NbAiLab/nb-asr-moonlight-sentiment

nb-asr-moonlight-sentiment

Curated 20,000 contrastive sentiment pair dataset derived from alexandrainst/sentiment (Trustpilot and consumer feedback).

Purpose

Designed for contrastive sentence embedding models to learn emotional valence and polarity across Scandinavian languages without benchmark contamination (specifically avoiding any NoRec or newspaper review data).

Composition (20,000 Total Pairs)

• —Norwegian (`no`): 14,000 pairs (7,000 positive / 7,000 negative)
• —Swedish (`sv`): 2,000 pairs (1,000 positive / 1,000 negative)
• —Danish (`da`): 2,000 pairs (1,000 positive / 1,000 negative)
• —English (`en`): 2,000 pairs (1,000 positive / 1,000 negative)

All pairs are strictly matched within language (anchor and positive share identical polarity and language, while opposing polarities serve as in-batch negatives).

Schema

• —task: Task identifier (equivalence)
• —anchor: Review text 1
• —positive: Review text 2 (same polarity and language)
• —negative: Empty string (in-batch negatives recommended)
• —direction: Language code (no, sv, da, en)
• —source: alexandrainst_sentiment
• —metadata_json: JSON string containing language, sentiment, and domain