=0.35)
Evaluating models
In order to evaluate models on your custom dataset you will need to annotate it. The annotation process can be done in 3 steps:
-
Convert your pandas DataFrame into a json file with SQuAD format:
from cdqa.utils.converters import df2squad json_data = df2squad(df=df, squad_version='v1.1', output_dir='.', filename='dataset-name') -
Use an annotator to add ground truth question-answer pairs:
Please refer to our
cdQA-annotator, a web-based annotator for closed-domain question answering datasets with SQuAD format. -
Evaluate the pipeline object:
from cdqa.utils.evaluation import evaluate_pipeline evaluate_pipeline(cdqa_pipeline, 'path-to-annotated-dataset.json') -
Evaluate the reader:
from cdqa.utils.evaluation import evaluate_reader evaluate_reader(cdqa_pipeline, 'path-to-annotated-dataset.json')
Notebook Examples
We prepared some notebook examples under the examples directory.
You can also play directly with these notebook examples using Binder or Google Colaboratory:, Notebook, Hardware, Platform, --------------------------------, ------------, ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------, [1] First steps with cdQA, CPU or GPU, , [2] Using the PDF converter, CPU or GPU, , [3] Training the reader on SQuAD, GPU, , Binder and Google Colaboratory provide temporary environments and may be slow to start but we recommend them if you want to get started with cdQA easily.
Deployment
Manual
You can deploy a cdQA REST API by executing:
export dataset_path=path-to-dataset.csv
export reader_path=path-to-reader-model
FLASK_APP=api.py flask run -h 0.0.0.0
You can now make requests to test your API (here using HTTPie):
http localhost:5000/api query=='your question here'
If you wish to serve a user interface on top of your cdQA system, follow the instructions of cdQA-ui, a web interface developed for cdQA.
Contributing
Read our Contributing Guidelines.
