Despite empirical successes of recurrent neural networks (RNNs) in natural language processing (NLP), theoretical understanding of RNNs is still limited due to intrinsically complex non-linear computations. We systematically analyze RNNs’ behaviors in a ubiquitous NLP task, the sentiment analysis of movie reviews, via the mapping between a class of RNNs called recurrent arithmetic circuits (RACs) and a matrix product state. Using the von-Neumann entanglement entropy (EE) as a proxy for information propagation, we show that single-layer RACs possess a maximum information propagation capacity, reflected by the saturation of the EE. Enlarging the bond dimension beyond the EE saturation threshold does not increase model prediction accuracies, so a minimal model that best estimates the data statistics can be inferred. Although the saturated EE is smaller than the maximum EE allowed by the area law, our minimal model still achieves training accuracies in realistic sentiment analysis data sets. Thus, low EE is not a warrant against the adoption of single-layer RACs for NLP. Contrary to a common belief that long-range information propagation is the main source of RNNs’ successes, we show that single-layer RACs harness high expressiveness from the subtle interplay between the information propagation and the word vector embeddings. Our work sheds light on the phenomenology of learning in RACs, and more generally on the explainability of RNNs for NLP, using tools from many-body quantum physics.