-
Project note on Listening Chat System (Keicho)pProjectIndex
-
Context: Scrapbox to manage project tasks.
Kanban in Scrapbox
- WIP
- HaveBetter
- đ€Allow footer to be specified when exporting to Scrapbox
- Dynalist
- The Phenomenon of Separate Conversations - Iâd like to discuss a bug that caused a mysterious log breakdown.
- Make questions that do not take other keywords score in active learning.
- Is local storage shared or not when âbring to home screenâ in standalone?
- 606da96aaff09e0000d54778
Memo - End Design - In the derivation, â[Allow questions to create command buttons. - Conversation log 2021-01-30 - I think the endgame of this one is probably that after the mode transitions to asking about relationships between symbols because it has a lot of keywords and considers them developed enough, it runs out of high scoring words, but itâs still in the asking about relationships mode. - Perhaps judging it from the distribution of keyword scores, - I propose an end. - Propose a new chat. - Return to search mode - I think it would be appropriate to do one of the following - question naturalness data set - The amount of features should be improved. - Related: Tuning questions that take two arguments. - Do not repeat symmetrical questions: done - Need to collect naturalness for questions that take two arguments as well. - Related: Good and bad feedback during conversation. - NG feedback is used as training data - â Not enough data and poor quality as training data. - I think Iâd be better off teaching through active learning.
- [[One-click import to Regroup]]
- [[IGNORE file to exclude inappropriate items from the analysis]]
- (e.g., logs that are not being used properly)
- [[Feedback button: Like]]
-
Do not tamper with the score to represent the subject of interest.
- This is causing confusion between âwhat is actually repeated in the userâs statementâ and âwhat the system determines to be noteworthyâ.
- If you were to create a task-organizing version of the listening chat system
-
In response to the input âTsu, Tsuyoshiâ.
- Not selected as a keyword because âstrongâ is an adjective.
- âWhat kind of âtoughâ is that âtoughâ?â is an ant question, so you might want to extract it.
- The âă€â can be a noun, so it is included in the keywords.
- The result is that you end up asking âoneâ digging question.
- Not selected as a keyword because âstrongâ is an adjective.
-
Play keywords that are not good.
- The CUI version had.
- Turn it off when you make the Mattermost version and keep it off.
- Use NGKW feedback to learn!
- âThere are quite a few cases where users are unaware of NGKW even though itâs a good keyword, and learning from users makes you a fool.
- I put it in the form of putting only the obvious no-noâs on the list.
-
What we want to feedback
- Keyword extraction is not the first step.
-
Youâre starting to buy into the subtlety of it.
-
Where is that w?
- NGKW command.
-
- NG keywords selected for input
- Example: The - midday sun
- Keyword extraction is fine, but not in the direction the user wants to drill down
- I hope you can say, âNo, not that way,â because this might not be something you can learn about in advance.
- It was realized in the form of the NGKW and UPKW commands.
- Not a good combination of keywords and questions chosen.
- NG feedback is used as training data
- Identical symbol relationship
-
If the answer is that the relationship between two symbols is identical, itâs not a good idea to ask questions about the relationship between symbols.
- Wouldnât it be nice if a button appeared that said âsameâ and when you pressed it, the two keywords were merged as a command?
- Iâll test it properly, because Iâm afraid that the disappearance of one of the keywords might cause some bug.
- Keyword extraction is not the first step.
-
State Transition Mechanism
- How should state transition diagrams be maintained?
- How to soften it up.
- Beliefs about the state
- partially observed Markov decision process
- Conversation Log: State Transitions
- Isnât it bad that the score you have for each word is one-dimensional to begin with? Thatâs what Iâm talking about. - Score = development, not attention
- Keyword and question selection is an internal volume caution.
- Described by a relatively small number of parameters
- I want to make the BETTER indicators as good as possible while adhering to the MUST constraints.
- parameter search
- The large number of variations in the choice of questions (all of which must appear at least once) is also a constraint.
- Order of appearance of the five basic questions
-
Include attenuation in cumulative score calculation?
- Wouldnât it be better to add attenuation? Iâve wondered about that.
- On the other hand, Iâm also inclined to think that if youâre responding well to the first âwhat do you hope will happen in this dialogueâ, you might want to remind yourself of the key words there from time to time.
- Often we observe cases where the first answer on this is messy.
- Itâs quite common for people to get more organized as they talk and it becomes clearer what they were looking for, so donât put too much emphasis on the first statementâŠ
-
- Hypothesis that questions that provoke long user input are good questions
- Questions that do not take keywords should take features from env.
- No means of recovery if keywords are split
- Judgment of Awareness
- Early design
- Someone is making utterances that do not contain keywords on the first move. Now Iâm asking questions without keywords anyway to avoid errors, but I might as well direct them to help since they donât seem to understand how to use it in the first place.
- There are cases where people donât understand âdigging for keywords,â and there are cases where people want to bully the program.
- Conversation termination from the programâs side Maybe we could.
- There are cases where people donât understand âdigging for keywords,â and there are cases where people want to bully the program.
- If you cut too much with NGKW before the keywords are developed, the âfountain has dried upâ.
- Maybe we are in a vicious cycle of using NGKW so much that the keywords are depleted, the program jumps to a keyword from the most recent statement and asks a question, which is off the main topic, so it is NGKWed again.
- How can we successfully lead them to the state of being crowded with lots of keywords?
- Show the percentage of the value to be achieved in each phase?
- Point out that there are too few keywords and the pace of NGKW is too fast?
- We could make it so that when there are few keywords, you have to wait a bit before you can press the button.
-
It would be nice if questions could specify some of the answers.
- Is it the same? No?â displays the âSameâ button
- When selected, it functions as a merge command for keywords.
- Is this a good place to end?â should be âyesâ or ânoâ.
- That selection can be used directly for training data.
- Is it the same? No?â displays the âSameâ button
-
- Read back and update
-
Add a RAKE-like element to keyword extraction
- Extract long sequences that appear in your own past statements/past logs/Scrapbox
This is software that helps you verbalize and retrieve âwhat is already in you but not yet verbalizedâ and is not intended to give you answers to your questions.
Keywords:.chatbot
-
The number of keywords taken is indefinite (0 or 1) only with respect to the phase.
- Maybe we should split it into two questions.
pKeicho-done Donât we need pairs of input rows and keyword extraction results? âdone
This page is auto-translated from /nishio/pKeicho using DeepL. 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. Iâm very happy to spread my thought to non-Japanese readers.