Agricultural Text-Pair Matching with Limited Labels: A Lightweight Two-Stage Semi-Supervised Model

Expert annotation is costly for agricultural text-pair matching, whereas unlabeled pairs are more readily available. We investigate a two-stage semi-supervised model, Semi-TM, that combines a bidirectional long short-term memory (BiLSTM) encoder based on the enhanced sequential inference model (ESIM) with supervised contrastive learning and a masked-view objective. We evaluate a Chinese sentence-pair dataset containing agricultural and general-domain questions and three public benchmarks. On the experimental dataset, at a labeled-data proportion of 30%, Semi-TM achieves a macro-F1 of 0.8943±0.0034 across five runs, compared with 0.8625±0.0025 for the cross-entropy baseline. In the public-benchmark evaluations, its macro-F1 is 0.8731±0.0030 on AFQMC, 0.8991±0.0045 on LCQMC, and 0.8447±0.0034 on PAWS-X under the dataset splits and labeled subset sizes used in this study. Ablations examine supervised contrastive learning, unlabeled examples, masking, and training order. A per-pair analysis further characterizes performance across sentence lengths and identifies representative agricultural matching errors. The evaluated model has fewer parameters and a smaller storage footprint than ESIM but has higher single-pair CPU latency.