LSTM-based sentiment analysis model for inter-organizational trust prediction

Sentiment analysis for inter-organizational trust prediction is crucial for collaborative relationships, as its dynamic and emotional analysis helps enhance partnerships. Yet, existing models mainly focus on unidimensional text polarity classification, missing temporal dependencies in trust relationships. To overcome this, an LSTM-based model is proposed. It has a trust fusion layer that uses convolution to transform sentence-level word vectors into N-gram matrices, capturing local context. The memory attention module with a BiLSTM extracts fused vectors of temporal dependencies and strengthens key content extraction via external memory. A sentence-level attention network boosts model sensitivity. On the Trust dataset, the model outperforms SOTA models, achieving 92.4% accuracy in commercial and 89.7% in university partnerships, and improving computational efficiency. It sets a new paradigm for trust evolution analysis and supports trust prediction and risk management.