This paper demonstrates a simultaneous analysis of construct measurements and a text dataset using a penalized neural network. Penalty methods using L1 (Lasso) or L2 (Ridge) loss functions are known as effective for network optimization and generalization in neural networks. These loss functions represent a priori assumption for the parameters in a model. The proposed method extends this idea to estimate an uncorrelated weight matrix to help interpret the parameter estimates of neural networks in line with consumers’ mental processes. In the empirical analysis, customer engagement measurements and customer review text were collected through a survey of hotel users. The results show that multidimensional customer engagement relates to several unique terms in the customer review text. While considering several limitations of the empirical models, the proposed approach can be used as a technique for understanding customer reviews.