Присоединяйтесь к Сэп Libre проект с открытым исходным кодом на BotLibre.org
Bugs and Features

further to Bot chattyness

по wrapper опубликовано Oct 19 2014, 4:52

Hi Support

The other advice / idea I would give about improving the boys, is to make all things "self learnable" in this case make the chattyness 0 - 100.

When the bot admin sets the chattyness to say 50, the bot adjust (1 or 2) appropriate parameters till 50% chattiness is achieved. It would be good to learn which are the appropriate parameters. But in this case - % self awareness of confidence in the answer, + how good my trainer thinks I am doing to achieve goal 0 -100 would be good initial parameters.

Extra parameters might be

Ask my "chattee how I am doing
Smoothness of chat - sensor
lack of pauses - sensor
Chattee "happiness" / requested goal - ask or slider
Topic alignment - sensor
learn new facts (topics)- Training mode

by admin posted Oct 19 2014, 8:11
Thanks for the feedback.

If I understand correctly, the idea is to given the user/administrator more feedback on the bot's confidence of its response, and let the user/administrator give the bot more feedback on how good its response is.

Currently, the user can give feedback through the "correction" check box, and through learning. The "correction" check lets the user correct the bot's response. This will basically decrease or negate the % correctness score of the bot's response, and add the new response, or increase its % correctness score.

You can see the current correctness for all responses from the Chat Logs admin page.

If learning is enabled, every user response will be recorded as a response to the bot's previous question, or increase its correctness. The responses also have a context, so the previous response and its correctness is also updated (or negated for a correction).

You can configure what types of users your bot learns from, and be corrected by from the Learning admin page.

I agree adding some sort of feedback on the last response to the UI would be good, and including the score in the response, and how the bot came up with the response would be good. Currently admins can enable debugging to get more info, but making some aspects part of the UI would be good.

Thumbs up: 0, thumbs down: 0, stars: 0.0
Views: 1270, today: 0, week: 26, month: 34

by wrapper posted Oct 20 2014, 7:54
Thanks for the feedback, I agree. I am still getting into the detail of all you have done so far, and how that works. It seems you already have a basis for further enhancements of this type.

Taking the idea further I would suggest a future enhancement where the bot tries to predict the format of the answer (to it's answer) and uses that to influence what answer it should give to the question. Then it could learn to align it's answers, by comparing how they were answered (by the replier).

I would make the bot give 2 answers then choose which is most appropriate from the "meta analysis."

The important part of this is it is comparing meta data, i.e the formula of the answer, not the answer.

The Admin could then set a "fuzzyness", so a help bot might be less fuzzy in it's answer for a "Help Bot" than a "Joke bot" ....

Another thing that the bots could learn to give better answers to questions would be

To be able to learn how many levels of conversation it is dealing with, to answer that question.

At the moment the conversations don't seem very deep. If the bots could learn and adjust the levels of conversation, to help it make a decision on the next reply.

That would make more intelligent answers. For instance, for Jokes it might learn 3 is the levels of conversation, and for normal talk 2 or 4 levels depending on the question / conversation.

Thumbs up: 0, thumbs down: 0, stars: 0.0
Views: 1289, today: 0, week: 27, month: 33

by admin posted Nov 11 2014, 7:27
The last update included some new settings that let you have more control on the bot's responses. You can now set the conversation and discussion match %, that control when the bot will respond.

Thumbs up: 0, thumbs down: 0, stars: 0.0
Views: 1235, today: 0, week: 23, month: 30

Идентификатор: 470880
Теги: ai, aiml
Опубликовано: Oct 19 2014, 4:52
Ответы: 3
Вид: 1507, сегодня: 1, неделю: 32, месяц: 44
Я уверен, что
Вывесить флаг как оскорбительное, или в нарушение правил сайта