Coordination of state transitions
Originally, the system was designed to "transition to the next state once the conversation has developed to a certain degree.
The "when it develops to some extent" was judged by "when the number of extracted keywords exceeds a certain level.
As a result, state transitions began to occur before they were sufficiently developed
I had foreseen this, but I didn't care and proceeded.
The case where keywords containing verbs are now extracted, if they are simply connected, the number of keywords extracted will naturally increase, and then the places that require a certain number of keywords to be extracted will all be triggered earlier than now.
I decided not to worry about it and integrate it.
I observed the speed at which keywords appeared, and since many users were running out of keywords in the first place, I figured I could increase them without worrying about it.
However, a negative effect was observed: "Questions that assume that they have developed before they are fully developed are asked."
Phase 3
Keywords with scores above 100, average 4.93, median 5
Keywords with scores over 200, average 1.06, median 1
Phase 2
Keywords with scores above 100, average 1.48, median 1
Keywords with scores over 200, average 0.69, median 1
Maximum score keywords, average 167
Therefore, we changed the index to use the number of high-scoring keywords and the keyword scores themselves, rather than the total number of keywords.
---
This page is auto-translated from /nishio/状態遷移の調整. 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.