Detecting Rhetorical Operations by Exploiting Their Inherent Properties

3 min readJul 31, 2021


Decentralized Socrates (Part 13)

Fudō Myōō, early 13th century, Kaikei

Previous post

The “ negative to positive” type of rhetorical manipulation mentioned in the previous entry can be detected automatically in principle for the following reasons.

In most position talk, the topic intended for a positive impression is not changed very often.

Therefore, it is possible to detect, without fact-checking, that multiple articles on a website consistently form a “negative to positive” shape regarding the news of one specific position. As long as the author of that documents does not change his or her position, we can expect stable detection results. Position talk is hard to change frequently a position he or she wants to impose. Propaganda is only effective when it persistently repeats the same stance, as Hitler said. This “positional adherence” is the inherent property of positional talk, making the following inference very tractable.

It is possible to infer “position × rhetoric” from multiple opinions and actions of a certain site or person.

Current achievements of AI technology permit inferences on uncomplicated interpretation and style of a text close to average humans’. What AI is not good at is inferences using real-world facts. For example, it is absurd to say “I went on an overseas tour trip” after the sentence “humanity has perished” (because if humanity had perished, there would be no one to carry out the tour trip), although detecting this type of contradiction is still challenging to current AI.

However, with rhetoric detection, such alignment between the physical world and the text is irrelevant. Rhetoric detection combines style detection in writing and inference of a position some speaker is pushing. The task of judging whether the evaluation of a particular item will end up being positive or negative from reviews is “sentiment estimation.” AI can do it quite well, even if it still ignores the reality outside texts. This technique is already familiar to us from the movie review classification. Style detection is another kind of task which AI is capable of treats. So, this rhetoric detection can be implemented by current AI technology.

Once position estimation and style detection are done, we can combine the two to infer that “Speaker A is a frequent user of style (rhetoric) type n in discussing topic X.” This is a rhetoric detection, and in the example above, topic X = the future of stock prices, and “type n” = “ negative to positive” type rhetoric. By calculating this for all topics of each speaker, we can see the rhetorical tendencies of that speaker. As I have pointed out repeatedly, this result has independent of whether the speaker is telling the truth or not.

Even if the authors of the position talk were aware of such a rhetorical detection system, there is no simple way around it.

Suppose a news producer uses a “ negative to a positive “ type of rhetorical manipulation to create an atmosphere that the stock price will continue to rise. He found out that the rhetoric detector caught his manipulation. However, assume he tries to fool the detector by simply using a “ positive to negative” order. In that case, the negative opinion will come last, and the impression it gives will be the opposite of the positive one he wants. Rhetorical manipulation is meaningless in this case.

Of course, it is possible to make an article more elaborate instead of “ negative to positive” or “positive to negative.” Yet, this would probably require more effort and be less effective than simple impression manipulation such as “ negative to positive.” The more complicated the rhetoric, the lower the chance of being recognized, the closer it is to noise.

Also, as mentioned earlier, the position talk is only compelling when it adheres to a fixed position. So it is difficult to avoid the situation by changing the position.

With such a system, it is possible to automatically detect whether a given site or person is prone to position talk, regardless of fact-checking.

Of course, we can mix this detection with other techniques, such as fact-checking or influence analysis using follow relationships on social networking sites.

Yet, as we can instantly see, this kind of verification requires enormous computational resources. Who can provide them? Or who should give it?

Our discussion will return to the DAO and decentralization to deal with this issue in the next post.