Abstractive Open IE

This repository contains the code for the paper:
Abstractive OpenIE
Kevin Pei, Ishan Jindal, Kevin Chen-Chuan Chang
EMNLP 2023

Overview

Open Information Extraction (OpenIE) is the task of extracting relation tuples from plain text. In its simplest form, OpenIE extracts information in the form of tuples consisting of subject(S), predicate(P), object(O), and any additional arguments(A). Abstractive OpenIE aims to expand on this task by including the extraction of inferred relations, which are relations where the tokens of the predicate or arguments do not exist within the sentence.

Data

https://drive.google.com/file/d/1j1DyVu9-U2fn2E_MKGH8xql3OBLrMbk8/view?usp=sharing

Training data include: oie4_train.tsv Base OIE4 training data de_backtranslated_oie4.tsv OIE4 with German back translated sentences sure_augmented_oie4.tsv OIE4 augmented with SuRE extractions de_backtranslated_sure_augmented_oie4.tsv OIE4 with both German backtranslated sentences and augmented by SuRE extractions

Dev sets include: lsoie_wiki_dev.tsv

Test sets include: WiRe57_test.txt Re-OIE2016_test.txt CaRB_test.txt lsoie_wiki_test.tsv CQ-W_QA_test.txt The test set used to evaluate on the downstream QUEST task

###Usage

python backtranslation.py [--FLAGS]

Creates paraphrases of sentences using back translation on the given training file. The language used is German.

python sure_data_creator.py [--FLAGS]

Creates inputs suitable for use with SuRE. The SuRE model can be accessed in the following github repository: https://github.com/luka-group/SuRE

Models

https://drive.google.com/file/d/1AZUmniWQ3irEXZkDbO5kbMZJztFSjylQ/view?usp=sharing

Models include: abstractive_oie4 T5 fine-tuned on base OIE4 training data abstractive_de_backtranslated_oie4 T5 fine-tuned on OIE4 with German back translated sentences abstractive_sure_augmented_oie4 T5 fine-tuned on OIE4 augmented with SuRE extractions abstractive_de_backtranslated_sure_augmented_oie4 T5 fine-tuned on OIE4 with both German backtranslated sentences and augmented by SuRE extractions

Usage

python main.py [--FLAGS]

Can be run in train or test mode --train_file and --dev_file are only necessary for training, and --test_file and --prediction_file are only necessary for testing

Evaluation

We provide code for entailment-based metrics. They are sentence-tuple entailment, combined tuple-tuple entailment, and tuple-tuple entailment.

Usage

python entailment_eval.py [--FLAGS]