📄 Manually Annotated Spelling Error Corpus for Amharic

Summary

Amharic is Ethiopia's official language and has many speakers, but it has few computational resources. Before this work, no shared benchmark existed for testing Amharic spell checkers. The authors argue that collecting errors from keystroke logs or typing games has drawbacks: crowdsourcing platforms support only some languages, and people playing a game may not write as they normally would. They therefore annotated real text by hand: 367 sentences from Deutsche Welle and Voice of America news, a retyped document, and the errata list of a well-known novel. Each error is labelled as either a non-word error or a real-word error, and the surrounding sentence is kept. Corrections follow what the original writer meant rather than prescriptive Ge'ez-based spelling rules. This matters because Ethiopic has homophonic characters that modern writers often swap. The corpus contains 372 misspellings, and about 94% of them are within two edits of their correction. The resource is small, but it is released openly and gives Amharic spelling correction a reference point for evaluation.

Key points

• The corpus has 367 annotated sentences with 372 misspellings: 287 non-word errors (77.15%) and 85 real-word errors (22.85%). It is published at github.com/andmek/ErrorCorpus.
• Damerau-Levenshtein distances between errors and corrections were 1 for 290 cases (77.96%), 2 for 59, 3 for 18 and 4 for 5. Roughly 94% fall within distance 2.
• Before measuring distance, the Ethiopic script was transliterated to Latin letters using the phonetic mappings of the Google and Keyman input methods.
• The Pearson correlation between misspellings and word length was 0.061, so the authors found no meaningful linear link.
• Words written in informal dialect forms were tagged as non-word errors. Many cognitive errors come from homophonic characters in Ethiopic.

Interesting ideas

• I1. Corrections record the form the writer intended, not a prescriptive standard, so the gold data reflects real usage instead of an ideal norm. — For Mova (tokens): Our normalizer could map noisy forms to the writer's likely target while keeping accepted spelling variants, such as US/UK pairs, as valid instead of "correcting" them.
• I2. Edit distance is measured after converting characters into the keystroke-level scheme people actually type with, so the distance reflects real slips rather than script units. — For Mova (tokens): Candidate generation for typo repair could score distances in a keyboard-aware or phoneme-aware space instead of raw characters.
• I3. Each error is stored inside its full sentence, which is what allows real-word errors to be detected at all, since the misspelled word is itself a valid word. — For Mova (learn): Evaluation sets for our repair rules should keep the whole sentence, so we can test whether a valid but wrong word is caught through its context in the tree or frames.
• I4. The authors cite the argument that a corrector able to handle poor spellers will likely handle typing slips too, but not the other way round. — For Mova (learn): We should evaluate on the harder error distribution, such as learner or noisy web text, so that robustness carries over to cleaner input.
• I5. Since most errors fall within two edits, a corrector can limit its candidate search to that radius and lose little coverage. — For Mova (tokens): Capping the lexicon lookup for out-of-vocabulary tokens at distance 2 keeps the search cheap, and it stays explainable because the edit path can be shown.