📄 An Amharic News Text classification Dataset

Summary

Amharic is spoken by about 100 million people in Ethiopia and is written in its own Fidel script. It has plenty of raw text but very little labeled data. Earlier Amharic topic-classification studies used small, single-source collections that were rarely shared. The authors build and openly release a news classification corpus of 50,706 articles gathered from eleven local and international Amharic news websites. Each article's category comes from the tag the website itself assigned; the authors checked these labels by hand and removed noise. Every record also keeps the source site, title, view count and publication date. As a reference point, they train Naive Bayes classifiers on raw term counts and on TF-IDF features. Preprocessing is whitespace tokenization, punctuation removal and normalization of interchangeable homophone characters. Both classifiers reach roughly 62% accuracy. The resource matters because it is at least five times larger than previous benchmarks and draws on many sources. It gives the Amharic community a shared testbed for stronger models and for experiments with class imbalance.

Key points

• The corpus has 50,706 news articles in 6 classes: local news (20,564), sport (9,812), politics (9,307), international news (6,515), business (3,873) and entertainment (635), so it is strongly imbalanced.
• The articles come from 11 sites, including Walta (8,764), Soccer Ethiopia (8,595), Fana Broadcasting (7,700), VOA Amharic (6,943), Ethiopian Reporter (6,280), Ethiopian Press (5,597) and BBC Amharic (816).
• Prior Amharic classification work used between 200 and 15,000 articles, each from a single source. Earlier studies had also complained that no standard corpus existed.
• The baselines are Naive Bayes with count-vectorizer features (62.2%) and with TF-IDF features (62.3%).
• The data and baseline scores are on GitHub. Planned future work covers word embeddings, transformer models, text augmentation and methods for handling imbalance.

Interesting ideas

• I1. Instead of annotating from scratch, take the category tags publishers already attach and then verify them by hand, which makes labeling much cheaper. — For Mova (learn): Existing structured markup such as Wikipedia infobox fields or genre tags could serve as cheap silver labels, with human or witness checks applied only to the disputed ones.
• I2. Merge script characters that sound the same and are used interchangeably, so one word does not show up as several vocabulary entries. — For Mova (tokens): The en crate could map spelling variants (British/American forms, typographic quotes, ligatures) to a canonical form before lemma lookup, while keeping the original surface form for explanations.
• I3. Collect from many outlets on purpose to diversify the text, and record the source of every item. — For Mova (learn): Keeping provenance on training and test items would allow evaluation splits by source, which shows whether a parser or reader has overfit to one outlet's style.
• I4. Publish a deliberately simple baseline next to the data so later work has a clear floor to beat. — For Mova (learn): Each Mova reader benchmark could report a bag-of-words score as a negative control, showing how much the explicit tree-and-frame pipeline actually adds.
• I5. Store per-article metadata such as posting date and source even when the main task does not need it, because other uses may appear later. — For Mova (world): Documents fed into the world model could carry date and publisher as provenance attributes, which helps with ordering events and judging conflicting claims across texts.