Assistance in dividing long sentences into stickies

---

Sentences written in a lighthearted manner

Ah, well, people who are not used to the process of making lots of stickies and doing the KJ method don't have a good idea of how granular the information should be when they make the stickies in the first place. That's where the software needs to help.

This is not the proper granularity.

https://gyazo.com/15e2094edc84a4e2694d774851748fda

Divide and conquer by experienced personnel

https://gyazo.com/0813639be5cc94dcbbac5d2d2ad9c220

https://gyazo.com/ca47482c3dfcbc2c4ce2cd0d0bd33018

This is a split and delete only pattern

With rewriting it looks like this:.

https://gyazo.com/024545873ed23e894dd2305b63569053

---

It worked for a while, but not great.

Instead of using the results after the engagement analysis, we thought it was necessary to devise a process of engagement analysis.

While trying to implement the co-reference analysis itself, I realized that there is no need for co-reference analysis in the first place.

v2: 4-level decomposition algorithm

consideration

For "words that appear at a distance but are connected in terms of their affinities."

Whether they are connected in terms of engagement or not, they can be separated as stickies because they appear at a distance!

Split by punctuation without length-dependent splitting, just split by punctuation.

It's good because I break it up with punctuation and if it's too long, I chop it up even more with conjunctive particles.

structure

Split by punctuation, parentheses, etc.

If the length is above the threshold, we'll split it with a conjunctive particle.

If the length exceeds the threshold, we'll split it with a cohesive particle.

If the length is above the threshold, we'll split it with a case particle.

I was thinking of doing machine learning based on the surrounding predispositions for where to divide, but when I looked at the predispositions with my eyes, I felt like this was the way to go.

v3: Algorithm to split recursively from the one with the highest split priority score

Observe and discuss v2 results

The problem that "come out" and "use it" are split by "te" and made into the original form to form "kuru" and "iru".

Same word, different division priority.

Change the score depending on the situation and split from the largest to the smallest.

Scores are now adjusted on a rule basis.

It will eventually fail.

Human beings cannot reconcile past decisions with consistency.

Moving to machine learning?

I'm collecting examples of cases that have not been divided well.

v4: If we were to move to machine learning

What is the best way to do it?

1: Binary classification of "to separate or not to separate" is performed for each word, and the words are divided from the one with the highest "score to separate" until the length constraint is satisfied.

Score calculation of the current score-based method becomes a form of machine learning.

2: CRF？LSTM？Transfomer？

If the system is too heavy, even if it is possible, I think it will be troublesome to operate.

Garbage Cleaning Algorithm

Not only do they divide, but they also remove unnecessary words.

E.g., "Oh, well..."

Currently, they're cutting it down to the dictionary base longest match.

Sticky note detail

https://gyazo.com/bcb23346a3e5a66bf74a2e3fab09a350

The default for images was not to auto-upgrade by default, but this has changed since M86. This auto-upgrade does not fall back to HTTP, so images served only via HTTP will not be viewable.

20 feels good.

API

---

2021-01-05

It's good up to the point where you inscribe it into a phrase, a little shorter, but you can remove the final particle, etc., from that phrase and make it a sticky note.

Another way to describe what we want to do is.

Sentences are a little too short for a phrase, so I'd like to combine as much as is acceptable.

I want to make a decision by looking at the length of the string of concatenated items.

There are so many words and phrases that don't need to be included in the deliverables that I want to ignore.

The Simple Way

If the area enclosed in parentheses is short enough, it is adopted.

This is useful, but not enough

----

---