Natural Language Processing for Extractive Search

Abstract

the paper presents the linguistic software for intelligent search systems. Large language models have been actively developing for the last decade. LLMs suitable for information search often require extensive resources and have redundant functionality when embedded into targeted information systems. Lightweight approaches to natural language processing are needed. We considered an extractive approach to a “question-answer” search intended to find sentences that answer a question in the specified document. For this, we proposed methods for analyzing the morphology, syntax, and semantics of the natural language. A corpus of Russian language texts containing 8,800 sentences was collected to implement graph-based syntax analysis with a weighting of a completely oriented graph by a forward-propagation artificial neural network. This corpus was also used to produce a set of syntax-oriented vector representations of words, applied in the semantic analysis by using a model based on a continuous bag of words architecture. The sentence ranking by relevance to the question is based on representing the semantics of the natural language text as a strongly connected directed graph, revealing implicit meaningful patterns within the language structures.