| Summary: | To effectively process textual data, many approaches have been proposed to create text representations. The transformation of a text into a form of numbers that can be computed using computers is crucial for further applications in downstream tasks such as document classification, document summarization, and so forth. In our work, we study the quality of text representations using statistical methods and compare them to approaches based on neural networks. We describe in detail nine different algorithms used for text representation and then we evaluate five diverse datasets: BBCSport, BBC, Ohsumed, 20Newsgroups, and Reuters. The selected statistical models include Bag of Words (BoW), Term Frequency-Inverse Document Frequency (TFIDF) weighting, Latent Semantic Analysis (LSA) and Latent Dirichlet Allocation (LDA). For the second group of deep neural networks, Partition-Smooth Inverse Frequency (P-SIF), Doc2Vec-Distributed Bag of Words Paragraph Vector (Doc2Vec-DBoW), Doc2Vec-Memory Model of Paragraph Vectors (Doc2Vec-DM), Hierarchical Attention Network (HAN) and Longformer were selected. The text representation methods were benchmarked in the document classification task and BoW and TFIDF models were used were used as a baseline. Based on the identified weaknesses of the HAN method, an improvement in the form of a Hierarchical Weighted Attention Network (HWAN) was proposed. The incorporation of statistical features into HAN latent representations improves or provides comparable results on four out of five datasets. The article presents how the length of the processed text affects the results of HAN and variants of HWAN models.
|
|---|
