<p>Credit evaluation can effectively control the financial risk of technology-based small and microenterprises (TSMEs) and promote their development. Deep learning has received extensive attention in the field of enterprise credit evaluation because of its excellent performance. However, the black box characteristic of deep learning has restricted its further promotion in the field. Improving the interpretability of credit evaluation models via deep learning has become the key to addressing this issue. For this reason, aligned with the characteristics of TSME, this paper introduces innovation theory and incorporates the innovation capabilities of TSMEs into the credit evaluation process. Then, we design a long short-term memory network model fused with an attention mechanism to extract credit features, with the use of good enterprise credit risk explanatory capabilities, from the unstructured text. Finally, we compare the proposed model with the benchmark model and explain, theoretically, why the proposed model has a better performance and stronger interpretability. The results indicate that attention mechanisms equip LSTM with the ability to focus on textual features of innovation capability that are highly predictive, thereby significantly improving the overall performance of the model. Meanwhile, the proposed model confirms the explanatory power of Schumpeter’s five innovation typologies in evaluating the credit risk of TSMEs.</p>