Journal of Data Science

Text clustering can streamline many labor-intensive tasks, but it creates a new challenge: efficiently labeling and interpreting the clusters. Generative large language models (LLMs) are a promising option to automate the process of naming text clusters, which could significantly streamline workflows, especially in domains with large datasets and esoteric language. In this study, we assessed the ability of GPT-3.5-turbo to generate names for clusters of texts and compared these to human-generated text cluster names. We clustered two benchmark datasets, each from a specialized domain: research abstracts and clinical patient notes. We generated names for each cluster using four prompting strategies (different ways of including information about the cluster in the prompt used to get LLM responses). For both datasets, the best prompting strategy beat the manual approach across all quality domains. However, name quality varied by prompting strategy and dataset. We conclude that practitioners should consider trying automated cluster naming to avoid bottlenecks or when the scale of the effort is enough to take advantage of the cost savings offered by automation, as detailed in our supplemental blueprint for using LLM cluster naming. However, to get the best performance, it is vital to test a variety of prompting strategies and perform a small test to identify which one performs best on each project’s unique data.

The North American Product Classification System (NAPCS) was first introduced in the 2017 Economic Census and provides greater detail on the range of products and services offered by businesses than what was previously available with just an industry code. In the 2022 Economic Census, NAPCS consisted of 7,234 codes and respondents often found that they were unable to identify correct NAPCS codes for their business, leaving instead written descriptions of their products and services. Over one million of these needed to be reviewed by Census analysts in the 2017 Economic Census. The Smart Instrument NAPCS Classification Tool (SINCT) offers respondents a low latency search engine to find appropriate NAPCS codes based on a written description of their products and services. SINCT uses a neural network document embedding model (doc2vec) to embed respondent searches in a numerical space and then identifies NAPCS codes that are close to the search text. This paper shows one way in which machine learning can improve the survey respondent experience and reduce the amount of expensive manual processing that is necessary after data collection. We also show how relatively simple tools can achieve an estimated 72% top-ten accuracy with thousands of possible classes, limited training data, and strict latency requirements.

The coronavirus disease 2019 (COVID-19) pandemic presented unique challenges to the U.S. healthcare system, particularly for nonprofit U.S. hospitals that are obligated to provide community benefits in exchange for federal tax exemptions. We sought to examine how hospitals initiated, modified, or disbanded community benefits programming in response to the COVID-19 pandemic. We used the free-response text in Part IV of Internal Revenue Service (IRS) Form 990 Schedule H (F990H) to assess health equity and disparities. We combined traditional key term frequency and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) clustering approaches with a novel Generative Pre-trained Transformer (GPT) 3.5 summarization approach. Our research reveals shifts in community benefits programming. We observed an increase in COVID-related terms starting in the 2019 tax year, indicating a pivot in community focus and efforts toward pandemic-related activities such as telehealth services and COVID-19 testing and prevention. The clustering analysis identified themes related to COVID-19 and community benefits. Generative Artificial Intelligence (GenAI) summarization with GPT3.5 contextualized these changes, revealing examples of healthcare system adaptations and program cancellations. However, GPT3.5 also encountered some accuracy and validation challenges. This multifaceted text analysis underscores the adaptability of hospitals in maintaining community health support during crises and suggests the potential of advanced AI tools in evaluating large-scale qualitative data for policy and public health research.

The National Association of Stock Car Auto Racing (NASCAR) is ranked among the top ten most popular sports in the United States. NASCAR events are characterized by on-track racing punctuated by pit stops since cars must refuel, replace tires, and modify their setup throughout a race. A well-executed pit stop can allow drivers to gain multiple seconds on their opponents. Strategies around when to pit and what to perform during a pit stop are under constant evaluation. One currently unexplored area is publically available communication between each driver and their pit crew during the race. Due to the many hours of audio, manual analysis of even one driver’s communications is prohibitive. We propose a fully automated approach to analyze driver–pit crew communication. Our work was conducted in collaboration with NASCAR domain experts. Audio communication is converted to text and summarized using cluster-based Latent Dirichlet Analysis to provide an overview of a driver’s race performance. The transcript is then analyzed to extract important events related to pit stops and driving balance: understeer (pushing) or oversteer (over-rotating). Named entity recognition (NER) and relationship extraction provide context to each event. A combination of the race summary, events, and real-time race data provided by NASCAR are presented using Sankey visualizations. Statistical analysis and evaluation by our domain expert collaborators confirmed we can accurately identify important race events and driver interactions, presented in a novel way to provide useful, important, and efficient summaries and event highlights for race preparation and in-race decision-making.

Networks are ubiquitous in today’s world. Community structure is a well-known feature of many empirical networks, and a lot of statistical methods have been developed for community detection. In this paper, we consider the problem of community extraction in text networks, which is greatly relevant in medical errors and patient safety databases. We adapt a well-known community extraction method to develop a scalable algorithm for extracting groups of similar documents in large text databases. The application of our method on a real-world patient safety report system demonstrates that the groups generated from community extraction are much more accurate than manual tagging by frontline workers.