Pinned toot

Hi, I am swan314. I'm from Japan. I am a natural language processing engineer. I'm interested in applied mathematics, especially, mathematical modeling and that's why I choose this instance. I am interested in seeing a wide range of information on mathematics. Maybe I will post less frequently, but I think I will post something about computer science. Nice to meet you.

I bought a book about quantum mechanics a few years ago, but was so busy with work that didn't have the motivation to read it. Now that have automated most of my work and have some free time, so will read it.
bookstore.ams.org/stml-47

In a memory game, if a player's memory capacity is M and he declares that he will only remember m<M, the possibility of overflow can be reduced, but what value m should be is a relationship between precision and recall.

I recently learned about the concept of naturaliness. We need a theory when we can no longer obtain new data from existing experiments, but we need criteria for selecting a theory. The idea that human existence is a constraint on physical structure seems to be known as the anthropic principle.

en.m.wikipedia.org/wiki/Natura

This paper is interesting for me to learn a case study of mathematical modeling. Ideally I want data of its experiment to check the models.

researchgate.net/publication/2

The book is Japanese translation version of "Mathematics: Frontiers and Perspectives" from AMS. It is written at Roger Penrose section.

In a book, I read that classical mechanics, special relativity, general relativity, quantum mechanics, and quantum field theory are characterized by which constants in {G, h, c^-1} are zero or not. I'm interested in physics recently, so these generalization is surprising for me to learn it.

This central affiliation bias results in the DK effect, which is no longer the explanation for psychological effects such as "overconfident people are less capable".

There are two variables, self evaluation(x) and real evaluation(y). DK effect assumes that the difference x-y negatively correlate y itself. If x and y are uniform distribution, it is artificially correlate negatively. And then, adding an assumption that "x is distributed around max(x)/2", the correlation is stronger.

The rule base that just stores sentence pairs that match perfectly against their company-made test data don't have generalization performance.

One day, a company claimed that they had achieved BLEU 88. Their method simply used a translation memory to perfectly match the sentences in the test data, so it is not an accurate representation of translation performance.

BLEU is used as a compromise in the evaluation of machine translation systems, but in many cases it is difficult to compare without matching the test data, so just saying "high BLEU" does not tell us anything.

The Dunning-Kruger effect can be easily and artificially simulated using simple uniform random number, so I don't think it is a psychological phenomenon.

Originally it was a system presented at an event called NLP2021, but the source code is not yet available, so I am trying to reproduce it personally.

anlp.jp/proceedings/annual_mee

Recently, I've tried a kind of interactive expert system by using ccg2lambda. If we input some premises and a goal, the system produce some subgoals, and I want to transform that subgoals to readable sentences for human, and then have tried to prove it interactively.

github.com/mynlp/ccg2lambda

Hi, I am swan314. I'm from Japan. I am a natural language processing engineer. I'm interested in applied mathematics, especially, mathematical modeling and that's why I choose this instance. I am interested in seeing a wide range of information on mathematics. Maybe I will post less frequently, but I think I will post something about computer science. Nice to meet you.

The social network of the future: No ads, no corporate surveillance, ethical design, and decentralization! Own your data with Mastodon!