This is a PyTorch implementation of the DrQA system described in the ACL 2017 paper Reading Wikipedia to Answer Open-Domain Questions.
DrQA is a system for reading comprehension applied to open-domain question answering. In particular, DrQA is targeted at the task of "machine reading at scale" (MRS). In this setting, we are searching for an answer to a question in a potentially very large corpus of unstructured documents (that may not be redundant). Thus the system has to combine the challenges of document retrieval (finding the relevant documents) with that of machine comprehension of text (identifying the answers from those documents).
Our experiments with DrQA focus on answering factoid questions while using Wikipedia as the unique knowledge source for documents. Wikipedia is a well-suited source of large-scale, rich, detailed information. In order to answer any question, one must first retrieve the few potentially relevant articles among more than 5 million, and then scan them carefully to identify the answer.
Note that DrQA treats Wikipedia as a generic collection of articles and does not rely on its internal graph structure. As a result, DrQA can be straightforwardly applied to any collection of documents, as described in the retriever README.
This repository includes code, data, and pre-trained models for processing and querying Wikipedia as described in the paper -- see Trained Models and Data. We also list several different datasets for evaluation, see QA Datasets. Note that this work is a refactored and more efficient version of the original code. Reproduction numbers are very similar but not exact.
Install DrQA and download our models to start asking open-domain questions!
Run python scripts/pipeline/interactive.py
to drop into an interactive session. For each question, the top span and the Wikipedia paragraph it came from are returned.
>>> process('What is question answering?')
Top Predictions:
+------+----------------------------------------------------------------------------------------------------------+--------------------+--------------+-----------+
| Rank | Answer | Doc | Answer Score | Doc Score |
+------+----------------------------------------------------------------------------------------------------------+--------------------+--------------+-----------+
| 1 | a computer science discipline within the fields of information retrieval and natural language processing | Question answering | 1917.8 | 327.89 |
+------+----------------------------------------------------------------------------------------------------------+--------------------+--------------+-----------+
Contexts:
[ Doc = Question answering ]
Question Answering (QA) is a computer science discipline within the fields of
information retrieval and natural language processing (NLP), which is
concerned with building systems that automatically answer questions posed by
humans in a natural language.
>>> process('What is the answer to life, the universe, and everything?')
Top Predictions:
+------+--------+---------------------------------------------------+--------------+-----------+
| Rank | Answer | Doc | Answer Score | Doc Score |
+------+--------+---------------------------------------------------+--------------+-----------+
| 1 | 42 | Phrases from The Hitchhiker's Guide to the Galaxy | 47242 | 141.26 |
+------+--------+---------------------------------------------------+--------------+-----------+
Contexts:
[ Doc = Phrases from The Hitchhiker's Guide to the Galaxy ]
The number 42 and the phrase, "Life, the universe, and everything" have
attained cult status on the Internet. "Life, the universe, and everything" is
a common name for the off-topic section of an Internet forum and the phrase is
invoked in similar ways to mean "anything at all". Many chatbots, when asked
about the meaning of life, will answer "42". Several online calculators are
also programmed with the Question. Google Calculator will give the result to
"the answer to life the universe and everything" as 42, as will Wolfram's
Computational Knowledge Engine. Similarly, DuckDuckGo also gives the result of
"the answer to the ultimate question of life, the universe and everything" as
42. In the online community Second Life, there is a section on a sim called
43. "42nd Life." It is devoted to this concept in the book series, and several
attempts at recreating Milliways, the Restaurant at the End of the Universe, were made.
>>> process('Who was the winning pitcher in the 1956 World Series?')
Top Predictions:
+------+------------+------------------+--------------+-----------+
| Rank | Answer | Doc | Answer Score | Doc Score |
+------+------------+------------------+--------------+-----------+
| 1 | Don Larsen | New York Yankees | 4.5059e+06 | 278.06 |
+------+------------+------------------+--------------+-----------+
Contexts:
[ Doc = New York Yankees ]
In 1954, the Yankees won over 100 games, but the Indians took the pennant with
an AL record 111 wins; 1954 was famously referred to as "The Year the Yankees
Lost the Pennant". In , the Dodgers finally beat the Yankees in the World
Series, after five previous Series losses to them, but the Yankees came back
strong the next year. On October 8, 1956, in Game Five of the 1956 World
Series against the Dodgers, pitcher Don Larsen threw the only perfect game in
World Series history, which remains the only perfect game in postseason play
and was the only no-hitter of any kind to be pitched in postseason play until
Roy Halladay pitched a no-hitter on October 6, 2010.
Try some of your own! Of course, DrQA might provide alternative facts, so enjoy the ride.
Setting up DrQA is easy!
DrQA requires Linux/OSX and Python 3.5 or higher. It also requires installing PyTorch version 1.0. Its other dependencies are listed in requirements.txt. CUDA is strongly recommended for speed, but not necessary.
Run the following commands to clone the repository and install DrQA:
git clone https://github.com/facebookresearch/DrQA.git
cd DrQA; pip install -r requirements.txt; python setup.py develop
Note: requirements.txt includes a subset of all the possible required packages. Depending on what you want to run, you might need to install an extra package (e.g. spacy).
If you use the CoreNLPTokenizer or SpacyTokenizer you also need to download the Stanford CoreNLP jars and spaCy en
model, respectively. If you use Stanford CoreNLP, have the jars in your java CLASSPATH
environment variable, or set the path programmatically with:
import drqa.tokenizers
drqa.tokenizers.set_default('corenlp_classpath', '/your/corenlp/classpath/*')
IMPORTANT: The default tokenizer is CoreNLP so you will need that in your CLASSPATH
to run the README examples.
Ex: export CLASSPATH=$CLASSPATH:/path/to/corenlp/download/*
.
If you do not already have a CoreNLP download you can run:
./install_corenlp.sh
Verify that it runs:
from drqa.tokenizers import CoreNLPTokenizer
tok = CoreNLPTokenizer()
tok.tokenize('hello world').words() # Should complete immediately
For convenience, the Document Reader, Retriever, and Pipeline modules will try to load default models if no model argument is given. See below for downloading these models.
To download all provided trained models and data for Wikipedia question answering, run:
./download.sh
Warning: this downloads a 7.5GB tarball (25GB untarred) and will take some time.
This stores the data in data/
at the file paths specified in the various modules' defaults. This top-level directory can be modified by setting a DRQA_DATA
environment variable to point to somewhere else.
Default directory structure (see embeddings for more info on additional downloads for training):
DrQA
├── data (or $DRQA_DATA)
├── datasets
│ ├── SQuAD-v1.1-<train/dev>.<txt/json>
│ ├── WebQuestions-<train/test>.txt
│ ├── freebase-entities.txt
│ ├── CuratedTrec-<train/test>.txt
│ └── WikiMovies-<train/test/entities>.txt
├── reader
│ ├── multitask.mdl
│ └── single.mdl
└── wikipedia
├── docs.db
└── docs-tfidf-ngram=2-hash=16777216-tokenizer=simple.npz
Default model paths for the different modules can also be modified programmatically in the code, e.g.:
import drqa.reader
drqa.reader.set_default('model', '/path/to/model')
reader = drqa.reader.Predictor() # Default model loaded for prediction
TF-IDF model using Wikipedia (unigrams and bigrams, 2^24 bins, simple tokenization), evaluated on multiple datasets (test sets, dev set for SQuAD):
Model | SQuAD P@5 | CuratedTREC P@5 | WebQuestions P@5 | WikiMovies P@5 | Size |
---|---|---|---|---|---|
TF-IDF model | 78.0 | 87.6 | 75.0 | 69.8 | ~13GB |
P@5 here is defined as the % of questions for which the answer segment appears in one of the top 5 documents.
Model trained only on SQuAD, evaluated in the SQuAD setting:
Model | SQuAD Dev EM | SQuAD Dev F1 | Size |
---|---|---|---|
Single model | 69.4 | 78.9 | ~130MB |
Model trained with distant supervision without NER/POS/lemma features, evaluated on multiple datasets (test sets, dev set for SQuAD) in the full Wikipedia setting:
Model | SQuAD EM | CuratedTREC EM | WebQuestions EM | WikiMovies EM | Size |
---|---|---|---|---|---|
Multitask model | 29.5 | 27.2 | 18.5 | 36.9 | ~270MB |
Our full-scale experiments were conducted on the 2016-12-21 dump of English Wikipedia. The dump was processed with the WikiExtractor and filtered for internal disambiguation, list, index, and outline pages (pages that are typically just links). We store the documents in an sqlite database for which drqa.retriever.DocDB
provides an interface.
Database | Num. Documents | Size |
---|---|---|
Wikipedia | 5,075,182 | ~13GB |
The datasets used for DrQA training and evaluation can be found here:
- SQuAD: train, dev
- WebQuestions: train, test, entities
- WikiMovies: train/test/entities (Rehosted in expected format from https://research.fb.com/downloads/babi/)
- CuratedTrec: train/test (Rehosted in expected format from https://github.com/brmson/dataset-factoid-curated)
The retriever/eval.py
, pipeline/eval.py
, and distant/generate.py
scripts expect the datasets as a .txt
file where each line is a JSON encoded QA pair, like so:
'{"question": "q1", "answer": ["a11", ..., "a1i"]}'
...
'{"question": "qN", "answer": ["aN1", ..., "aNi"]}'
Scripts to convert SQuAD and WebQuestions to this format are included in scripts/convert
. This is automatically done in download.sh
.
The reader
directory scripts expect the datasets as a .json
file where the data is arranged like SQuAD:
file.json
├── "data"
│ └── [i]
│ ├── "paragraphs"
│ │ └── [j]
│ │ ├── "context": "paragraph text"
│ │ └── "qas"
│ │ └── [k]
│ │ ├── "answers"
│ │ │ └── [l]
│ │ │ ├── "answer_start": N
│ │ │ └── "text": "answer"
│ │ ├── "id": "<uuid>"
│ │ └── "question": "paragraph question?"
│ └── "title": "document id"
└── "version": 1.1
Some datasets have (potentially large) candidate lists for selecting answers. For example, WikiMovies' answers are OMDb entries while WebQuestions is based on Freebase. If we have known candidates, we can impose that all predicted answers must be in this list by discarding any higher scoring spans that are not.
DrQA is not tied to any specific type of retrieval system -- as long as it effectively narrows the search space and focuses on relevant documents.
Following classical QA systems, we include an efficient (non-machine learning) document retrieval system based on sparse, TF-IDF weighted bag-of-word vectors. We use bags of hashed n-grams (here, unigrams and bigrams).
To see how to build your own such model on new documents, see the retriever README.
To interactively query Wikipedia:
python scripts/retriever/interactive.py --model /path/to/model
If model
is left out our default model will be used (assuming it was downloaded).
To evaluate the retriever accuracy (% match in top 5) on a dataset:
python scripts/retriever/eval.py /path/to/format/A/dataset.txt --model /path/to/model
DrQA's Document Reader is a multi-layer recurrent neural network machine comprehension model trained to do extractive question answering. That is, the model tries to find the answer to any question as a text span in one of the returned documents.
The Document Reader was inspired by, and primarily trained on, the SQuAD dataset. It can also be used standalone on such SQuAD-like tasks where a specific context is supplied with the question, the answer to which is contained in the context.
To see how to train the Document Reader on SQuAD, see the reader README.
To interactively ask questions about text with a trained model:
python scripts/reader/interactive.py --model /path/to/model
Again, here model
is optional; a default model will be used if it is left out.
To run model predictions on a dataset:
python scripts/reader/predict.py /path/to/format/B/dataset.json --model /path/to/model
The full system is linked together in drqa.pipeline.DrQA
.
To interactively ask questions using the full DrQA:
python scripts/pipeline/interactive.py
Optional arguments:
--reader-model Path to trained Document Reader model.
--retriever-model Path to Document Retriever model (tfidf).
--doc-db Path to Document DB.
--tokenizer String option specifying tokenizer type to use (e.g. 'corenlp').
--candidate-file List of candidates to restrict predictions to, one candidate per line.
--no-cuda Use CPU only.
--gpu Specify GPU device id to use.
To run predictions on a dataset:
python scripts/pipeline/predict.py /path/to/format/A/dataset.txt
Optional arguments:
--out-dir Directory to write prediction file to (<dataset>-<model>-pipeline.preds).
--reader-model Path to trained Document Reader model.
--retriever-model Path to Document Retriever model (tfidf).
--doc-db Path to Document DB.
--embedding-file Expand dictionary to use all pretrained embeddings in this file (e.g. all glove vectors to minimize UNKs at test time).
--candidate-file List of candidates to restrict predictions to, one candidate per line.
--n-docs Number of docs to retrieve per query.
--top-n Number of predictions to make per query.
--tokenizer String option specifying tokenizer type to use (e.g. 'corenlp').
--no-cuda Use CPU only.
--gpu Specify GPU device id to use.
--parallel Use data parallel (split across GPU devices).
--num-workers Number of CPU processes (for tokenizing, etc).
--batch-size Document paragraph batching size (Reduce in case of GPU OOM).
--predict-batch-size Question batching size (Reduce in case of CPU OOM).
DrQA's performance improves significantly in the full-setting when provided with distantly supervised data from additional datasets. Given question-answer pairs but no supporting context, we can use string matching heuristics to automatically associate paragraphs to these training examples.
Question: What U.S. state’s motto is “Live free or Die”?
Answer: New Hampshire
DS Document: Live Free or Die “Live Free or Die” is the official motto of the U.S. state of New Hampshire, adopted by the state in 1945. It is possibly the best-known of all state mottos, partly because it conveys an assertive independence historically found in American political philosophy and partly because of its contrast to the milder sentiments found in other state mottos.
The scripts/distant
directory contains code to generate and inspect such distantly supervised data. More information can be found in the distant supervision README.
We provide a number of different tokenizer options for convenience. Each has its own pros/cons based on how many dependencies it requires, overhead for running it, speed, and performance. For our reported experiments we used CoreNLP (but results are all similar).
Available tokenizers:
- CoreNLPTokenizer: Uses Stanford CoreNLP (option: 'corenlp'). We used v3.7.0. Requires Java 8.
- SpacyTokenizer: Uses spaCy (option: 'spacy').
- RegexpTokenizer: Custom regex-based PTB-style tokenizer (option: 'regexp').
- SimpleTokenizer: Basic alpha-numeric/non-whitespace tokenizer (option: 'simple').
See the list of mappings between string option names and tokenizer classes.
Please cite the ACL paper if you use DrQA in your work:
@inproceedings{chen2017reading,
title={Reading {Wikipedia} to Answer Open-Domain Questions},
author={Chen, Danqi and Fisch, Adam and Weston, Jason and Bordes, Antoine},
booktitle={Association for Computational Linguistics (ACL)},
year={2017}
}
This implementation of the DrQA Document Reader is closely related to the one found in ParlAI. Here, however, the work is extended to interact with the Document Retriever in the open-domain setting. On the other hand, the implementation in ParlAI is more general, and follows the appropriate API to work in more QA/Dialog settings.
Hamed Zaghaghi has provided a wrapper for a Web UI.
DrQA is BSD-licensed.