📄 Analysis of the Ethiopic Twitter Dataset for Abusive Speech in Amharic

Summary

The paper looks at abusive language in Amharic social media. Amharic is a less-resourced language, and online hate has been linked to real-world unrest in Ethiopia. The authors have harvested tweets in the Ethiopic (Fidel) script every day since mid-August 2014. This produced the Ethiopic Twitter Dataset (ETD), about three million tweets from 154,477 users.

Because Fidel serves many languages, they built an n-gram language identifier following Cavnar and Trenkle. It was trained on general reference corpora for Amharic, Tigrinya and Ge'ez, and tweets matching none of the three were set aside as "other."

Five native speakers drew abusive keywords from Facebook, Twitter and YouTube. The keywords were split into hate and offensive categories using Ethiopia's 2019 draft hate-speech law and the distinction from Davidson et al. Normalized frequencies show that:

• both Amharic tweets and keyword-bearing tweets grow over the years;
• hate keywords far outnumber offensive ones;
• abusive terms are denser on Twitter than in the reference corpus.

The authors explicitly state that keyword counts do not amount to classifying abuse. The work is a groundwork resource for future automatic Amharic abuse detectors.

Key points

• ETD-AM, the Amharic portion, holds about 26.3M tokens (≈1.1M types). GRC-AM, the general reference corpus, holds about 46.4M tokens. Smaller portions exist for Tigrinya (ETD-TI) and Ge'ez (ETD-GE).
• Reference text came from several sources:
  • Amharic: a Hamburg focused-crawler web corpus, OPUS, and the Amharic Web Corpus.
  • Tigrinya: OPUS and the Tigrinya Web Corpus.
  • Ge'ez: STEP and Ge'ez Frontier Foundation data.
• There are 99 hate and 48 offensive keywords. The conclusion's figure of "around 144" does not quite match this total.
• All charts are normalized to parts per million. Tweets from 2015 were excluded because an encoding fault corrupted their storage.
• Named organizations behave differently by source. TPLF, common in mainstream news, is more frequent in GRC-AM. OLF, which mainstream media framed as abusive, shows up more on Twitter.

Interesting ideas

• I1. When several languages share one script, train character n-gram profiles on reference corpora, and send unmatched text to an explicit "other" bucket instead of forcing a label.
• I2. Compare ppm-normalized frequencies in a target domain with those in a general reference corpus to expose terms that are overrepresented in that domain.
• I3. Assign each category through a written external definition (here a legal draft), with everything that fails it falling into a second class.
• I4. The same entity name can be neutral in one source and loaded in another, so its distribution across sources reveals framing.
• I5. The authors state openly that keyword hits measure coverage, not verified abuse.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): A similar profile check could flag foreign or code-switched spans before tagging, so they are marked as unknown rather than silently mis-lemmatized.
• I2 (learn): Contrasting silver-parsed domain text with treebank text this way could surface vocabulary and constructions where the parsers are likely undertrained.
• I3 (frames): Speech-act or irony labels could likewise be tied to explicit, citable criteria, so each label carries its own justification.
• I4 (world): Entities could record the source or register of their mentions, letting the world model note when a referent is being presented pejoratively.
• I5 (learn): Evaluation reports could keep lexical-match statistics clearly apart from validated judgments, which avoids treating surface hits as ground truth.