Abstract: At present, the supply and demand matching of scientific and technological achievements faces the problem of low accuracy. In order to effectively improve the transformation efficiency of scientific and technological achievements, a model of supply and demand matching of scientific and technological achievements based on ERNIE is constructed. First, the ERNIE model is used for in-depth knowledge mining of scientific and technological achievements and enterprise requirements. Word vector representation is input into Transformer structure to capture deep semantic information of text. Multi-head attention mechanism is added to enhance the model’s processing of complex features and significantly improve the accuracy of semantic understanding. Secondly, the semantic representation ability of the model is improved by adding global average pooling and multi-layer feedforward neural network. Finally, the Sentence-BERT model, twin network architecture and contrast loss function are introduced to optimize the sentence embedding representation, and the matching similarity is calculated with cosine similarity. According to the evaluation index, the training accuracy of the model is 2 percentage points higher than that of other models, and the actual accuracy, loss rate and F1 value are increased by nearly 1 percentage point, reaching 94.5%.

Key words: scientific and technological achievements, Supply-demand matching model, Transformer, attention mechanism, Sentence-BERT, text similarity

摘要： 针对科技成果供需匹配准确率偏低的问题,为有效提升科技成果转化效率,构建了一种基于ERNIE的科技成果供需匹配模型。首先,对科技成果和企业需求的文本使用ERNIE模型进行知识深度挖掘,将词向量表示输入Transformer结构捕捉文本的深层语义信息,加入了多头注意力机制增强模型对复杂特征的处理,显著提高语义理解的准确性;其次,通过添加全局平均池化和多层前馈神经网络提升模型的语义表示能力;最后,引入了Sentence-BERT模型,孪生网络架构和对比损失函数对句子嵌入表示进行优化,并结合余弦相似度计算匹配相似度。经评估指标验证,模型的训练准确率高于其他模型2个百分点,实际准确率、损失率以及F1值提高近1个百分点,达到94.5%。

关键词: 科技成果, 供需匹配模型, Transformer, 注意力机制, Sentence-BERT, 文本相似度