It is known that in the tasks of natural language processing, the representation of texts by vectors of fixed length using word-embedding models makes sense in cases where the vectorized texts are short.The longer the texts being compared, the worse the approach works. This situation is due to the fact that when using word-embedding models, information is lost when converting the vector representations of the words that make up the text into a vector representation of the entire text, which usually has the same dimension as the vector of a single word.This paper proposes an alternative way for using pre-trained word-embedding models for text vectorization. The essence of the proposed method consists in combining semantically similar elements of the dictionary of the existing text corpus by clustering their (dictionary elements) embeddings, as a result of which a new dictionary is formed with a size smaller than the original one, each element of which corresponds to one cluster. The original corpus of texts is reformulated in terms of this new dictionary, after which vectorization is performed on the reformulated texts using one of the dictionary approaches (TF-IDF was used in the work). The resulting vector representation of the text can be additionally enriched using the vectors of words of the original dictionary obtained by decreasing the dimension of their embeddings for each cluster.A series of experiments to determine the optimal parameters of the method is described in the paper, the proposed approach is compared with other methods of text vectorization for the text ranking problem – averaging word embeddings with TF-IDF weighting and without weighting, as well as vectorization based on TF-IDF coefficients.
Pretrained Transformers for Text Ranking: BERT and Beyond
