BanglaFig: A Hierarchical Dataset of Bengali Literal and Figurative Language

Description

This repository contains the raw BanglaFig dataset, developed to support automatic detection of figurative language in Bengali, a low-resource, code-mixed setting. The raw dataset comprises seven category files: 6,000 Literal sentences and 1,000 sentences for each of six figurative subtypes (Idiom, Metaphor, Hyperbole, Simile, Personification, Sarcasm/Irony), totaling 12,000 sentences. Sentences were manually collected sentence-by-sentence from a deliberately varied set of sources to capture native Bengali script, fully Romanized Bengali ("Banglish"), and hybrid code-mixed text as it naturally occurs. Literal sentences were drawn from news outlets (BSS, The Daily Ittefaq, Prothom Alo), reference writing (Banglapedia, Bangla Wikipedia), and NCTB textbooks. Figurative sentences were drawn from Bangla Academy references, the writings of Tagore, Nazrul, and Jibanananda Das, Humayun Ahmed's novels, folk sources and proverbs, and social media (Facebook posts and blogs). Each sentence was independently annotated by three native Bengali speakers, first for the literal/figurative distinction (Stage 1) and then, for figurative sentences, for one of the six figurative subtypes (Stage 2). Disagreements were resolved by majority vote or, where all three annotators differed, by adjudication from a senior annotator. Inter-annotator agreement was strong at both stages (Stage 1 Cohen's kappa = 0.89; Stage 2 Fleiss' kappa = 0.78). Each raw file contains a sentence column (the Bengali sentence text) along with label and figurative_type columns identifying its category. Downstream preprocessing — removing rows with missing sentence or label values and deduplicating sentences that had been sourced into more than one category — reduces this raw set of 12,000 sentences to a final cleaned dataset of 11,947 sentences (6,000 Literal, 5,947 Figurative, distributed as 1,000 Personification, 1,000 Sarcasm, 1,000 Metaphor, 999 Idiom, 980 Simile, and 968 Hyperbole). Figurative sentences were deliberately spared aggressive text cleaning at the raw-collection stage, since idiomatic phrasing and unusual word order are central to the phenomena the dataset is intended to capture. BanglaFig is the first Bengali resource to unify literal-versus-figurative detection with fine-grained figurative subtype classification within a single hierarchically structured dataset, rather than covering a single figurative phenomenon as prior Bengali datasets have done. It is intended to support benchmarking of classical machine learning, deep learning, and transformer-based models for figurative language detection, as well as downstream research in machine translation, sentiment analysis, dialogue systems, and computational sociolinguistics.

Steps to reproduce

1. Literal sentences were collected from Bangladeshi news outlets (BSS, The Daily Ittefaq, Prothom Alo), reference sources (Banglapedia, Bangla Wikipedia), and NCTB textbooks. 2. Figurative sentences were collected category-by-category from literary sources (Bangla Academy references, and the works of Tagore, Nazrul, and Jibanananda Das for Metaphor, Simile, and Personification), Humayun Ahmed's novels, folk sources and proverbs, and social media (Facebook posts and blogs, primarily for Hyperbole and Sarcasm). 3. Collection was performed manually and sentence-by-sentence, discarding unclear or ambiguous sentences on the spot, and shifting sourcing away from any single author or outlet once a category neared its target size. 4. All sentences were normalized to Unicode NFC form with whitespace collapsed. Duplicate, empty, and artifact-only rows were removed, and sentences under four words were filtered out as unreliable fragments. 5. Three native Bengali speakers, not involved in data collection, independently annotated every sentence: first for the literal/figurative distinction (Stage 1), then, for figurative sentences, for one of six subtypes (Stage 2) — Idiom, Metaphor, Hyperbole, Simile, Personification, or Sarcasm/Irony. 6. Annotation followed a guideline refined after a pilot round. Full three-way agreement was kept as-is; two-of-three agreement was resolved by majority vote; remaining disagreements were adjudicated by a senior annotator. 7. The final annotated sentences were compiled into a single CSV file containing the sentence text, literal/figurative label, and figurative subtype (where applicable).

Institutions

• Hamdard University Bangladesh

  Dhaka Division, Dhaka
• Chittagong University of Engineering & Technology

  Chittagong, Chittagong

Categories

Licence