ChristianHugD/altlex

---

Dataset Description

Dataset Summary

This dataset is derived from the SemEval-2010 Task 8: "Identifying the Cause-Effect Relation in Text". It focuses on identifying and classifying causal relationships between entities in sentences. The original task aimed to promote research in relation extraction, specifically focusing on the detection and classification of semantic relations between pairs of nominals.

This particular version provides the data in a ready-to-use CSV format with three configurations tailored for common NLP tasks:

• sequence-classification: For classifying the presence of a causal relation at the sentence level.
• pair-classification: For classifying the causal relationship between specific entity pairs within text (potentially using text with marked pairs).
• token-classification: For detecting and labeling "Cause" and "Effect" entities as spans within text (e.g., using IOB format).

Supported Tasks and Leaderboards

This dataset can be used to train and evaluate models for:

• Text Classification: For determining if a sentence expresses a causal relationship (sequence-classification config).
• Relation Extraction / Text Classification: For classifying the type of relationship between two marked nominals (pair-classification config).
• Named Entity Recognition (NER) / Token Classification: For identifying and tagging cause and effect entities within sentences (token-classification config).

Given its origin, it's suitable for benchmarking performance on relation extraction tasks. You might find relevant leaderboards on the original SemEval-2010 Task 8 website or other platforms dedicated to relation extraction.

Languages

The dataset is entirely in English (en).

---

Dataset Structure

Data Instances

Each instance in the dataset typically represents a sentence or a segment of text with associated labels. For the token-classification setup, sentences are tokenized.

Here's an example for the token-classification config:

Configurations Overview

This dataset offers the following configurations, each tailored for a specific causal extraction task. You select the desired configuration when loading the dataset using load_dataset(). All configurations share the same underlying data files (train.csv, validation.csv, test.csv), but interpret specific columns for their respective tasks.

1. sequence-classification Config

This configuration provides text and a binary label indicating whether a causal relation is present in the text. It is designed for sequence classification tasks.

Key Columns

• text: string - The input text, representing the document or sentence to be classified. This serves as the input feature for models.
• seq_label: int - The binary label indicating the presence (1) or absence (0) of a causal relation. This is the target label for classification.
  • 0: negative_causal_relation (No causal relation detected)
  • 1: positive_causal_relation (A causal relation is present)

Data Instance Example

{
  "text": "We have gotten the agreement of the Chairman and the Secretary, preliminary to any opening statements, to stay until 1 p.m. We will probably have some votes, so we will maximize our time.",
  "seq_label": 1
}

2. pair-classification Config

This configuration focuses on classifying the causal relationship between two pre-defined text spans within a larger text. It is designed for pair-classification tasks where the input often highlights the potential cause and effect arguments.

Key Columns

• text_w_pairs: string - The text where the potential causal arguments are explicitly marked (e.g., using special tags like <ARG0> and <ARG1>). This is the input feature for models.
• pair_label: int - The binary label indicating whether the relationship between the marked pair is causal (1) or not (0). This is the target label for classification.
  • 0: negative_causal_relation (No causal relation between the pair)
  • 1: positive_causal_relation (A causal relation exists between the pair)

Data Instance Example

{
  "text_w_pairs": "We have gotten the agreement of the Chairman and the Secretary, preliminary to any opening statements, to stay until 1 p.m. <ARG0>We will probably have some votes</ARG0>, so <ARG1>we will maximize our time</ARG1>.",
  "pair_label": 1
}

3. token-classification Config

This configuration provides pre-tokenized text and corresponding token-level labels (BIO tags) that mark the spans of Causes and Effects. It is suitable for token classification (span detection) tasks.

Key Columns

• text: string - The original raw text (provided for context).
• tokens: list[str] - The pre-tokenized version of the text. This is the input feature for models.
• labels: list[int] - A list of integer IDs, where each ID corresponds to a BIO tag for the respective token in the tokens list. This is the target label for span detection.
  • 0: O (Outside of any annotated causal-span)
  • 1: B-Cause (Beginning of a Cause span)
  • 2: I-Cause (Inside a Cause span)
  • 3: B-Effect (Beginning of an Effect span)
  • 4: I-Effect (Inside an Effect span)

Data Instance Example

{
  "text": "The heavy rain caused flooding in the streets.",
  "tokens": ["The", "heavy", "rain", "caused", "flooding", "in", "the", "streets", "."],
  "labels": [0, 1, 2, 0, 3, 4, 4, 4, 0] # Example BIO tags for Cause "heavy rain" and Effect "flooding in the streets"
}