Dataset and Code for Automating Systematic Literature Review with Responsible AI: A PRISMA + NLP Framework Applied to Intergenerational Pedestrian Environments
Description
This dataset accompanies a systematic literature review that investigates how pedestrian environments support intergenerational exchange, with a focus on extracting design-relevant features using Natural Language Processing (NLP) tools. The study integrates the PRISMA framework with AI-assisted techniques such as semantic similarity filtering, question answering (QA) models, and TF-IDF-based keyword analysis. The dataset includes: Metadata of 5,571 academic articles retrieved from Scopus A filtered list of 60 semantically relevant articles Extracted QA responses for five thematic domains (physical environment, social perception, design strategies, behavioral aspects, and barriers) TF-IDF keyword importance scores and categorized outputs Python scripts that implement the entire NLP-based analysis pipeline The data helps identify key spatial, perceptual, and behavioral elements that facilitate or inhibit intergenerational interaction in urban pedestrian spaces. It can be reused for benchmarking NLP-based literature analysis, urban design research, or Responsible AI implementation in systematic reviews. All intermediate results are traceable, and the analysis is reproducible using the provided source code and instructions.
Files
Steps to reproduce
This dataset was produced as part of a systematic literature analysis on intergenerational pedestrian environments using a PRISMA-guided and NLP-enhanced pipeline. Data Collection: We collected 5,571 academic papers from the Scopus database using predefined Boolean queries. Semantic Screening: Titles, abstracts, and author keywords were embedded using the SentenceTransformer model (all-MiniLM-L6-v2) and filtered based on cosine similarity to the topic vector. QA-based Information Extraction: A RoBERTa-based QA model (deepset/roberta-base-squad2) was used to extract sentence-level answers to five domain-specific research questions. TF-IDF Analysis: Extracted QA answers were categorized and vectorized using TF-IDF to compute keyword importance. Quantile-based thresholds were used to classify terms into four salience levels. Code Execution: All analysis steps are implemented in Python 3 and organized under the src/ directory. Required packages are listed in requirements.txt. To reproduce the full analysis, run src/qa_keyword_pipeline_en.py after installing dependencies. Intermediate datasets and processed outputs are included in this repository.
Categories
Funders
- Korean Researcher Information