BENCHMARKING DYNAMIC CONVOLUTIONAL NEURAL NETWORK WITH LANGUAGE MODELING PRE-TRAINING FOR SENTIMENT AND QUESTION CLASSIFICATION TASKS.

One-dimensional convolutional models are used for various natural language processing tasks. This study revisits Dynamic Convolutional Neural Network (DCNN) architecture. The study investigates the effect of language modeling pre-training on Wikicorpus on published experiment results for DCNN. Therefore, the reference study integrates a top layer for the language-modeling training into DCNN. Also, benchmarks were reported for the original DCNN compared to the pre-trained language model version. The revisited model was then benchmarked for sentiment classification and question classification tasks. Benchmarks include transfer learning from pre-trained DCNN for language modeling and ground-up trained versions of DCNN on Stanford Sentiment Tree Bank and TREC Question Classification datasets. [ABSTRACT FROM AUTHOR]