Challenges of shopping cart content categorization with LSTM neural networks

This case study investigates the performance of different neural network architectures on shopping cart content categorization. Two Bag-of-Words and Four Long Short-Term Memory unit (LSTM) models were tested on the corpus. The noisy corpus consists of brand names, nouns, and adjectives, where the word order in some cases might be cardinal. Although bigrams can effectively capture the existence of word sequences, the models using LSTM units turn out to outperform the other models as they result in higher accuracy. This paper describes how an effective text categorization model can be developed on a corpus with peculiarities and focuses on the differences between the tested models.