This repository puts together recent models and data sets for sentence-level relation extraction using knowledge bases (i.e., distant supervision). In particular, it contains the source code for WWW'17 paper CoType: Joint Extraction of Typed Entities and Relations with Knowledge Bases.
Please also check out our new repository on handling shifted label distribution in distant supervision
Task: Given a text corpus with entity mentions detected and heuristically labeled using distant supervision, the task aims to identify relation types/labels between a pair of entity mentions based on the sentence context where they co-occur.
For evaluating on sentence-level extraction, we processed (using our data pipeline) three public datasets to our JSON format. We ran Stanford NER on training set to detect entity mentions, mapped entity names to Freebase entities using DBpediaSpotlight, aligned Freebase facts to sentences, and assign entity types of Freebase entities to their mapped names in sentences:
PubMed-BioInfer: 100k PubMed paper abstracts as training data and 1,530 manually labeled biomedical paper abstracts from BioInfer (Pyysalo et al., 2007) as test data. It consists of 94 relation types (protein-protein interactions) and over 2,000 entity types (from MESH ontology). (Download)
NYT-manual: 1.18M sentences sampled from 294K New York Times news articles which were then aligned with Freebase facts by (Riedel et al., ECML'10) (link to Riedel's data). For test set, 395 sentences are manually annotated with 24 relation types and 47 entity types (Hoffmann et al., ACL'11) (link to Hoffmann's data). (Download)
Wiki-KBP: the training corpus contains 1.5M sentences sampled from 780k Wikipedia articles (Ling & Weld, 2012) plus ~7,000 sentences from 2013 KBP corpus. Test data consists of 14k system-labeled sentences from 2013 KBP slot filling assessment results. It has 7 relation types and 126 entity types after filtering of numeric value relations. (Download)
Please put the data files in corresponding subdirectories under
Performance comparison with several relation extraction systems over KBP 2013 dataset (sentence-level extraction).
|Mintz (our implementation, Mintz et al., 2009)||0.296||0.387||0.335|
|LINE + Dist Sup (Tang et al., 2015)||0.360||0.257||0.299|
|MultiR (Hoffmann et al., 2011)||0.325||0.278||0.301|
|FCM + Dist Sup (Gormley et al., 2015)||0.151||0.498||0.300|
|HypeNet (our implementation, Shwartz et al., 2016)||0.210||0.315||0.252|
|CNN (our implementation, Zeng et at., 2014)||0.198||0.334||0.242|
|PCNN (our implementation, Zeng et at., 2015)||0.220||0.452||0.295|
|LSTM (our implementation)||0.274||0.500||0.350|
|Bi-GRU (our implementation)||0.301||0.465||0.362|
|SDP-LSTM (our implementation, Xu et at., 2015)||0.300||0.436||0.356|
|Position-Aware LSTM (Zhang et al., 2017)||0.265||0.598||0.367|
|CoType-RM (Ren et al., 2017)||0.303||0.407||0.347|
|CoType (Ren et al., 2017)||0.348||0.406||0.369|
Note: for models that trained on sentences annotated with a single label (HypeNet, CNN/PCNN, LSTM, SDP/PA-LSTMs, Bi-GRU), we form one training instance for each sentence-label pair based on their DS-annotated data.
We will take Ubuntu for example.
$ pip install pexpect ujson tqdm
$ cd code/DataProcessor/ $ git clone [email protected]:stanfordnlp/stanza.git $ cd stanza $ pip install -e . $ wget http://nlp.stanford.edu/software/stanford-corenlp-full-2016-10-31.zip $ unzip stanford-corenlp-full-2016-10-31.zip
We have included compilied binaries. If you need to re-compile
retype.cpp under your own g++ environment
$ cd code/Model/retype; make
As an example, we show how to run CoType on the Wiki-KBP dataset
Start the Stanford corenlp server for the python wrapper.
$ java -mx4g -cp "code/DataProcessor/stanford-corenlp-full-2016-10-31/*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer
Feature extraction, embedding learning on training data, and evaluation on test data.
For relation classification, the "none"-labeled instances need to be first removed from train/test JSON files. The hyperparamters for embedding learning are included in the run.sh script.
Dataset to run on.
- KBP: -negative 3 -iters 400 -lr 0.02 -transWeight 1.0 - NYT: -negative 5 -iters 700 -lr 0.02 -transWeight 7.0 - BioInfer: -negative 5 -iters 700 -lr 0.02 -transWeight 7.0
Hyperparameters for relation classification are included in the run.sh script.
Evaluates relation extraction performance (precision, recall, F1): produce predictions along with their confidence score; filter the predicted instances by tuning the thresholds.
$ python code/Evaluation/emb_test.py extract KBP retype cosine 0.0 $ python code/Evaluation/tune_threshold.py extract KBP emb retype cosine
The last command in run.sh generates json file for predicted results, in the same format as test.json in data/source/$DATANAME, except that we only output the predicted relation mention labels. Replace the second parameter with whatever threshold you would like.
$ python code/Evaluation/convertPredictionToJson.py $Data 0.0
Code for producing the JSON files from a raw corpus for running CoType and baseline models is here.
You can find our implementation of some recent relation extraction models under the Code/Model/ directory.