Summary of generated sentences

Table 1 presents a summary of the sentences generated using four different models. Representative examples of these generated sentences are illustrated in Table 2.

Cosine similarity

The results of the similarity analysis in question 1 are shown in Figure 1. Notably, a high degree of similarity was observed in the English outputs. Regardless of language or model differences, a generally high similarity trend was evident compared to the other two query texts.

The results of the similarity analysis in question 2 are presented in Figure 2. Compared to question 1, there was an overall decrease in similarity. Particularly, a trend of reduced similarity between languages was observed.

The results of the similarity analysis in question 3 are depicted in Figure 3. Compared to question 2, there was a decrease in similarity across different languages. A notable high similarity in the Japanese version of GPT-4 was observed (cosine similarity: 0.89-0.96).

Items mentioned

For question 1, 24 topics were identified (as shown in Table 3). Topics such as 'magnetic fields & radio waves', 'imaging sites (tissues)', and 'no radiation exposure' were mentioned in almost all generated texts. A frequent topic in the English model was 'image creation algorithms'. 'Claustrophobia' was only observed in the English outputs. Although limited, topics exclusive to the Japanese outputs included 'operated by a specialist' and 'allergy'. Mentions of 'implant' were found in both Japanese and English outputs, particularly in the more recent models.

For question 2, 34 topics were identified (as shown in Table 4). Topics like 'clothing', 'implant', 'pregnant woman', and 'claustrophobia' were mentioned in almost all texts. Descriptions characteristic to the English texts included 'renal function' and 'medication intake'. Specific to the English outputs, albeit limited, were topics such as 'intraorbital metal', 'child's examination', and 'monitoring during the examination'. Unique to the Japanese outputs were 'electronic devices', 'examination time', 'magnetic field sickness', and three other topics. 'Tattoo' was mentioned in both languages, especially in the newer models.

For question 3, 33 topics were identified (as shown in Table 5). Descriptions of 'clothing' appeared in all texts, while 'food and drink' were mentioned in all but one. English-exclusive topics encompassed 'arrival time', 'means of returning home', 'communication with doctors and radiographers', among ten others. Conversely, Japanese-exclusive topics included 'magnetic fields & RF waves', 'electronic devices', 'examination appointment', and two others.