This project provides the implementation of distantly supervised relation extraction with (bag-level) soft-label adjustment.
Details of soft-label adjustment can be found here. The implementation is based on Tensorflow 1.0.0 and Python 2.7.
Distant supervision automatically generates training examples by aligning entity mentions in plain text with those in KB and labeling entity pairs with their relations in KB. If there's no relation link between certain entity pair in KB, it will be labeled as negative instance (NA).
The automatic labeling by KB inevitably accompanies with wrong labels because the relations of entity pairs might be missing from KBs or mislabeled. Multi-instances learning (MIL) is proposed to combat the noise. The method divides the training set into multiple bags of entity pairs (shown in the figure above) and labels the bags with the relations of entity pairs in the KB (bag-level DS label). Each bag consists of sentences mentioning both head and tail entities.
Much effort has been made in reducing the influence of noisy sentences within the bag, including methods based on at-least-one assumption and attention mechanisms over instances.
As shown in the figure above, due to the absence of (Jan Eliasson, Sweden)(Jan Eliasson is a Swedish diplomat.) from the Nationality relation in the KB,the entity pair is mislabeled as NA.
Actually, no matter how we design the weight calculation of the sentences (in that bag) for bag representation, the bag would be a noisy instance during training. So we try to solve the problem from a different point of view. Since the bag-level DS label can be mislabeled, we design a soft-label adjustment on the bag-level DS label to correct the ill-labeled cases.
Although our model can be trained on a CPU, using GPU can greatly accelerate the speed. So we strongly recommended using GPUs to train the model.
Our code is based on Tensorflow 1.0.0. You can find the installation instructions here.
requirements.txt summarize the dependencies of our code. You can install these dependencies by:
pip install -r requirements.txt
The benchmark dataset we use is FB+NYT (New York Times corpus as text resources; Freebase as distant supervision) which is also a widely benchmark dataset for many previous work.
The dataset we use can be downloaded via Google Drive or Baidu Yunpan.
Please unzip the data.zip in the soft-label-RE directory.
To avoid negative effects of dominant NA instances in the begining of training, soft-label method is adopted after about 3000 steps of parameter updates. During pre-training, the use_pre_train_model and use_soft_label should be set as False.
tf.app.flags.DEFINE_boolean("use_pre_train_model", False,'use pre-trained model or label')
tf.app.flags.DEFINE_boolean("use_soft_label", False,'use soft label or not')
tf.app.flags.DEFINE_string("load_model_name", 'pretrain/model.ckpt-3300','the path of pre-trained model without soft-label')
We provide two pre-trained models for ONE and ATT configurations, respectively. The pre-trained models can be downloaded via Google Drive or Baidu Yunpan.
Please unzip the pretrain.zip in the soft-label-RE directory.
Then, you can load the ONE model by setting:
tf.app.flags.DEFINE_boolean("use_pre_train_model", True,'use pre-trained model or label')
tf.app.flags.DEFINE_string("load_model_name", 'pretrain/one/model.ckpt-3300','the path of pre-trained model without soft-label')
or load the ATT model by setting:
tf.app.flags.DEFINE_boolean("use_pre_train_model", True,'use pre-trained model or label')
tf.app.flags.DEFINE_string("load_model_name", 'pretrain/att/model.ckpt-3600','the path of pre-trained model without soft-label')
You can choose ATT (selective attention) or ONE (at-least-one) configuration as the bag representation by setting the one_or_att to att or one, respectively.
tf.app.flags.DEFINE_string("one_or_att",'one','at-least-one or selective attention model')
Then run the main.py.
python main.py
The results can be found in res/TIME_STAMP/log.txt.
If you find the code and data resources helpful, please cite the following paper:
@inproceedings{liu2017soft,
title={A Soft-label Method for Noise-tolerant Distantly Supervised Relation Extraction},
author={Liu, Tianyu and Wang, Kexiang and Chang, Baobao and Sui, Zhifang},
booktitle={Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing},
pages={1790--1795},
year={2017}
}