Advancements in Natural Language Processing: Sentiment Analysis and Named Entity Recognition

Abstract

Sentiment Analysis (SA) and Named Entity Recognition (NER) are fundamental tasks in Natural Language Processing that have traditionally been addressed as independent problems. However, real-world text understanding demands joint modeling of emotional tone and entity-level semantics—knowing that a review praises a specific product feature or that a news article expresses negative sentiment toward a named organization yields richer insight than either task alone. This paper presents SANER (Sentiment-Aware Named Entity Recognizer), a unified multi-task transformer architecture that jointly performs sentiment analysis and named entity recognition through shared contextual representations and cross-task attention fusion. SANER employs a domain-adaptive pre-training strategy, task-specific attention heads, and a conditional random field (CRF) decoding layer for structured entity prediction. Evaluated on six benchmark datasets spanning product reviews, social media, financial news, and biomedical text, SANER achieves state-of-the-art results on both tasks simultaneously: 96.2% F1 on CoNLL-2003 NER (surpassing the previous best by 1.6 points), 94.8% accuracy on SST-2 sentiment classification, and 89.7% accuracy on Twitter sentiment analysis. We further demonstrate that multi-task learning produces mutual improvements—NER F1 increases by 1.4 points when jointly trained with SA, and SA accuracy improves by 2.3 points when jointly trained with NER—confirming that entity awareness and sentiment understanding are synergistic capabilities. Extensive few-shot experiments show that SANER maintains strong performance in low-resource settings, achieving 86.4% F1 with only 10 labeled examples per entity type.