ARTime

Automatic Rule Generation for Time Expression Normalization (Findings of EMNLP, 2021)

arXiv link: https://arxiv.org/abs/2108.13658

The codes are obtained in the software directory.

The evaluationResults.zip file includes the main results reported in our paper.

Citation

@inproceedings{ding-etal-2021-automatic-rule,
    title = "Automatic rule generation for time expression normalization",
    author = "Ding, Wentao  and
      Chen, Jianhao  and
      Li, Jinmao  and
      Qu, Yuzhong",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2021",
    month = nov,
    year = "2021",
    doi = "10.18653/v1/2021.findings-emnlp.269",
    pages = "3135--3144",
}

Building

ARTime is maven project built with IntelliJ IDEA. It runs with java-11 and scala-2.12.11.

Program Entry

First, mark software/src/main as Sources Root. The main function to reproduce the results of our method is artime.Main.main(). The train process and test process have been annotated. You need to choose the datasets from the commented codes for training and testing to reproduce all kinds of results. The code gives an example which treats Tweets-modified trainingset as trainingset and treats Tweets-modified testset as testset to get the results of ARTime normalization with gold mention on Tweets-modified.

Switch to ARTime+H

If you want to reproduce the results of ARTime+H, you need to make 2 small modifications to our codes:

In artime.struct.TexToken.scala, line 76:

val On=true

In artime.rule.RuleHandler.scala, line 199:

val ruleSorted: IndexedSeq[Rule] = getRules(100)

Evaluation tools

We use the standard evaluation tools of TempEval-3 Task to evaluate our performance， which is included in software/evaluation_tools.