Through AI-based analysis of the minutes of the Japanese Diet, this essay visualizes groupthink within political parties, the political leanings of newspaper editorials, and ideological bias across academic fields at major universities.
It is an important piece that counters academic preconceptions and political assumptions with objective data and scientific methods grounded in predictive power.

2019-04-13
Striking back at academic preconceptions with artificial intelligence…
Many of the results obtained in this survey accord with my own empirical observations.

What follows is a continuation of the previous chapter.

Striking back at academic preconceptions with artificial intelligence.

Many of the results obtained in this survey accord with my own empirical observations.
As an undergraduate, the author studied biochemistry in the Faculty of Science, and in graduate school worked on research related to artificial intelligence in a Graduate School of Engineering.
From that time, he had a strong interest in the theory of scholarship and the philosophy of science.
Accordingly, he read many books on the philosophy of science and also took part in study groups in related laboratories, and he still vividly remembers being astonished by a certain remark he encountered there.

“This must be correct because Professor ○○ says so.”

It was the moment when I felt disillusioned with authoritarianism.
Such a remark would be unthinkable in a science or engineering laboratory.
But that incident was only the prologue.

After taking up a teaching position at a university, I came to realize that there exists a kind of “scholarship” that operates on values entirely different from the education we had received.
In the natural sciences, one is rigorously trained to interpret experimental results while excluding subjectivity as much as possible.
Naturally, it is out of the question to tamper with experimental results for political reasons or to distort their interpretation.
That is precisely why scholarship must remain independent from politics.
And yet, within the humanities and other non-scientific disciplines, I noticed that there was a proliferation of “scholarship” that, from the standpoint of someone in the sciences, made it impossible to tell whether one was engaging in scholarship or in politics.

Though all of them call themselves scholarship, activities based on entirely different norms are being carried out.
To resolve this confusion, one must begin by defining scholarship itself.
That is why I wrote What Is Scholarship? in 2005.
In that book, I argued that disciplines bearing the name of “science,” such as the humanities, social sciences, and natural sciences, must satisfy the requirement of being “systematic knowledge possessing the power to predict.”
A discipline that conveniently manipulates experimental results naturally cannot possess predictive power.

In recent years, research ethics have come under stricter scrutiny, prompted by the misconduct scandal involving Ms. Obokata.
The other day, however, at a symposium on promoting education in research ethics, I again encountered a shocking scene.
A man from a medical background who had served for many years as president of a national university made the following remark from the stage as he discussed the background to research misconduct.

“When people fall into groupthink, misconduct occurs.
It is just like the Liberal Democratic Party.”

When I think that a person who cannot perceive the inappropriateness of making such a political remark in a setting like this stands at the center of research ethics education, I cannot help feeling utterly disheartened.
Of course, it would be another matter if he were saying this on the basis of objective evidence that the Liberal Democratic Party is more prone to groupthink than other parties.
But in my own research, at least, I have obtained results showing the opposite.

For about ten years, our research group has been conducting document analysis using information engineering, in order to bring scientific methods into the humanities and social sciences.
Generally, this field is called big data, data mining, or text mining.
Much of this research assumes business applications, but I have worked chiefly on the analysis of political documents, such as the minutes of the National Diet.
As one part of that work, we created an artificial intelligence system that identifies from which party a statement made in the Diet came.
This connects to the chapter I published on 2018-05-05 under the title, The highest accuracy was for the Communist Party at 93%, while the lowest was for the Democratic Party at 65%.
What follows is a continuation of the previous chapter.

When we had a classification system trained by machine learning on Diet minutes from 1999 to 2008 determine, from five choices — the Liberal Democratic Party, Komeito, the Democratic Party of Japan, the Social Democratic Party, and the Communist Party — to which party each legislator belonged, the highest accuracy was for the Communist Party at 93%, while the lowest was for the Democratic Party at 65%.

In other words, the Communist Party’s members held the most uniform opinions, making it easy for artificial intelligence to identify them as “Communist Party,” whereas the Democratic Party’s members had the most diverse opinions and shared the fewest common characteristics, so the accuracy rate declined.
The fact that the opinions of legislators belonging to the Communist Party were so uniform can also be taken to mean that the party had the strongest tendency toward groupthink.
Conversely, the diversity of opinion among Democratic Party legislators can also be taken to mean that it was a party prone to splitting, and that prediction was splendidly borne out by the later split of the Democratic Party’s successor formations.

The Liberal Democratic Party came next to the Democratic Party, with a relatively low accuracy rate of 70%.
Based on these objective data, the criticism that the Liberal Democratic Party is characterized by groupthink is not correct.

Based on these research results, I then considered whether the same system might be used to measure the political bias of discourse in society.
That is because it can numerically indicate which party’s parliamentary statements various forms of discourse most closely resemble.
The first attempt was to apply it to newspaper editorials.
When we analyzed the five papers — Asahi Shimbun, Mainichi Shimbun, Nihon Keizai Shimbun, Yomiuri Shimbun, and Sankei Shimbun — the result was generally in line with common perception: Asahi Shimbun showed the highest similarity to the three opposition parties of the time, namely the Democratic Party, the Social Democratic Party, and the Communist Party, while Sankei Shimbun showed the lowest.
However, all five newspapers showed greater similarity to the opposition parties than to the ruling parties.
This is likely because newspaper editorials are basically often written from a critical standpoint toward something.

Comparing the bias of elite universities.

The next thing we attempted was an analysis of texts published on university websites.
We collected and analyzed texts such as messages from university faculty members and statements of educational philosophy from graduate schools and departments, sorting them by academic field.
The universities and texts collected and analyzed were as follows: the University of Tokyo, with 44 texts in the humanities and social sciences, 51 in science and engineering, and 38 in life sciences; Kyoto University, with 72, 37, and 40 respectively; the University of Tsukuba, with 51, 32, and 32; Waseda University, with 55 in the humanities and social sciences and 29 in science and engineering; and Keio University, with 47 and 32 respectively.
Waseda and Keio were excluded for life sciences because there were too few relevant organizations to gather a sufficient number of texts.

We divided the collected texts into three fields — humanities and social sciences, science and engineering, and life sciences — and determined which party’s statements each university’s texts most closely resembled.
The results are shown in Figures 1 to 3 on the left.
All were judged to lean toward the ruling parties because the collected texts were intended to promote the universities, and therefore contained more positive expressions than critical or negative ones.
Even so, it became clear that all academic fields at Kyoto University, as well as the humanities at Keio University, showed greater similarity to opposition-party discourse, and that by field, science and engineering leaned most toward the ruling parties, while life sciences leaned most toward the opposition parties.

However, whether a text resembles ruling-party statements or opposition-party statements is also influenced by the balance between positive and negative expressions, so it is insufficient in many respects as a criterion for judging whether a way of thinking is conservative or progressive.
For that reason, research into the ideological bias of universities had temporarily stalled.
But the previously mentioned “Scholars’ Association Opposing the Security-Related Laws” provided excellent material for overcoming that difficulty.

Political bias in the natural sciences.

As already stated, all disciplines bearing the name of science — the humanities, social sciences, and natural sciences alike — aim at constructing systematic knowledge with predictive power.
To achieve that aim, scholarship must be politically neutral.
For example, scholars with a pro-China political stance claimed that once China became a member of the international community through WTO accession and the like, it would begin to respect intellectual property rights and would no longer display territorial ambitions.
But that prediction was completely wrong.
If one had objectively analyzed the history and political institutions of the Chinese Communist Party, one could have fully anticipated such an outcome.

Regrettably, political bias exists even in the natural sciences.
Among scholars promoting nuclear power, there were some who said that nuclear power generation was absolutely safe, but scientifically speaking there is no such thing as absolute safety.
In reality, the Fukushima Daiichi Nuclear Power Plant suffered a major accident.
On the other hand, not a few researchers advertise natural energy as environmentally friendly in order to obtain research budgets, yet because of its low energy density, making natural energy a major power source entails environmental destruction through the development of vast areas.
Recently, the destruction of nature caused by mega-solar projects has begun to attract criticism, but these developments could easily have been predicted if viewed without the ideological spectacles of any particular politics or ideology.
To be continued.