Estimating Work Mentality from One-on-one Meeting Memo

Abstract

It is an important issue for HR managers to assess the work mentality of employees to provide appropriate support, for the sake of better human capital management. While employee surveys about their own work mentality is commonly used, it is often difficult to conduct such surveys frequently due to the burden on the employees. Thus, it will practically useful if mentality can be estimated without conducting employee surveys. In this study, we developed models to estimate the work mentality of employees from memos from “one-on-one” meetings held between a manger and a member, which have been popular in recent years. With 3186 memos from actual one-on-one meetings and results from mentality surveys, we compared three machine learning models, including a model with Doc2Vec, BERT-sentiment, and fine-tuned BERT. The highest performance model was fine-tuned BERT with F1 0.69 and Accuracy 0.75. Furthermore, we compared the classification performance between fine-tuned BERT and human estimations from the same memos. Results indicated that fine-tuned BERT outperformed non-experts and was comparable with an expert.