TFWIN
Source code and dataset of TheWebConf 2019 paper: "A Novel Unsupervised Approach for Precise Temporal Slot Filling from Incomplete and Noisy Temporal Contexts" [pdf]
If you use the code/data please cite this paper:
@inproceedings{wang2019novel,
title={A Novel Unsupervised Approach for Precise Temporal Slot Filling from Incomplete and Noisy Temporal Contexts},
author={Wang, Xueying and Zhang, Haiqiao and Li, Qi and Shi, Yiyu and Jiang, Meng},
booktitle={Proceedings of TheWebConf 2019},
year={2019}
}
Overview
In this work, we proposed an unsupervised approach of two modules that mutually enhance each other: one is a reliability estimator on fact extractors conditionally to the temporal contexts; the other is a fact trustworthiness estimator based on the extractor’s reliability. The iterative learning process reduces the noise of the extractions.
Data
This folder "data" contains 4 structured datasets for termpoal facts extraction.
data_post_CP: country's president data with post time signal
data_post_SP: sport's player data with post time signal
data_text_CP: country's president data with text time signal
data_text_CP: sport's player data with text time signal
Format of these four datasets are same, for each line:
RANKPATTERN \t PATTERN \t ENTITYPOS \t VALUEPOS \t RANKENTITYVALUETM \t ENTITY \t VALUE \t TM \t COUNT
Besides, ground truth for "country's president" is collected in "groundtruth_president".
Codes
The source codes of our models are in "tfwin.py", run with python 2.7
Contact
Xueying Wang (xwang41@nd.edu)