AI Zone Admin Forum Add your forum
What is a conversation and what is a CONTEXT ?

What is a conversation and what is a conversational context?

I would be glad to hear how different forum members think about this. I am not interested in technical details but in a general picture and ideas.

For example, with the chatbot I am working on, I have a single data structure that I want to fill by eliciting information from a client. So the chatbot always knows what it wants to talk about. Perhaps modeling the client’s thinking is the correct combination of models: an internal chatbot data model and an external client ‘thought’ model.

How do you approach this?


  [ # 1 ]

There are 3 parts of context:
What the chatbot wants to talk about.
What the user wants to talk about.
The steam of the conversation (topics discussed in the current session).

You could also add topics of previous conversations with this user.
The ability to maintain the conversation stream (related volleys) or to successfully change to a topic that the user is willing to engage in, will often define how well a bot handles “context”.


  [ # 2 ]

Very interesting question. There is not one type of conversation, but several conversation depending of the aim of each participant.

Some links : In french but Google Translate is your friend wink

The paper is focused on man machine dialogue (MMD)
domain. The main contribution is about the
formalization of the dialogue process between an user
and the machine, in terms of goals and strategies. The
approach is based on the beliefs, the goals and the acts
and contains implicitly a deontic logic (as « social »
obligations of the partners). Our model take place in the
general framework of “game theory”. The paper
presents essentially some features of a formulation of
dialogue game theory that attempts to mirror and
expand a theory of conversations proposed by
Vanderveken in the context of illocutionary logic.



  [ # 3 ]

I think a good conversation is any dialogue wherein two people contribute to eachother’s understanding, like two players taking turns building a house of cards. I do not think of context as topics or a general client model, but as tracking what has been recently said. The bare minimum is being able to figure out pronouns and shorthand replies, just to know what the user it talking about -at all-. Most chatbots don’t even manage that, which means the house of cards collapses at the base every time.

I approach the matter through psychology. Real humans are not often aware, or care about, eachother’s hidden goals in a conversation, and often those goals change on a dime. This means that in my opinion it isn’t necessary to distill the user’s long-term goal and plan a strategy. But rather to figure out their short-term goals: To guess what the user is trying to get at with the last two or three exchanges. For this I monitor a bunch of long-range variables, like how much the user seems to like the subjects they mention, how much they seem to know about them, the overall emotional state of the user, whether they state a problem or opinion, etc. These variables help decide what type of response would be most appropriate: asking information, giving information, feigning interest, summarising, suggesting solutions, showing sympathy, or small talk, to name some options. There’s been some previous discussion here.


  [ # 4 ]

Thank you Don. After three sentences I am already thinking - “yes” this is what I am wondering about. Putting this in computer-ease: the past conversation must be a data structure which can be “cast” as a pronoun [or a pair, or as anything which can become implicit in the next few phrases]. A good thought.
The list of response types is also quite interesting.


  [ # 5 ]

That’s what I do anyway, I keep the last 100 sentences of bot and user around as a grammar-parsed structure, and look back which recent words are of the grammatical type(s) that correspond with a pronoun. Insofar as I know, ChatScript also provides a grammatical framework for this, and AIML can handle pronouns to some extent by manually marking words to allow referring to later. Stanford also developed free tools for pronoun resolution. The same structure can help out with shorthand responses where e.g. the verb is omitted.

Denis’ first linked paper mentions Searle’s Discourse Analysis research on types of responses called “speech acts”. I personally disagree with their arbitrary categories, but you could check them out for inspiration. Mine I got from sociology books on how to be a good conversationalist.


  [ # 6 ]

The difference between context and conversation is sound, I think.

What do you think?


  [ # 7 ]

Hi Guru: I may not be using “conversation” and “context” the same way as you, so I may not be understanding you. I am asking about context for a text-based exchange . I suppose that sound context is forgotten at about the same rate that text based context is.


  [ # 8 ]

Peter, you are correct!  The difference between context and conversation is not always sound.

In Latin “con” means with or together.  Both context and converse share “con” as a prefix.
Latin “textus” means literary style, which relates to Latin “versus” meaning, a line of writing.
So, “con-textus” and “con-versus”, both have “with” and “writing” in common.

By modern definition, verse is a verb for “speak” which may involve sound, but not always.




  [ # 9 ]

Guru are you imitating a chatbot? I see you use the same word format in other posts. What do you think? I think someone is playing a joke on us.


  [ # 10 ]

I guess you could say that’s just my “textus” (Latin for writing style), Peter.

Congratulations on starting a good topic to analyze for discussion purposes.


  login or register to react