📄 Semantic Enrichment of Nigerian Pidgin English for Contextual Sentiment Classification

Summary

Social media text from Nigeria often mixes Standard English with Nigerian Pidgin, within a single sentence. Many Pidgin words look like English but carry different meanings: "ginger" signals motivation, and "tank" expresses thanks. Off-the-shelf English sentiment tools therefore misjudge such text. The authors adapt VADER, a rule-based, lexicon-driven sentiment scorer built for social media. Fluent Pidgin speakers translated entries of the English VADER lexicon into one-word Pidgin equivalents, which yielded 300 Pidgin tokens with VADER-style scores. The original and the extended lexicon were then each used to label roughly 14,000 Pidgin tweets, and their outputs were compared with labels from expert Pidgin speakers. The extended lexicon produced labels closer to the human judgments. Sample tweets show neutral or wrong-polarity scores flipping to the correct polarity. Humans still caught nuances that the rules missed. The work matters because it releases two resources for an under-served, widely spoken contact language: a VADER-compatible Pidgin sentiment lexicon and a gold-labelled tweet corpus.

Key points

• Released resources: 300 Pidgin sentiment tokens with scores that plug into VADER's roughly 9,000-entry lexicon, and about 14,000 Nigerian Pidgin tweets with gold sentiment labels.
• Nigerian Pidgin is the primary language of an estimated 3–5 million people and a second language for up to 75 million Nigerians.
• Most English sentiment words translate into Pidgin phrases rather than single tokens, which limited lexicon coverage. Some Pidgin words, such as kasala, gbege and para, correspond to several English entries, so their scores were averaged.
• Three annotators from Nigeria's South South region labelled the data, and at least one other person checked each label.
• Evaluation is qualitative: a table of example tweets shows compound scores before and after the update alongside the expert label. No aggregate accuracy figure is reported.

Interesting ideas

• I1. When one dialect word corresponds to several entries in a source lexicon, giving it the mean of their scores is a cheap way to inherit values without new annotation.
• I2. Look-alike words with shifted meanings ("false friends") are fixed by layering a small dialect overlay on top of an existing lexicon, not by retraining.
• I3. The evaluation lines up three judgments per sentence: the old system, the updated system, and a human expert. This makes every flip visible and attributable to the lexicon change.
• I4. Each label is created by one annotator and confirmed by a second, which is a low-cost quality gate for gold data in a niche variety.
• I5. Translating a lexicon showed that many concepts surface as multiword phrases in the target variety. Restricting to single tokens loses coverage.

For Mova

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

• I1 (world): compiled value or polarity tables could assign dialect or slang lemmas the averaged values of their WordNet-linked standard synonyms, with the source entries recorded for explainability.
• I2 (tokens): a register- or dialect-keyed lemma/sense overlay in crate en could redirect forms like "tank" to their local sense before tagging.
• I3 (learn): rule or lexicon changes could be evaluated as per-sentence before/after diffs against gold, so each fix and each regression can be explained.
• I4 (learn): small gold sets that check silver data from the witness parsers could use the same two-person check.
• I5 (frames): sentiment- or attitude-bearing expressions in prag could be stored as multiword entries matched on the UD tree, not as single tokens.