[Paper Review] Universal Model for Paraphrasing -- Using Transformation Based on a Defined Criteria --

Masaki Murata, Hitoshi Isahara|ArXiv.org|Dec 5, 2001

Natural Language Processing Techniques9 references19 citations

TL;DR

This paper proposes a universal paraphrasing model that generates diverse paraphrases by applying different transformation criteria—such as similarity, length, frequency, and grammatical validity—using rule-based transformations and corpus-based evaluation. The system successfully implements multiple paraphrasing tasks, including sentence compression, polishing, spoken-language conversion, and question answering, with no incorrect transformations in spoken-language experiments, demonstrating the model's versatility and reusability across NLP applications.

ABSTRACT

This paper describes a universal model for paraphrasing that transforms according to defined criteria. We showed that by using different criteria we could construct different kinds of paraphrasing systems including one for answering questions, one for compressing sentences, one for polishing up, and one for transforming written language to spoken language.

Motivation & Objective

• To develop a universal framework for paraphrasing that supports multiple NLP tasks using a single adaptable model.
• To address the challenge of creating diverse paraphrasing systems without rebuilding from scratch for each task.
• To enable easy adaptation of paraphrasing systems by modifying only the evaluation criteria, not the core transformation rules.
• To demonstrate the model's effectiveness across different paraphrasing types, including question answering, sentence compression, and written-to-spoken language transformation.
• To validate that corpus-based frequency and rule-based transformation can jointly enable accurate, meaning-preserving paraphrasing.

Proposed method

• The model uses a two-module architecture: a transformation module that generates candidate rewrites using hand-written or automatically mined rules, and an evaluation module that selects the best transformation based on predefined criteria.
• Transformation rules are constrained to preserve semantic meaning, ensuring that all rewrites are paraphrases by definition.
• The evaluation module selects transformations using criteria such as semantic similarity, sentence length, frequency in corpora, and grammatical plausibility.
• For frequency-based evaluation, the system compares occurrence counts of word forms in corpora (e.g., 'summarization' vs. 'summarisation') to favor more common, natural forms.
• For spoken-language transformation, the system uses spoken-language corpora to favor expressions like 'ma' and 'toiu' that are common in spoken Japanese.
• The model supports machine translation by applying translation rules and using target-language corpora for probability scoring.

Experimental results

Research questions

• RQ1 Can a single model architecture support multiple types of paraphrasing by varying only the evaluation criteria?
• RQ2 How can transformation rules be designed to preserve meaning while enabling diverse paraphrase types?
• RQ3 To what extent can corpus frequency be used to improve sentence quality, grammaticality, and naturalness?
• RQ4 Can the same transformation rules be reused across different paraphrasing tasks with minimal reconfiguration?
• RQ5 How effective is the model in transforming written language into spoken language without introducing errors?

Key findings

• The model successfully generated paraphrases for four distinct tasks—question answering, sentence compression, sentence polishing, and written-to-spoken language transformation—using the same core transformation rules.
• In spoken-language transformation experiments, no incorrect transformations occurred, and the output included natural spoken-language features like 'ma' and 'toiu'.
• The system achieved a low recall rate in spoken-language transformation, indicating a need for more comprehensive rule sets despite high accuracy in individual cases.
• The use of frequency-based criteria effectively improved sentence quality, such as converting 'summarisation' to 'summarization' when the latter was more frequent in corpora.
• The model demonstrated reusability: components like transformation rules could be repurposed across tasks by changing only the evaluation criteria.
• The integration of rule-based transformations with corpus-based evaluation enabled robust, meaning-preserving paraphrasing across diverse linguistic styles and domains.