Modality estimation methods in internal training reflection texts using BERT.

Abstract

In-house training includes opportunities for employees to reflect on the training through the writing of written reflections. Its purpose is to create opportunities to look back and think about "how the content learned in the training could be utilised in subsequent work," and to allow employees to discover their own issues and set future action goals. In addition, it is possible to measure the effectiveness of the training, and the planners can also discover issues in the training itself, such as the appropriateness of the content, which is useful for improving future training programs. However, there has been little research on these employee reflection texts, especially text analysis using natural language processing (NLP), and there is a need for methods to find useful uses for the texts. By taking a NLP approach to the accumulated data of reflective statements, it is possible to estimate the degree of understanding of the training content and to grasp the degree of growth of employees over time. In addition, there is a potential for proposing training content suited to each employee and improving the efficiency of training as a further development of this approach. As a first step towards the realization of these possibilities, this study proposes automatic modality estimation methods focusing on sentence-final expressions in retrospective self-reflection sentences. Modality in Japanese includes the speaker’s subjective semantic content, which indicates the speaker’s judgement of facts and his/her attitude towards speech and communication to the listener. Therefore, we believe that the modality part of self-reflection sentences on training texts contains important information that leads to the understanding of each employee’s behavior and intentions. Experimental results show that the proposed methods enables modality estimation with higher accuracy than manually formulated rules based on sentence structure. The results also show the possibility of developing the proposed methods to be applied to a deeper analysis approach to reflection sentences.