Search

ChatGPT is an AI language model trained on vast amounts of text data, including scientific papers, providing a comprehensive understanding of catalysis. However, its reliability in catalysis research is unknown. To evaluate reliability, we compared a ChatGPT-generated review article on heterogeneous catalysts for higher alcohols synthesis by CO2 hydrogenation to published peer-reviewed papers. Although the ChatGPT review article covers most necessary parts, it lacks sufficient discussion of the reaction mechanism. The core sections are too general, being not specific enough to the topic, and contain errors. The lack of citations further increases unreliability. While ChatGPT can provide much content on catalysis, it is insufficient and inaccurate for research on specific topics.

This paper presents the development and testing of a text data mining tool to assist in the selection and assignment of adjunct faculty to teach STEM courses. The tool scores the resume of a faculty member against course descriptions in a STEM graduate program. The tool returned a similarity score between a resume and course descriptions, which was then used as an indicator of faculty suitability to teach courses in the program. We enhanced the original tool with an improved user interface and deployed it to search for new faculty searches and in the process of assigning courses. A TD-IDF text analytic technique was used for similarity scoring. Our research question was to investigate whether a similarity-scoring tool for faculty resumes against course descriptions would be useful in the search and assignment process to hire faculty to teach specific courses. As part of our methods, we developed a friendly user interface to the existing tool using a student-centered coding contest. We applied the tool to the hiring and assignment of adjunct faculty. We measured success as the processing of a large number of open positions in a relatively short period of time and found a significantly high number of good fits between faculty and their course assignments. We investigated whether the scoring system positively correlated with the courses assigned to them.We successfully filled over 50 unassigned courses with appropriate faculty over a period of three months, where 30% were new hires. In the process, we discovered that the vast majority of the incumbent’s similarity scores positively correlated to the courses assigned to them. This generated sufficient confidence that the description scoring system has been integrated as part of our faculty hiring and assignment processes in our programs.

A literature search can be an arduous process, especially for novice researchers. We have developed a tool that allows a researcher to rank order a list of references that are returned by a keyword-based search engine, based on similarity to known exemplars. This significantly accelerates literature searches by novices. Our research question was: can we produce a text-analytic tool that, when used by an inexperienced scholar, rank-orders a list of references against an exemplar, so that the time needed to find relevant literature is reduced, and the literature survey section of their paper will be superior. An experiment was set up where one course section used the tool to produce the literature review section of a thesis proposal, and the other class used traditional literature research tools. We surveyed both sections to self-report the time used for the literature search. We found some time savings by some of the students using the tool. We also provided blind, randomly selected pairs of completed proposals to SME faculty who teach that same class to assess the quality of the literature sections of the samples. We found that the tool-using section of students reported significantly less time to do the literature search, and the quality of their literature review produced had a significantly higher quality.

We applied text data mining techniques from machine learning to position (job) descriptions posted on NYU’s job search site, Bureau of Labor Statistics (BLS) standard U.S. job descriptions, course descriptions, and curricula descriptions. Our work compared Term Frequency-Inverse Document Frequency (TD-IDF) to Latent Semantic Indexing (LSI) and found that TD-IDF was preferred in this application. We used TD-IDF to measure the extent of coherence among the collections of our documents. We then leveraged those measurements to developed novel approaches to assist students and curricula designers in answering these questions: (1) for students, given an interest in specific jobs, which degrees and courses are most relevant; (2) for students, given courses that have been taken, which jobs are most likely to result in initial interviews; (3) for curricula designers, how aligned are degree programs with specific groups of jobs (for example, with STEM jobs); (4) for curricula designers, to what extent do current and proposed degrees address different job opportunities. Other similar applications are possible by composing our Python and JMP code. Our work could be extended by providing open source implementation of the algorithms.

We applied text data mining techniques from machine learning to position (job) descriptions posted on a university job search site, Bureau of Labor Statistics (BLS) standard U.S. job descriptions, ...