BanglaCineAspect: A dataset for aspect-based sentiment analysis of Bangla movie reviews

Published: 26 October 2026| Version 1 | DOI: 10.17632/zv6mms5vfh.1
Contributors:
, Anindya Majumder , Tafhimul Azwad, Mahfuzulhoq Chowdhury

Description

BanglaCineAspect is a dataset of 1,421 Bangla movie reviews annotated for aspect-based sentiment analysis (ABSA). Each review is labelled for nine aspects of a film: Direction, Acting, Songs & Music, Screenplay, Story, Cinematography, Background Music, Editing and Dialogue. An aspect is coded 1 if the review expresses a positive opinion about it, 0 if the opinion is negative, and left blank if the review does not mention it. The dataset contains 3,786 aspect labels (2,696 positive and 1,090 negative), with 2.66 aspects per review on average. Reviews were collected from five sources: a Google Form shared with Bangla-speaking viewers, reviews gathered in person, YouTube comments, and user reviews on IMDb and Letterboxd. Every review comes with English text. Two annotators labelled the reviews, and 100 reviews were labelled by both to measure agreement (Cohen's kappa = 0.775; observed agreement = 88.56%). The repository contains two archives: 1. RawDataset.zip: the reviews and labels as originally collected and annotated (1,428 rows), before any cleaning. 2. BanglaCineAspect_preprocessed_dataset.zip: the cleaned dataset used for modelling, containing: - Bangla_Movie_Reviews_ABSA_preprocessed.csv and .xlsx: 1,421 reviews with 14 columns (review_id, review_bn_clean, review_bn_original, review_en, source_excel_row and the nine aspect columns). The CSV is UTF-8 with BOM so Bangla displays correctly in Excel. - data_dictionary.csv: description of every column and of the label coding. - dropped_rows.csv: the 7 exact-duplicate rows removed during cleaning, with their original row numbers. - README.txt: a short summary of the dataset and cleaning steps. The data can be used to train and evaluate aspect-level sentiment classifiers for Bangla, to compare deep learning and transformer models on a low-resource language, and to study cross-lingual sentiment analysis using the parallel English text.

Files

Steps to reproduce

1. Collection: Reviews were gathered from a Google Form, in-person collection, YouTube, IMDb and Letterboxd. Only reviews that discussed at least one of the nine aspects were kept. Usernames, profile links and other identifying details were not collected. [Add how the English text was produced.] 2. Cleaning (applied to the Bangla text to create review_bn_clean): (a) Unicode NFC normalisation; (b) removal of URLs; (c) replacement of zero-width spaces, byte-order marks and non-breaking spaces with normal spaces; (d) removal of emojis, symbols and characters outside the Bangla block, Latin letters, digits and basic punctuation (. , ! ? । ' -); (e) collapsing of repeated punctuation, whitespace normalisation and lower-casing of embedded English words. The original text is kept unchanged in review_bn_original. 3. Filtering: rows with empty text or no aspect label were checked for (none found), and 7 rows that were exact duplicates in both text and labels were removed, leaving 1,421 reviews. Removed rows are listed in dropped_rows.csv. 4. Annotation: For each review, the annotator read the full text and marked each aspect as 1 (positive), 0 (negative) or blank (not mentioned). 100 reviews were labelled by both annotators to compute agreement. 5. Loading the data (Python): import pandas as pd df = pd.read_csv("Bangla_Movie_Reviews_ABSA_preprocessed.csv", encoding="utf-8-sig") To build a task for one aspect, keep the rows where that column is 0 or 1, e.g. sub = df[df["Acting"].notna()][["review_bn_clean", "Acting"]] 6. Benchmark setup used in the accompanying article: for each aspect, a stratified 70/30 train/test split, with 15% of the training part used for validation, repeated with random seeds 42, 43 and 44. Macro-F1 is reported as the main metric.

Institutions

Categories

Natural Language Processing

Licence