The motivation is straightforward: online public opinion has become a major factor affecting the stability of universities. Traditional monitoring approaches rely mainly on keyword filtering and statistical models, which the authors say fail to capture the contextual semantic features of text and lack the ability to predict the dynamic evolution of sudden opinion events. Rule-based filters and shallow machine learning classifiers can perform basic sentiment sorting, but they generalize poorly when sentences carry long-distance semantic dependencies, implicit expressions, or abrupt topic shifts—precisely the hallmarks of a campus controversy unfolding in real time.

At the heart of the new system sits a bidirectional LSTM, a recurrent neural network that reads each text sequence twice, once forward and once backward. This bidirectional encoding lets the model capture contextual semantic relationships in both directions, solving the problem of unidirectional networks that ignore future information. Inside each LSTM unit, a set of gates—the forget gate, input gate, and output gate—selectively adds or removes information from a memory cell, allowing the network to decide which words matter across long stretches of text. The forget gate discards unnecessary information, the input gate adds new content to the memory cell, and the output gate adjusts the cell state with a tanh activation function to determine what influences the current prediction.

But a strong architecture is only half the battle. Hyperparameters such as the learning rate, the number of hidden layer nodes, and the batch size directly affect model performance, and tuning them by hand is slow and unreliable. The study turns to an improved whale optimization algorithm, a nature-inspired metaheuristic that mimics the bubble-net hunting behavior of humpback whales. The IWOA introduces a nonlinear convergence factor, which decreases with the square of the iteration ratio, keeping the search broad in early iterations and accelerating local convergence later. It also employs a random differential mutation strategy, updating individual positions by taking differences between randomly selected individuals and the current best solution, which maintains population diversity and helps the algorithm avoid getting trapped in local optima. The algorithm searches the Bi-LSTM hyperparameter space using classification accuracy on the validation set as its fitness function.

The second innovation is the attention mechanism, borrowed from the Transformer architecture. Rather than treating every word equally, the self-attention module computes relevance between words through a Query-Key-Value structure, dynamically assigning higher weights to opinion-bearing terms and reducing noise interference. Crucially, the attention weights are learned autonomously from the training data rather than searched externally. Because Transformers excel at global semantic relationships but are comparatively weak at local feature capture, the researchers folded in a recurrent neural network that extracts initial features through a dense layer and activation before passing them into the Transformer encoder, which applies multi-head attention, feedforward transformations, and layer normalization across six encoder layers with eight attention heads.

The full system is a dual-path ensemble. The IWOA-Bi-LSTM path produces temporal feature vectors that encode how context evolves across a sequence, while the RNN-Transformer path produces semantic feature vectors that capture word- and phrase-level sentiment associations. In the feature fusion layer, these two vectors are concatenated into a unified representation, then mapped by a fully connected layer, with a Softmax function outputting probability values for each sentiment category and risk level. The rationale, the authors explain, is that understanding public opinion texts requires both grasping the evolutionary logic of context in time and identifying sentiment tendencies at the word level; a single model often struggles to extract both kinds of information simultaneously, while the parallel architecture preserves the original structure of each feature type before integration.

Training relied on a cross-entropy loss function measuring the difference between predicted and true label distributions, optimized with the Adam optimizer, which adaptively adjusts learning rates per parameter. The team validated the optimization component on the ChnSentiCorp dataset of roughly 9,600 short Chinese sentiment texts and the Weibo Emotions Corpus of 50,000 sentiment-annotated texts across seven categories, both split 8:1:1 into training, validation, and test sets. Running on an NVIDIA RTX 3090 GPU with TensorFlow 2.6.0, the IWOA-Bi-LSTM held a recall of 0.8 at a 0.6 confidence threshold, where rival optimizers combined with LSTM—using grey wolf, genetic algorithm, and particle swarm optimization—declined more steeply. Loss dropped below 0.2 within the first 200 iterations on ChnSentiCorp, with accuracy stabilizing near 0.99.

For the full monitoring model, the researchers compared against Bi-LSTM, WOA-Bi-LSTM, Attention-Bi-LSTM, Transformer, and BERT on THUCNews, a collection of roughly 74,000 Chinese news articles across ten categories, and on about 15,000 education-related tweets. Because THUCNews labels are topic categories rather than sentiment polarities, two annotators with Chinese linguistics backgrounds independently labeled a 20 percent sample, reaching a Cohen’s Kappa consistency coefficient of 0.82. The proposed model’s F1 score of 0.81 edged past BERT’s 0.76 on THUCNews, though its recall of 0.72 fell just short of BERT’s 0.73. On the Twitter dataset, the model led all comparisons in both F1 and recall. In sentiment intensity analysis, the model scored 0.82 on 10,000 data points, 0.07 higher than the second-best model, and its text analysis efficiency reached 78.35 percent on THUCNews and 82.47 percent on Twitter, far above the comparative models, which ranged from roughly 32 to 49 percent.

Those efficiency numbers may be the study’s most striking claim: the dual-path design preserved key semantic information in texts truncated or padded to fixed length far better than any single-path baseline, suggesting real deployment value for the informal, mixed-length writing typical of campus social media. The authors argue the results demonstrate high accuracy and applicability in sentiment classification of campus public opinion and provide a new approach for monitoring that supports development toward efficient systems. The model’s advantages widened as data volume grew, indicating that large-scale opinion analysis is where the ensemble architecture truly pays off.

Limitations remain, and the researchers are candid about them. The data sources were confined to the text modality, so the model does not integrate images or videos from multiple platforms, and coverage of rare dialects and newly emerged internet slang is insufficient. Future work, they write, will explore a multimodal public opinion analysis framework and introduce a real-time incremental learning mechanism to continuously enhance generalization and prediction performance. As universities worldwide grapple with how to keep online discourse from tipping into crisis, the study offers a technical blueprint: read the conversation in both directions, weight the words that carry the emotion, and let an algorithm inspired by whales tune the machinery automatically.

Subject of Research: Artificial Intelligence

Article Title: Campus public opinion monitoring method based on Bi-LSTM and attention mechanism

Article References: Hai, M., & Mo, X. (2026). Campus public opinion monitoring method based on Bi-LSTM and attention mechanism. Discover Artificial Intelligence, 6(1), Article 1523. https://doi.org/10.1007/s44163-026-02413-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02413-7

Keywords: Campus, public, opinion, monitoring, method, based, Bi-LSTM, attention, mechanism, scientific research