• Stars
    star
    420
  • Rank 103,194 (Top 3 %)
  • Language
    Python
  • License
    Apache License 2.0
  • Created over 7 years ago
  • Updated almost 6 years ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

Document QA

This repo contains code for our paper Simple and Effective Multi-Paragraph Reading Comprehension. It can be used to train neural question answering models in tensorflow, and in particular for the case when we want to run the model over multiple paragraphs for each question. Code is included to train on the TriviaQA and SQuAD datasets.

A demo of this work can be found at documentqa.allenai.org

Small forewarning, this is still much more of a research codebase then a library. we anticipate porting this work in allennlp where it will enjoy a cleaner implementation and more stable support.

Setup

Dependencies

We require python >= 3.5, tensorflow 1.3, and a handful of other supporting libraries. Tensorflow should be installed separately following the docs. To install the other dependencies use

pip install -r requirements.txt

The stopword corpus and punkt sentence tokenizer for nltk are needed and can be fetched with:

python -m nltk.downloader punkt stopwords

The easiest way to run this code is to use:

export PYTHONPATH=${PYTHONPATH}:`pwd`

Data

By default, we expect source data to be stored in "~/data" and preprocessed data to be stored in "./data". The expected file locations can be changed by altering config.py.

Word Vectors

The models we train use the common crawl 840 billion token GloVe word vectors from here. They are expected to exist in "~/data/glove/glove.840B.300d.txt" or "~/data/glove/glove.840B.300d.txt.gz".

For example:

mkdir -p ~/data
mkdir -p ~/data/glove
cd ~/data/glove
wget http://nlp.stanford.edu/data/glove.840B.300d.zip
unzip glove.840B.300d.zip
rm glove.840B.300d.zip

SQuAD Data

Training or testing on SQuAD requires downloading the SQuAD train/dev files into ~/data/squad. This can be done as follows:

mkdir -p ~/data/squad
cd ~/data/squad
wget https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json
wget https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json

then running:

python docqa/squad/build_squad_dataset.py

This builds pkl files of the tokenized data in "./data/squad"

TriviaQA Data

The raw TriviaQA data is expected to be unzipped in "~/data/triviaqa". Training or testing in the unfiltered setting requires the unfiltered data to be download to "~/data/triviaqa-unfiltered".

mkdir -p ~/data/triviaqa
cd ~/data/triviaqa
wget http://nlp.cs.washington.edu/triviaqa/data/triviaqa-rc.tar.gz
tar xf triviaqa-rc.tar.gz
rm triviaqa-rc.tar.gz

cd ~/data
wget http://nlp.cs.washington.edu/triviaqa/data/triviaqa-unfiltered.tar.gz
tar xf triviaqa-unfiltered.tar.gz
rm triviaqa-unfiltered.tar.gz

To use TriviaQA we need to tokenize the evidence documents, which can be done by

python docqa/triviaqa/evidence_corpus.py

This can be slow, we support multi-processing

python docqa/triviaqa/evidence_corpus.py --n_processes 8

This builds evidence files in "./data/triviaqa/evidence" that are split into paragraphs, sentences, and tokens. Then we need to tokenize the questions and locate the relevant answers spans in each document. Run

python docqa/triviaqa/build_span_corpus.py {web|wiki|open} --n_processes 8

to build the desired set. This builds pkl files "./data/triviaqa/{web|wiki|open}"

Training

Once the data is in place our models can be trained by

python docqa/scripts/ablate_{triviaqa|squad|triviaqa_wiki|triviaqa_unfiltered}.py

See the help menu for these scripts for more details. Note that since we use the Cudnn RNN implementations, these models can only be trained on a GPU. We do provide a script for converting the (trained) models to CPU versions:

python docqa/scripts/convert_to_cpu.py

Modifying the hyper-parameters beyond the ablations requires building your own train script.

Testing

SQuAD

Use "docqa/eval/squad_eval.py" to evaluate on paragraph-level (i.e., standard) SQuAD. For example:

python docqa/eval/squad_eval.py -o output.json -c dev /path/to/model/directory

"output.json" can be used with the official evaluation script, for example:

python docqa/squad/squad_official_evaluation.py ~/data/squad/dev-v1.1.json output.json

Use "docqa/eval/squad_full_document_eval.py" to evaluate on the document-level. For example

python docqa/eval/squad_full_document_eval.py -c dev /path/to/model/directory output.csv

This will store the per-paragraph results in output.csv, we can then run:

python docqa/eval/ranked_scores.py output.csv

to get ranked scores as more paragraphs are used.

TriviaQA

Use "docqa/eval/triviaqa_full_document_eval.py" to evaluate on TriviaQA datasets, like:

python docqa/eval/triviaqa_full_document_eval.py --n_processes 8 -c web-dev --tokens 800 -o question-output.json -p paragraph-output.csv /path/to/model/directory

Then the "question-output.json" can be used with the standard triviaqa evaluation script, the "paragraph-output.csv" contains per-paragraph output, we can run

python docqa/eval/ranked_scores.py paragraph-output.csv

to get ranked scores as more paragraphs as used for each question, or

python docqa/eval/ranked_scores.py --per_doc paragraph-output.csv

to get ranked scores as more paragraphs as used for each (question, document) pair, as should be done for TrivaQA web.

User Input

"docqa/scripts/run_on_user_documents.py" serves as a heavily commented example of how to run our models and pre-processing pipeline on other kinds of text. For example:

python docqa/scripts/run_on_user_documents.py /path/to/model/directory "Who wrote the satirical essay 'A Modest Proposal'?" ~/data/triviaqa/evidence/wikipedia/A_Modest_Proposal.txt ~/data/triviaqa/evidence/wikipedia/Jonathan_Swift.txt

Pre-Trained Models

We have four pre-trained models

  1. "squad" Our model trained on the standard SQuAD dataset, this model is listed on the SQuAD leaderboard as BiDAF + Self Attention

  2. "squad-shared-norm" Our model trained on document-level SQuAD using the shared-norm approach.

  3. "triviaqa-web-shared-norm" Our model trained on TriviaQA web with the shared-norm approach. This is the model we used to submit scores to the TriviaQA leader board.

  4. "triviaqa-unfiltered-shared-norm" Our model trained on TriviaQA unfiltered with the shared-norm approach. This is the model that powers our demo.

The models can be downloaded here

The models use the cuDNN implementation of GRUs by default, which means they can only be run on the GPU. We also have slower, but CPU compatible, versions here.

More Repositories

1

allennlp

An open-source NLP research library, built on PyTorch.
Python
11,751
star
2

OLMo

Modeling, training, eval, and inference code for OLMo
Python
4,535
star
3

RL4LMs

A modular RL library to fine-tune language models to human preferences
Python
2,101
star
4

longformer

Longformer: The Long-Document Transformer
Python
2,022
star
5

bilm-tf

Tensorflow implementation of contextualized word representations from bi-directional language models
Python
1,621
star
6

scispacy

A full spaCy pipeline and models for scientific/biomedical documents.
Python
1,618
star
7

bi-att-flow

Bi-directional Attention Flow (BiDAF) network is a multi-stage hierarchical process that represents context at different levels of granularity and uses a bi-directional attention flow mechanism to achieve a query-aware context representation without early summarization.
Python
1,533
star
8

scibert

A BERT model for scientific text.
Python
1,495
star
9

open-instruct

Python
1,185
star
10

ai2thor

An open-source platform for Visual AI.
C#
1,160
star
11

dolma

Data and tools for generating and inspecting OLMo pre-training data.
Python
961
star
12

XNOR-Net

ImageNet classification using binary Convolutional Neural Networks
Lua
839
star
13

s2orc

S2ORC: The Semantic Scholar Open Research Corpus: https://www.aclweb.org/anthology/2020.acl-main.447/
Python
817
star
14

mmc4

MultimodalC4 is a multimodal extension of c4 that interleaves millions of images with text.
Python
793
star
15

scitldr

Python
734
star
16

objaverse-xl

🪐 Objaverse-XL is a Universe of 10M+ 3D Objects. Contains API Scripts for Downloading and Processing!
Python
701
star
17

papermage

library supporting NLP and CV research on scientific papers
Python
692
star
18

natural-instructions

Expanding natural instructions
Python
690
star
19

visprog

Official code for VisProg (CVPR 2023 Best Paper!)
Python
686
star
20

science-parse

Science Parse parses scientific papers (in PDF form) and returns them in structured form.
Java
611
star
21

pdffigures2

Given a scholarly PDF, extract figures, tables, captions, and section titles.
Scala
593
star
22

writing-code-for-nlp-research-emnlp2018

A companion repository for the "Writing code for NLP Research" Tutorial at EMNLP 2018
Python
558
star
23

tango

Organize your experiments into discrete steps that can be cached and reused throughout the lifetime of your research project.
Python
528
star
24

allennlp-models

Officially supported AllenNLP models
Python
521
star
25

specter

SPECTER: Document-level Representation Learning using Citation-informed Transformers
Python
506
star
26

dont-stop-pretraining

Code associated with the Don't Stop Pretraining ACL 2020 paper
Python
488
star
27

unified-io-2

Python
471
star
28

macaw

Multi-angle c(q)uestion answering
Python
451
star
29

lumos

Code and data for "Lumos: Learning Agents with Unified Data, Modular Design, and Open-Source LLMs"
Python
433
star
30

scholarphi

An interactive PDF reader.
Python
418
star
31

deep_qa

A deep NLP library, based on Keras / tf, focused on question answering (but useful for other NLP too)
Python
404
star
32

acl2018-semantic-parsing-tutorial

Materials from the ACL 2018 tutorial on neural semantic parsing
402
star
33

unifiedqa

UnifiedQA: Crossing Format Boundaries With a Single QA System
Python
384
star
34

pawls

Software that makes labeling PDFs easy.
Python
380
star
35

OLMoE

OLMoE: Open Mixture-of-Experts Language Models
Jupyter Notebook
374
star
36

kb

KnowBert -- Knowledge Enhanced Contextual Word Representations
Python
359
star
37

PeerRead

Data and code for Kang et al., NAACL 2018's paper titled "A Dataset of Peer Reviews (PeerRead): Collection, Insights and NLP Applications"
Python
354
star
38

reward-bench

RewardBench: the first evaluation tool for reward models.
Python
346
star
39

naacl2021-longdoc-tutorial

Python
342
star
40

openie-standalone

Quality information extraction at web scale. Edit
Scala
327
star
41

Holodeck

CVPR 2024: Language Guided Generation of 3D Embodied AI Environments.
Python
319
star
42

python-package-template

A template repo for Python packages
Python
318
star
43

allenact

An open source framework for research in Embodied-AI from AI2.
Python
316
star
44

ir_datasets

Provides a common interface to many IR ranking datasets.
Python
314
star
45

s2orc-doc2json

Parsers for scientific papers (PDF2JSON, TEX2JSON, JATS2JSON)
Python
302
star
46

acl2022-zerofewshot-tutorial

291
star
47

OLMo-Eval

Evaluation suite for LLMs
Python
280
star
48

procthor

🏘️ Scaling Embodied AI by Procedurally Generating Interactive 3D Houses
Python
257
star
49

fm-cheatsheet

Website for hosting the Open Foundation Models Cheat Sheet.
JavaScript
255
star
50

FineGrainedRLHF

Python
243
star
51

beaker-cli

A collaborative platform for rapid and reproducible research.
Go
230
star
52

comet-atomic-2020

Python
228
star
53

spv2

Science-parse version 2
Python
225
star
54

scifact

Data and models for the SciFact verification task.
Python
217
star
55

objaverse-rendering

📷 Scripts for rendering Objaverse
Python
206
star
56

ScienceWorld

ScienceWorld is a text-based virtual environment centered around accomplishing tasks from the standardized elementary science curriculum.
Scala
197
star
57

unified-io-inference

Jupyter Notebook
196
star
58

allennlp-demo

Code for the AllenNLP demo.
TypeScript
191
star
59

citeomatic

A citation recommendation system that allows users to find relevant citations for their paper drafts. The tool is backed by Semantic Scholar's OpenCorpus dataset.
Jupyter Notebook
189
star
60

cartography

Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics
Jupyter Notebook
188
star
61

savn

Learning to Learn how to Learn: Self-Adaptive Visual Navigation using Meta-Learning (https://arxiv.org/abs/1812.00971)
Python
175
star
62

vampire

Variational Methods for Pretraining in Resource-limited Environments
Python
173
star
63

vila

Incorporating VIsual LAyout Structures for Scientific Text Classification
Python
172
star
64

s2-folks

Public space for the user community of Semantic Scholar APIs to share scripts, report issues, and make suggestions.
171
star
65

hidden-networks

Python
164
star
66

cord19

Get started with CORD-19
161
star
67

mmda

multimodal document analysis
Jupyter Notebook
158
star
68

PRIMER

The official code for PRIMERA: Pyramid-based Masked Sentence Pre-training for Multi-document Summarization
Python
150
star
69

catwalk

This project studies the performance and robustness of language models and task-adaptation methods.
Python
141
star
70

dnw

Discovering Neural Wirings (https://arxiv.org/abs/1906.00586)
Python
139
star
71

deepfigures-open

Companion code to the paper "Extracting Scientific Figures with Distantly Supervised Neural Networks" 🤖
Python
133
star
72

tpu_pretrain

LM Pretraining with PyTorch/TPU
Python
132
star
73

allentune

Hyperparameter Search for AllenNLP
Python
128
star
74

SciREX

Data/Code Repository for https://api.semanticscholar.org/CorpusID:218470122
Python
128
star
75

scidocs

Dataset accompanying the SPECTER model
Python
127
star
76

lm-explorer

interactive explorer for language models
Python
127
star
77

pdffigures

Command line tool to extract figures, tables, and captions from scholarly documents in PDF form.
C++
125
star
78

OpenBookQA

Code for experiments on OpenBookQA from the EMNLP 2018 paper "Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering"
Python
121
star
79

peS2o

Pretraining Efficiently on S2ORC!
120
star
80

gooaq

Question-answers, collected from Google
Python
116
star
81

allennlp-as-a-library-example

A simple example for how to build your own model using AllenNLP as a dependency.
Python
113
star
82

embodied-clip

Official codebase for EmbCLIP
Python
111
star
83

multimodalqa

Python
109
star
84

alexafsm

With alexafsm, developers can model dialog agents with first-class concepts such as states, attributes, transition, and actions. alexafsm also provides visualization and other tools to help understand, test, debug, and maintain complex FSM conversations.
Python
108
star
85

allennlp-semparse

A framework for building semantic parsers (including neural module networks) with AllenNLP, built by the authors of AllenNLP
Python
107
star
86

scicite

Repository for NAACL 2019 paper on Citation Intent prediction
Python
106
star
87

ai2thor-rearrangement

🔀 Visual Room Rearrangement
Python
104
star
88

commonsense-kg-completion

Python
102
star
89

medicat

Dataset of medical images, captions, subfigure-subcaption annotations, and inline textual references
Python
102
star
90

real-toxicity-prompts

Jupyter Notebook
101
star
91

s2search

The Semantic Scholar Search Reranker
Python
99
star
92

aristo-mini

Aristo mini is a light-weight question answering system that can quickly evaluate Aristo science questions with an evaluation web server and the provided baseline solvers.
Python
96
star
93

gpv-1

A task-agnostic vision-language architecture as a step towards General Purpose Vision
Jupyter Notebook
92
star
94

flex

Few-shot NLP benchmark for unified, rigorous eval
Python
91
star
95

elastic

Python
91
star
96

manipulathor

ManipulaTHOR, a framework that facilitates visual manipulation of objects using a robotic arm
Jupyter Notebook
88
star
97

spoc-robot-training

SPOC: Imitating Shortest Paths in Simulation Enables Effective Navigation and Manipulation in the Real World
Python
85
star
98

S2AND

Semantic Scholar's Author Disambiguation Algorithm & Evaluation Suite
Python
85
star
99

propara

ProPara (Process Paragraph Comprehension) dataset and models
Python
82
star
100

ARC-Solvers

ARC Question Solvers
Python
82
star