Most of the content on my blog is in Japanese; some posts are written in English. In Google Chrome and Safari, you can translate a webpage into a language of your choice by right-clicking and choosing Translate. They do a surprisingly good job, and I appreciate it because none of us speaks everything. (If things don’t make sense, feel free to let me know. I’ll try to explain what I mean.)

Having said that, there’s a caveat. Neural machine translation like Google Translate will alter the sentiment of the source text. And the sentiment that comes out will likely be more negative than the original.

I’ve always wondered why machine-translated texts often have this slightly depressed (or depressing or dissatisfied) tone I don’t detect in the original sentences. Japanese and English are the only two languages in which I’m proficient enough to read between the lines, feeling the nuance. So I wasn’t sure if this issue was particular to Japanese-English translation.

A few weeks ago I was looking at the English translation of a Japanese newsletter, noticed this again, and finally googled why. There were many pieces of research for various language combinations. The data analysis parts of an academic paper I have no idea what they are talking about, but I do enjoy reading the discussion parts.

So here’s this AI layperson’s narrative on how to interpret Google Translate’s job.

1. We humans are naturally biased to be negative, because a lion in the bush.

2. Machine imitates humans, thus will translate a lot like us.

3. For accurate translation, turn up the positivity just a little.  Setting > Sentiment Adjustment. :)







happy new year 2022


1日1冊:丹野清志「写真集のつくり方」、Alison Bechdel「Fun Home」各つづき。