Forced to use it to organize tasks
I was told that the listening chat system is not suitable for organizing tasks when there are multiple tasks because it digs into a single task, so I tried a crazy experiment: if that's the case, why not open a chat for each task?
→Conclusion: Objective achieved
consideration
Open a separate chat for each task and you can drill down individually.
That's not the way it was done in the Mattermost era.
Crystallization of the previous task to a question of value will have a positive impact on the later task.
The language in the #2 chat about the importance of the rate of keyword increase gave a strong raison d'être to task #5. Before I started, I thought maybe number 4 would be next.
The more detailed the individual tasks, the easier it was for humans to prioritize, and the zero chat was in phase mode and not asking questions.
It would have been hard to do a recap on each of the five tasks if we had had questions.
nisbot.iconWhat would you like to see happen in this conversation?
nishio.iconI'm told that the current system is not suitable for selecting the next task from multiple tasks, so I'd like to try to see what happens if I drill down to the individual tasks then.
nisbot.iconWhat type of task is the task?
nisbot.iconWhat would you like to see happen in this conversation?
nishio.iconI want to think about the design of the end
nisbot.iconWhat type of design is that?
nishio.iconHow to end this dialogue? First idea, "What just happened?" with the appearance of "I see" and verbalization of the content of the awareness, then "Is this a good place to end?" and if "yes", it ends; second idea, "What happened?" or "Is this a good place to end?" with machine learning.
nisbot.iconWhat type of dialogue is that?
nishio.iconThe human side gets tired, and the quality of the questions usually declines and the end of the session becomes unpleasant. It's exhausting on the human side, and the quality of the questions generally declines as you run out of good keywords, leading to an unpleasant ending.
nisbot.iconWhat type of side is that side?
nishio.iconNGKW side
nisbot.iconWhat type of question is that gradual question?
nishio.iconThe longer you keep doing this, the worse it will get unless there is a sufficient supply of good keywords from humans.
nisbot.iconWhat type of task is the task?
nishio.iconListen first.
nisbot.iconYes, sir.
nishio.iconAnother thing that came to mind during the first chat, looking at the extracted keywords for each step in the early stages, there may be a difference between cases that develop well and those that do not. We could use that to detect situations that don't seem to work and give advice on how to make them better.
nisbot.iconthe opening(s) (e.g. in a game of go or chess)
(PS: I saw this keyword extraction and thought, "Well, what I want to do is 'design the early stages'.")
nisbot.iconWhat kind of person is that person?
nishio.iconAside from that, do we do it with machine learning or rule-based?
nisbot.iconWhat type of machine learning is that?
nishio.iconBinary classification from a variety of features. I'm not sure what kind of features we need, but maybe the last input, the question, the keyword exhaustion, the length of the conversation, etc.
nisbot.iconWhat type of question is that?
nishio.iconI think the probability of prompting awareness depends on the question.
nisbot.iconIs there anything else about that question?
nishio.icon(initiate a second chat).
nishio.iconI remembered, depending on the question, I wanted to push the answer choice. Specifically, "Is this a good place to end?" I want to give a "yes" and then press it to finish.
nisbot.iconWhat is it that you value?
nishio.iconIt feels good to end a stonkingly clean session. It's better than cutting off the slog that goes on and on.
nisbot.iconWhat type of feelings are those feelings?
nishio.iconIt looks like it was done properly. Which reminds me, I want to delete the input field on the chat system side of the decision.
nisbot.iconWhat type of chat system side is that chat system side?
nishio.iconI mean at the discretion of the program.
nisbot.iconWhat type of decision is that decision?
nishio.iconIf you decide that the dialogue is over here, delete the input field and end it.
nisbot.iconIs there anything else about that dialogue?
nishio.icon(Leave for a moment)
nisbot.iconWhat would you like to see happen in this conversation?
nishio.iconI would like to think about the design of the early stages of the project.
nisbot.iconWhat type of early stage is that early stage?
nishio.iconIn the early stages of this dialogue, some people don't get on track well and crash, and I want to help them get on track well.
nisbot.iconIs there anything else you can tell us about that early stage?
nishio.iconThe chat system asks in-depth questions in response to keywords, but in extreme cases, for example, there are people who enter sentences with no keywords at all. In such cases, of course, there are no keywords, so it is impossible to ask a proper question.
nisbot.icon
In the early stages of this dialogue, some people don't get on track well and crash, I want to help them get on track well.
Where is this early stage?
nishio.iconEspecially the first time you use it, I recommend looking at the log, but you start it up without looking at it and put "what do you want to happen?" and then you start up without looking at the log and put "Nothing" in the "What do you want to happen? Well, aside from such extreme cases, even people who have read the logs to some extent sometimes use NGKW too much and run out of keywords.
nisbot.icon
In the early stages of this dialogue, some people don't get on track well and crash, I want to help them get on track well.
How do you know these early stages?
nishio.icon(Why such a focus on the early stages?)
nishio.iconAs for how to know that the early stages are not going well, perhaps you can compare the logs that are going well with those that are not and see if there is a significant difference in the growth of the number of keywords extracted, and then you can determine that.
nisbot.iconWhere is the beginning of this process?
nishio.icon🙁
nisbot.icon
How do you know that the early stages are not going well? If you compare the logs that are going well with the logs that are not going well, there is probably a big difference in the growth of the number of keywords extracted, and that's how you can tell.
What does this early stage look like?
nishio.iconIt's already there, but it's like launching a rocket, you have to get a certain speed or you'll crash.
nisbot.iconWhat type of rocket is that?
nishio.iconRocket to reach the moon
nisbot.iconIs there anything else you can tell us about that rocket?
nishio.iconI can detect lack of speed, but I don't know what to do when I detect it.
nisbot.iconWhat type of speed is that?
nishio.iconSpeed of keyword growth
nisbot.iconIs there anything else you can tell us about its speed?
nishio.iconOn the other hand, some people reduce it more and more.
nisbot.iconWhat kind of person is he/she?
nishio.iconI'm too rigid in my evaluation of keyword extraction, and if something is even a little bit wrong, I'll NGKW more and more. I'm not in the business of creating training data for a keyword extraction engine, but using this chat system is supposed to help me reach awareness (month), but I'm acting counterproductive.
nisbot.iconWhat type of chat system is that?
nishio.iconTake out what humans have not yet verbalized.
nisbot.iconWhat is it that you value?
nishio.iconBy using this chat system, I can verbalize what I have not been able to verbalize yet, notice my blind spots, etc., and I have a high probability of this happening, so I want other users to be able to use it well and experience the awareness.
nisbot.iconIs there anything else you can tell us about that chat system?
nishio.iconHmmm, does it take practice to get used to using it?
nisbot.icon
By using this chat system, I can verbalize things that I have not been able to verbalize yet, or notice blind spots, or something like that, and I have a high probability of this happening, so I want other users to be able to use it well and experience the awareness.
What is this chat system like?
nishio.iconI am a counselor who can chat with him 24/7, but his Japanese is a bit awkward. If I had to point out every single mistake in her Japanese, the conversation would never get anywhere.
nisbot.icon
By using this chat system, I can verbalize things that I have not been able to verbalize yet, or notice blind spots, or something like that, and I have a high probability of this happening, so I want other users to be able to use it well and experience the awareness.
Where does this chat system come from?
nishio.icon(Once here, go to the next task)
nisbot.iconWhat would you like to see happen in this conversation?
nishio.iconI would like to think about a feature that plays bad keywords.
nisbot.iconWhat type of keywords are those?
nishio.iconThe one this system uses for questions.
nisbot.iconWhat kind is that one?
nishio.iconNGKW
nisbot.iconWhat type of system is that?
nishio.iconDig deeper by asking questions for keywords.
nisbot.iconIs there anything else you can tell us about those keywords?
nishio.iconBut as we talked about in another chat, people often get too few keywords... Maybe we should prioritize increasing keywords over correcting the direction of reducing them?
nisbot.iconWhat do those keywords seem to be?
nishio.iconHooks for catching fish
nisbot.iconWhat type of fish hook is that?
nishio.iconThe more you have, the more likely you are to catch fish, but more bad hooks don't mean anything.
nisbot.icon
But as we talked about in another chat, often people get too few keywords... maybe the priority should be to increase keywords rather than to modify them in the direction of reducing them?
Where is this keyword?
nishio.iconI've found a few keywords that are obviously no good, about 110 when I search the logs with NGKW. Some of them are not no-no's from my point of view, but some of them are no-no's, like "つ"
nisbot.iconWhere are those keywords located?
nishio.iconThey are scattered in the data file and can be collected by writing a tally script.
nisbot.icon
I've found a few keywords that are obviously not good, I've searched the logs for NGKW and found about 110. Some of them aren't no-no's from my point of view, but some of them are no-no's, like "つ"
How do you know these keywords?
nishio.iconThings that are subject to NGKW commands. The rest may be candidates even if they are not explicitly commanded, but their scores are not increasing.
nisbot.iconWhat type of tally script is that tally script?
nishio.iconI'll go through all the logs, collect the keywords that are the subject of the command, and output them.
nisbot.iconWhat type of log is that log?
nishio.iconStored logs for this system
nisbot.iconIs there anything else you can tell us about that system?
nishio.iconVarious failed keywords can be found by various people using them.
nisbot.iconWhat type of failure keywords are those failure keywords?
nishio.iconA failure to extract keywords, which was cut out in an odd way.
nisbot.iconWhat type of keyword extraction is that?
nishio.iconIt is rather good for some long sentences, but for short words, it tries to force extraction from them and sometimes extracts strange things. The goodness of the keywords found is normalized to a maximum of 100 points, so if the input is such that only one garbage keyword is found, it will be a garbage keyword with 100 points.
nisbot.iconWhat is it that you value?
nishio.iconIt is important to be able to extract good keywords
nisbot.iconWhere do these keywords come from?
nishio.iconShould we just lower the score when the user's input is short or something? Not necessarily, I guess...sometimes you have to put in a perfect keyword in one word. I still have to machine learn how good the keywords are.
nisbot.iconWhat should those keywords happen to be?
nishio.iconBad keywords, even if they are extracted, should score low.
nisbot.iconWhat needs to happen to those keywords for that to happen?
nishio.iconWe just need to determine that the keyword is bad.
nisbot.iconCan that keyword do it?
nishio.iconYou can do that by putting together that kind of machine learning.
nisbot.iconWhat type of machine learning is that?
nishio.iconWe can use a binary classification to learn whether NG or not, and use the probability values as coefficients.
nisbot.iconWhat type of coefficient is that?
nishio.icon1 for good keywords, 0 for bad keywords.
nisbot.iconWhat type of good keywords are those good keywords?
nishio.iconI guess I could choose one that has a longer user response to the keyword.
nisbot.iconWhat type of user is this user?
nishio.iconPeople who generate input
nisbot.iconWhat type of input is that?
nishio.iconAnswers to Questions
nisbot.iconWhat type of question is that?
nishio.icon🙁
nisbot.icon
1 for good keywords, 0 for bad keywords
Anything else about this good keyword?
nishio.icon(Move on to the next task)
nisbot.iconWhat would you like to see happen in this conversation?
nishio.icon I would like to think about [Naturalness determination of questions
nisbot.iconWhat type of judgment is that judgment?
nishio.iconDetermine if the combination of keywords and questions is natural based on training data
nisbot.iconWhat type of question is that?
nishio.iconSome questions given
nisbot.iconIs there anything else about that question?
nishio.iconI think there's a funny kind of word to put in there, obviously, like "what happens when ~?"
nisbot.iconWhat type of word is the word?
nishio.iconI wondered if I could make a decision from the data without considering that, but it doesn't seem to be working out so well, so I guess I should think about what I would do if I were rule-based and devise a feature set.
nisbot.icon
Some questions given
Where is this question?
nishio.iconPreset in the system.
nisbot.iconWhat type of preset is that preset?
nishio.iconAdapted from a technique called Clean Language.
nisbot.iconWhat type of clean language is that clean language?
nishio.icon(Go to next task)
nisbot.iconWhat would you like to see happen in this conversation?
nishio.iconI want to think about movement extraction.
nisbot.iconWhat type of extraction is that?
nishio.iconWhen we say keyword extraction, the world often extracts noun phrases. It's the same with keywords in papers and so on. But when I think about using them in this chat, verbs are often accompanied by more concrete images than nouns, and are too good to throw away.
nisbot.iconWhat type of noun is it?
nishio.iconIf you extract the noun, it would look like this. With verbs, it would be "What kind of 'use' is that 'use'?" And so on. It's a question, isn't it?
nisbot.iconIs there anything else about that noun?
nishio.iconNouns often tend to be abstract concepts, and when asked about abstract concepts, they often tend to be even more abstract, which is not very good!
nisbot.iconHow do you know that noun?
nishio.iconMorphological analysis will tell you.
nisbot.icon
Nouns often tend to be abstract concepts, and when asked about abstract concepts, they often tend to be even more abstract, which is not very good!
Where is this noun located?
nishio.iconI'll take out the verbs in the user input and add them to the keywords for now, but some of them, like "say" and "see" are not very good, so I need to filter them out.
nisbot.iconWhat type of keywords are those?
nishio.iconAlso, where do you have the information that it's better to put it in hook brackets.
nisbot.iconWhat type of hook brackets are those hook brackets?
nishio.iconIt's better to group not only verbs but also words that contain particles, so you can either judge by the display part.
nisbot.iconWhat type of verb is that verb?
nishio.iconIt's not about "being" or anything, but about the user's active behavior that is important.
nisbot.iconWhat type of user is this user?
nishio.iconWe've delved into each of the five tasks so far in individual chats
nisbot.icon(Internet) chat
nishio.iconThe first is the design of the end, which is intended to have the value of "it feels good to end nicely." Specifically, "allow questions to produce answer choices," "ask the question, "Is this a good place to end, rule-based," "if yes, delete the input field," and "yes, If yes, then delete the input field, and if no, then use the information for machine learning in the future.
nisbot.iconoptions
nishio.iconSecond is the design of the early stage, which is to "have many users experience that by using this chat system, they can get language and awareness of their blind spots," and to solve the problem of not getting off track in the early stage for this purpose.
nisbot.iconthe opening(s) (e.g. in a game of go or chess)
nishio.iconHe wrote that he can detect the slow rate of keyword increase, but does not know what to do about it. Is it better to visualize it? A rocket launch metaphor, a percentage indication of how far the value to be achieved in each phase has been achieved?
nisbot.iconmetaphor
nishio.iconThe third is the ability to play bad keywords. The value is to be able to extract good keywords. This is machine learning of good and bad keywords.
nisbot.iconmachine learning
nishio.iconThe fourth is the naturalness determination of the question, which has not yet reached the value question. Specifically, I am thinking that I should devise a feature to do this in a rule-based manner.
nisbot.iconvalue
nishio.iconThe fifth is motion extraction, which is also not a value question. It is easy enough to just extract verbs, but not everything, so a filter must be applied.
nisbot.iconvalue
nishio.iconAnd well, after thinking this far, I decided that the first of the five should be movement extraction. Because it will have a positive impact on the way the early stages of the project will proceed by increasing the number of keywords extracted. Next is a mechanism to play the keywords that are not good. This improves the quality of keywords and reduces the likelihood of stumbling in the early stages.
nisbot.iconthe opening(s) (e.g. in a game of go or chess)
nishio.iconThe early design does not have a clear language of action. It would be a good idea to try to verbalize it again after the first two are done. And if you come up with one, implement it, and if not, proceed to the end design.
nisbot.iconthe opening(s) (e.g. in a game of go or chess)
nishio.iconImprovements to the naturalness determination of the questions are pending, but I'm okay with putting it off until later because the training data is growing just by leaving it alone.
nisbot.iconquestion
nishio.iconthe end
nisbot.iconWhat would you like to see happen in this conversation?
nishio.iconI wanted the tasks to be organized, but I got what I needed to do first and next, so the objective was achieved.
nisbot.iconIs there anything else about that task?
---
This page is auto-translated from /nishio/タスク整理に無理やり使う. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I'm very happy to spread my thought to non-Japanese readers.