FICLE Dataset

The dataset can be loaded and utilized through the following:

from datasets import load_dataset
ficle_data = load_dataset("tathagataraha/ficle")

Dataset card for FICLE

Dataset Summary

The FICLE dataset is a derivative of the FEVER dataset, which is a collection of 185,445 claims generated by modifying sentences obtained from Wikipedia. These claims were then verified without knowledge of the original sentences they were derived from. Each sample in the FEVER dataset consists of a claim sentence, a context sentence extracted from a Wikipedia URL as evidence, and a type label indicating whether the claim is supported, refuted, or lacks sufficient information.

Languages

The FICLE Dataset contains only English.

Dataset Structure

Data Fields

• Claim (string): A statement or proposition relating to the consistency or inconsistency of certain facts or information.
• Context (string): The surrounding information or background against which the claim is being evaluated or compared. It provides additional details or evidence that can support or challenge the claim.
• Source (string): It is the linguistic chunk containing the entity lying to the left of the main verb/relating chunk.
• Source Indices (string): Source indices refer to the specific indices or positions within the source string that indicate the location of the relevant information.
• Relation (string): It is the linguistic chunk containing the verb/relation at the core of the identified inconsistency.
• Relation Indices (string): Relation indices indicate the specific indices or positions within the relation string that highlight the location of the relevant information.
• Target (string): It is the linguistic chunk containing the entity lying to the right of the main verb/relating chunk.
• Target Indices (string): Target indices represent the specific indices or positions within the target string that indicate the location of the relevant information.
• Inconsistent Claim Component (string): The inconsistent claim component refers to a specific linguistic chunk within the claim that is identified as inconsistent with the context. It helps identify which part of the claim triple is problematic in terms of its alignment with the surrounding information.
• Inconsistent Context-Span (string): A span or portion marked within the context sentence that is found to be inconsistent with the claim. It highlights a discrepancy or contradiction between the information in the claim and the corresponding context.
• Inconsistent Context-Span Indices (string): The specific indices or location within the context sentence that indicate the inconsistent span.
• Inconsistency Type (string): The category or type of inconsistency identified in the claim and context.
• Fine-grained Inconsistent Entity-Type (string): The specific detailed category or type of entity causing the inconsistency within the claim or context. It provides a more granular classification of the entity associated with the inconsistency.
• Coarse Inconsistent Entity-Type (string): The broader or general category or type of entity causing the inconsistency within the claim or context. It provides a higher-level classification of the entity associated with the inconsistency.

Data Splits

The FICLE dataset comprises a total of 8,055 samples in the English language, each representing different instances of inconsistencies. These inconsistencies are categorized into five types: Taxonomic Relations (4,842 samples), Negation (1,630 samples), Set Based (642 samples), Gradable (526 samples), and Simple (415 samples).

Within the dataset, there are six possible components that contribute to the inconsistencies found in the claim sentences. These components are distributed as follows: Target-Head (3,960 samples), Target-Modifier (1,529 samples), Relation-Head (951 samples), Relation-Modifier (1,534 samples), Source-Head (45 samples), and Source-Modifier (36 samples).

The dataset is split into train, validation, and test.

• train: 6.44k rows
• validation: 806 rows
• test: 806 rows

Dataset Creation

Curation Rationale

We propose a linguistically enriched dataset to help detect inconsistencies and explain them. To this end, the broad requirements are to locate where the inconsistency is present between a claim and a context and to have a classification scheme for better explainability.

Data Collection and Preprocessing

The FICLE dataset is derived from the FEVER dataset, using the following- ing processing steps. FEVER (Fact Extraction and VERification) consists of 185,445 claims were generated by altering sentences extracted from Wikipedia and subsequently verified without knowledge of the sentence they were derived from. Every sample in the FEVER dataset contains the claim sentence, evidence (or context) sentence from a Wikipedia URL, and a type label (‘supports’, ‘refutes’, or ‘not enough info’). Out of these, we leverage only the samples with the ‘refutes’ label to build our dataset.

Annotations

You can see the annotation guidelines here.

In order to provide detailed explanations for inconsistencies, extensive annotations were conducted for each sample in the FICLE dataset. The annotation process involved two iterations, with each iteration focusing on different aspects of the dataset. In the first iteration, the annotations were primarily "syntactic-oriented." These fields included identifying the inconsistent claim fact triple, marking inconsistent context spans, and categorizing the six possible inconsistent claim components. The second iteration of annotations concentrated on "semantic-oriented" aspects. Annotators labeled semantic fields for each sample, such as the type of inconsistency, coarse inconsistent entity types, and fine-grained inconsistent entity types. This stage aimed to capture the semantic nuances and provide a deeper understanding of the inconsistencies present in the dataset.

The annotation process was carried out by a group of four annotators, two of whom are also authors of the dataset. The annotators possess a strong command of the English language and hold Bachelor's degrees in Computer Science, specializing in computational linguistics. Their expertise in the field ensured accurate and reliable annotations. The annotators' ages range from 20 to 22 years, indicating their familiarity with contemporary language usage and computational linguistic concepts.

Personal and Sensitive Information

Considerations for Using the Data

Social Impact of Dataset

Discussion of Biases

Other Known Limitations

Additional Information

Citation Information

@misc{raha2023neural,
      title={Neural models for Factual Inconsistency Classification with Explanations}, 
      author={Tathagata Raha and Mukund Choudhary and Abhinav Menon and Harshit Gupta and KV Aditya Srivatsa and Manish Gupta and Vasudeva Varma},
      year={2023},
      eprint={2306.08872},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Contact