Explicitly Defined Sampling Categories for ESA on a Bipartite Graph
Mateusz Ozga, Julian Szymański · 2020
This paper presents a study of extensions of the Explicit Semantic Analysis (ESA) used for text representation. The standard ESA algorithm leads to allocation of blocks of words in (O |V2| × n) time on average, where n is the size of the words in text corpora being the subject of analysis and |V2| stands for the size of the vocabulary. Proposed extensions have been based on the selection of training data for ESA and employs for that purpose the category structure of Wikipedia called CESA. The paper proposes the metrics for evaluation of the quality and test the performance of the methods in the function of the training data size. We also study the influence of these methods on the quality of the representation. We established that the total number of queries in case of training is (O |D ⊆ V2| × n). Furthermore, the CESA method leads to allocation of blocks of words in O (|V1| × |V2| × n) time on average, and O (|V1| ×|V2| × n) time on worse case.