Since Hinton and Salakhutdinov published their landmark science paper in 2006 ending the previous neural-network winter, research in neural networks has increased dramatically. Researchers have applied neural networks seemingly successfully to various topics in the field of computer science. However, there is a risk that we overlook other methods. Therefore, we take a recent end-to-end neural-network-based work (Dhingra et al., 2018) as a starting point and contrast this work with more classical techniques. This prior work focuses on the LAMBADA word prediction task, where broad context is used to predict the last word of a sentence. It is often assumed that neural networks are good at such tasks where feature extraction is important. We show that with simpler syntactic and semantic features (e.g. Across Sentence Boundary (ASB) N-grams) a state-ofthe- art neural network can be outperformed. Our discriminative language-model-based approach improves the word prediction accuracy from 55.6% to 58.9% on the LAMBADA task. As a next step, we plan to extend this work to other language modeling tasks.