This data descriptor presents two fully synthetic corpora for sentiment analysis and named entity recognition (NER) in Uzbek. The first corpus contains 12,000 hybrid synthetic sentences generated from templates with lexical randomization, automatic insertion of named entities (PER/ORG/LOC), lexicon-based polarity scoring, and a controlled emoji distribution. The second corpus includes 3000 “manual-style” sentences designed to resemble short, naturally structured messages. Although the manual-style subset was initially intended to be emoji-free, the released version includes a 39.6% emoji presence (sentences containing at least one emoji) to maintain comparability in emotional markers across corpora. Both corpora are released in CSV, XLSX, and JSONL formats and share a unified schema (id, text, sentiment, entities, entity_type, polarity_score, polarity_source, token_count, emojis, emoji_position, emoji_sentiment, conflict_flag, sentiment_from_polarity_score, split). The dataset is publicly available via Mendeley Data (DOI: 10.17632/y2d5pcyrzz.3).
The Sentiment140 collection, which contains 1.6 million labeled tweets, is used in this research, to study sentiment classification. Cleaning, tokenization, and sequence padding to 50 tokens were among the many preprocessing steps that were applied to the data. Three models were investigated: aUtilizing the VADER model as a baseline, a 1D Convolutional Neural Network (CNN), and a dual-directional Long ShortTerm Memory (BiLSTM) network. While the BiLSTM contained $\mathbf{6 4}$ units with dropout for regularization, the CNN was made up of dense layers, a conversion layer with 128 filters to perform global peak pooling, and a layer to embed with 128 dimensions. Both models had a batch size of 128 and were trained over 20 epochs. The BiLSTM dramatically outperformed VADER, exhibiting an accuracy of roughly 8587% on the test set with low loss and successfully identifying specific and sequential developments in informal web-based text.