A Multidisciplinary Multimodal Aligned Dataset for Academic Data Processing.

Area of Science:

• Bibliometrics
• Scientometrics
• Data Science

Background:

• Academic data processing relies heavily on textual data in scientometrics and bibliometrics.
• Existing datasets overlook the significance of visual elements in scholarly articles.
• There is a need for comprehensive datasets that integrate both textual and visual information.

Purpose of the Study:

• Introduce a novel, multidisciplinary multimodal aligned dataset (MMAD) for academic data processing.
• Enhance research capabilities in areas like trend analysis and citation recommendation by incorporating visual data.
• Provide a foundation for new research avenues in academic data analysis.

Main Methods:

• Compiled a dataset of over 1.1 million peer-reviewed scholarly articles.
• Integrated text with aligned visual elements and associated metadata.
• Developed a Language Model-based quality validation method using specific prompts to assess text-to-visual alignment accuracy.
• Assessed dataset representativeness by comparing country/region distribution with SCImago benchmarks.

Main Results:

• The multimodal aligned dataset (MMAD) contains over 1.1 million scholarly articles with aligned text and visuals.
• A novel validation method using Language Models demonstrates effective quality control for text-to-visual alignment.
• MMAD's representativeness was evaluated against established benchmarks.

Conclusions:

• The introduction of MMAD addresses the gap in existing datasets by incorporating aligned visual elements.
• MMAD facilitates advanced academic data processing, enabling new research in automated caption generation and figure trend analysis.
• This dataset offers a fertile ground for future advancements in scientometrics and bibliometrics.