README.md

metadata

dataset_info:
  features:
    - name: factuality_value
      dtype: string
    - name: predicat@xml:space
      dtype: string
    - name: predicat@charOffset
      dtype: string
    - name: predicat@headOffset
      dtype: string
    - name: predicat@id
      dtype: string
    - name: predicat@text
      dtype: string
    - name: predicat@type
      dtype: string
    - name: predicat@charOffsetMin
      dtype: int64
    - name: predicat@charOffsetMax
      dtype: int64
    - name: subject@xml:space
      dtype: string
    - name: subject@charOffset
      dtype: string
    - name: subject@headOffset
      dtype: string
    - name: subject@id
      dtype: string
    - name: subject@text
      dtype: string
    - name: subject@type
      dtype: string
    - name: subject@charOffsetMin
      dtype: int64
    - name: subject@charOffsetMax
      dtype: int64
    - name: object@xml:space
      dtype: string
    - name: object@charOffset
      dtype: string
    - name: object@headOffset
      dtype: string
    - name: object@id
      dtype: string
    - name: object@text
      dtype: string
    - name: object@type
      dtype: string
    - name: object@charOffsetMin
      dtype: int64
    - name: object@charOffsetMax
      dtype: int64
    - name: id
      dtype: string
    - name: raw_sent_text
      dtype: string
    - name: sent_charOffset
      dtype: string
    - name: sent_charOffsetMin
      dtype: int64
    - name: sent_charOffsetMax
      dtype: int64
    - name: formated_sentence
      dtype: string
  splits:
    - name: train
      num_bytes: 2278527
      num_examples: 3149
    - name: test
      num_bytes: 1559577
      num_examples: 2179
  download_size: 1308178
  dataset_size: 3838104
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
task_categories:
  - text-classification

Introduction

Factuality classification/quantification is one of the most difficult tasks in NLP. As apposed to sentiment analysis or other NLP tasks with statistical patterns, this task requires syntactic dependency patterns (aka, paradigmatics). In fact, N. Jiang et al have demonstrated BERTs inability to recognize paradigmatics.

Dataset Description

This dataset was constructed by H. Kilicoglu et al to predict the factuality expressed in text about a certain event/triple. Each triple is composed out of a subject-predicate-object. The dataset contains the position of each triple in a sentence, the raw sentence and a masked sentence where those positions are marked with special characters. It also contains the factuality value assigned by the H. Kilicoglu et al.

The sentences are taken from the PubMed biomedical abstracts.

The dataset factuality classes belong to a factuality scale introduced by H. Kilicoglu et al. The following figure shows this factuality scale. Counterfact, Doubtful, Possible, Probable, Certain represent varyibg levels of certainty, while Uncommited and Conditional represent a lack of information that would express factuality regarding a claim or an event.

Tasks

The main task that this data was designed for is factuality classification.