Statement Verification and Evidence Finding with Tables

This repository contains the solution code for the shared task which is part of the SemEval 2021 shared tasks.

Tasks

• Task A: Statement Verification
• Task B: Evidence Finding

Task A evaluation:

• 2 way: Statements are considered only if its ground truth verification is either entailed or refuted.
• 3 way: Statements with unknown ground truth verification also considered.

How to run?

Below are the example commands.

• Run over the dataset

       python -u -m src.dataset --data_dir <data_dir> --data_split <data_split> --submit_dir <submit_dir> --flag_cell_span --flag_cell_span --flag_approx_string_match

• Evaluate metrics

       python -u -m src.evaluate --input_dir <input_dir> --output_dir ./output/score/ --statistics_dir ./output/statistics --data_split <data_split>

  Directory structure expected by evaluate script

  input_dir
  └── ref (ground truthed xml files)
  └── res (predicted xml files)
  

Results

F1-score averaged over the tables.

Metrics/Split	Train	Dev	Test
# tables	981	52	52
task_a_2way_f1	0.3493	0.3564	0.3770
task_a_3way_f1	0.3493	0.4174	0.4695
task_b_f1	NA	0.3538	0.3641

Publication

KaushikAcharya at SemEval-2021 Task 9: Candidate Generation for Fact Verification over Tables (SemEval)

Citation

If you find this implementation helpful, please consider citing:

@inproceedings{acharya-2021-kaushikacharya,
    title = "{K}aushik{A}charya at {S}em{E}val-2021 Task 9: Candidate Generation for Fact Verification over Tables",
    author = "Acharya, Kaushik",
    booktitle = "Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021)",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.semeval-1.181",
    doi = "10.18653/v1/2021.semeval-1.181",
    pages = "1271--1275",
    abstract = "This paper describes the system submitted in the SemEval-2021 Statement Verification and Evidence Finding with Tables task. The system relies on candidate generation for logical forms on the table based on keyword matching and dependency parsing on the claim statements.",
}