Look at a recent conversation with the http://neuralconvo.huggingface.co/ chatbot recently:
Left hand side is the robot, and right hand side is me.
What can be said about the chatbot?
1. Lacking of breath of topics for conversation. It is almost always saying "I don't know" or its equivalent all the time. The state of art for Eliza (https://en.wikipedia.org/wiki/ELIZA) is much more advanced than this.
2. Lacking of ability to pursue depth in any given topic of conversation.
3. Lacking of many "common sense" logical sentences, or "Tautology" in a sense. (saying the same thing in many different ways).
4. Inability to derive or insinuate indirect meanings of phrases used.
5. Inability to reinterpret the same sentences, word for word identical, but interepting it in a different way: eg, same word can be used as actual name, or general name, or adjectves vs noun switch, or meaning switch when the same word have double meaning etc.
6. Network of knowledge: its database of knowledge is almost non-existent. nouns which are linked via relationship to other nouns - bound by timing, location, action, relationship.
7. Pursuing step (3) above further, for every word or concept: there is always another concept that generalized it (widen it to more cases), or specialize it (narrow down to more special cases).
8. Language is just a means of communicating deeper knowledge. So if deeper knowledge is represented as a graph, then the language should treated as joining the different entities in the knowledge graph - with adjectives/nouns, or verbs etc.
9. Piecing a statement together will require: subjective matter + nouns + verbs (action) + objective matter + adjectives + adverbs (to describe state of action) + past/present timing. Emotional wordings like "Aha" or "Wow!!!" or "Hm...." may help to change the emotional tone of the conversation. To behave like a logical person, the changes should be logical and consistent, and not too abrupt.
9. Piecing a statement together will require: subjective matter + nouns + verbs (action) + objective matter + adjectives + adverbs (to describe state of action) + past/present timing. Emotional wordings like "Aha" or "Wow!!!" or "Hm...." may help to change the emotional tone of the conversation. To behave like a logical person, the changes should be logical and consistent, and not too abrupt.
10. Using analogies from "Affective Computing", the chatbot could have infer a lot more information about the statements I have written and then use it to generate more conversation.
11. Since the store of knowledgebase may be very large, question is how to selectively choose the relevant portion of the knowledgebase and usdde that to focus on generating the conversation contents (without revisiting all uninteresting nodes/sites/graphs etc). This "attention" masks can be repeatedly be use for converging on the right topics to discuss and talk.
12. The retrieval vs storing of knowledge should occurred separately from each other - so that it is possible to have them occurring CONCURRENTLY at the same time.: It should be possible to retrieve knowledge with a certain focus or attention mask, and then after that if there is a new piece of knowledge not stored before, it will be stored immediately.
References:
https://research.google.com/pubs/pub46055.html
https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/46055.pdf
https://www.researchgate.net/profile/Eric_Atwell/publication/267575064_Using_dialogue_corpora_to_train_a_chatbot/links/54536a9b0cf2bccc490a0890/Using-dialogue-corpora-to-train-a-chatbot.pdf
http://eprints.whiterose.ac.uk/81930/1/AComparisonBetweenAliceElizabeth.pdf
https://www.slideshare.net/saumyasrivastava2212/affective-computing
https://pdfs.semanticscholar.org/7092/6d4b8909db07b90909a4a394c16bb64a3ecd.pdf
http://www.telegraph.co.uk/technology/2016/01/21/affective-computing-how-emotional-machines-are-about-to-take-ove/
https://en.wikipedia.org/wiki/Affective_computing
No comments:
Post a Comment