Sentence transition matrix: An efficient approach that preserves sentence semantics

Sentence embedding is an influential research topic in natural language processing (NLP). Generation of sentence vectors that reflect the intrinsic meaning sentences crucial for improving performance various NLP tasks. Therefore, numerous supervised and unsupervised sentence-representation approaches have been proposed since advent distributed representation words. These evaluated on semantic textual similarity (STS) tasks designed to measure degree information preservation; neural network-based models typically deliver state-of-the-art performance. However, these limitations they learnable parameters thus require large amounts specific types labeled training data. Pretrained model-based approaches, which become a predominant trend field, alleviate this issue some extent; however, it still necessary collect sufficient data fine-tuning process necessary. Herein, we propose efficient approach learns transition matrix tuning vector capture latent meaning. Our method has two practical advantages: (1) can be applied any method, (2) robust STS with only few examples.