Legal case precedents have a considerable impact on the development of litigation strategies. This research uses the neural network language modeling (NNLM) approach to analyze and identify judgment documents of US trademark (TM) litigation cases as precedents of a given target case. In this research, the NNLM has been trained using 4835 TM litigation documents. There are more than 800,000 words in the entire training text set including more than 150,000 vocabularies. The words in TM legal documents are vectorized to train the NN model for e-discovery of semantically correlated precedents and their features. Specifically, non-supervised machine learning (ML) methods, including clustering and Latent Dirichlet Allocation (LDA), are applied to form the TM legal document clusters, topics, and key terminologies used to characterize the TM case descriptions and precedents. The definition of the clusters, topics and corresponding key terms enhance the ability of the system to recommend and explain similar case judgments for any given TM case of interest or a cease and desist letter with detailed claims of infringement. Further, the intelligent approach provides macro and micro views for companies to research TM litigation trends as a means to better protect their brand equity.