Research on Sentiment Classification of Online Comments Using Multilayer Fusion with ELMO-BiLSTM

In the era of the rapid expansion of the Internet, individuals are enabled to freely articulate their perspectives on a diverse range of topics via platforms various social media. This phenomenon has significantly elevated the importance of sentiment classification in online reviews. However, traditional models have problems in semantic capture and sequence processing, making it difficult to handle lexical items in metaphors that depend on context, and there are limitations in dealing with the ambiguity and vagueness of text. Consequently, this research presents a novel multi-layer fusion architecture integrating Bi-LSTM and ELMO. At the foundational level, ELMO is incorporated. Primarily tasked with generating word vectors, it possesses the remarkable capacity to dynamically formulate word embeddings in light of the contextual nuances of words, thereby enhancing the precision of semantic capture. Leveraging its potent semantic comprehension capabilities, ELMO undertakes in-depth semantic encoding of the text, furnishing abundant semantic data for subsequent processing steps. The intermediate layer deploys the Bi-LSTM, which is widely acclaimed for its proficiency in handling sequence data and adeptly extracting the anterior and posterior contextual details of the text. It proceeds to conduct further feature extraction based on the semantic information furnished by ELMO. The uppermost layer employs a fully connected layer to refine and amalgamate the fused features, culminating in a more emblematic emotional feature representation. Upon conducting comparative experiments with established models like RNN, CNN, and GRU, the outcomes unequivocally demonstrate that the model devised in this paper exhibits superior performance across all evaluation metrics and is more proficient in addressing the sentiment classification task of online reviews, benefiting businesses in product improvement and marketing.