Work

A Rule-Based Parser in Comparison with Statistical Neuronal Approaches in Terms of Grammar Competence

Year: 2024

Type: article

Abstract: The “Easy Language” standard was created to help individuals with cognitive disabilities understand texts more easily. Typically, text simplification is performed by language experts and is available for limited materials. We introduce a new software tool designed to analyze and simplify any text according to the “Easy Language” rules. This tool uses a rule-based system, conducting a full grammatical analysis of each sentence and then simplifying it into a grammatically correct form. Unlike neuronal approaches, which are based on statistics and are very popular today, our rule-based approach explicitly addresses language ambiguities by examining all possible interpretations and eliminating the incorrect ones. The purpose of the present study is to compare the performance of our rule-base parser with two state-of-the-art statistical parsers, one based on dependencies between words (SpaCy parser) and the other based on linguistic constituents (Stanford parser). Although large language models (LLMs), which are the technical basis of the software ChatGPT, were not designed specifically for grammatical parsing, because of their popularity, users, especially language learners, often ask them grammatical questions as well. Therefore, we use LLMs as supplementary models for comparison. LMMs produce grammatically correct text on any topic; however, their grammar knowledge is implicit within the trained weights. To evaluate how well state-of-the-art methods can perform a grammatical analysis, we parse ten sentences with our tool, the statistical parsers from SpaCy and Stanford, and ask two LLMs equivalent grammar questions. The results show that our rule-based method provides a more informative and reliable grammatical analysis compared to these two parsers and outperforms LLMs in that specific task.

Source: Applied Sciences

DOI: 10.3390/app15010087

Authors: Simon M. Strübbe, Alexander T. D. Grünwald, Irina Sidorenko, Renée Lampe

Institutions: TUM Klinikum, Technical University of Munich

Language: English

Topic: Natural Language Processing Techniques

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $2,728

Funder: Stiftung Würth