[Vicco]

I'm writing a play about chatbots and intelligent assistants (a la Siri), and so I've come here to ask questions and maybe speed up my research.

My main question is this: what would it look like technologically if one were to imagine a very modest sci-fi leap in chatbot sophistication so that it could handle follow up questions in conversation such as, human: "Where is the Eiffel Tower?" bot: "Paris, France" human:"How tall is it?"  bot: "One thousand sixty three feet." What improvement might one imagine in the design (coding) of the chatbot that would make it recognize "it" as referring to the Eiffel Tower in the second question? Or does that already exist? Could one imagine such a mobile-device chatbot gaining its semantic capabilities through a connection to a cloud-located knowledge base? Or word net?

I'm not sure if I'm phrasing the question with enough clarity or precision yet, but there's the best I can do so far...

[lrh9]

The noun a pronoun refers to is called the "antecedent". There is currently no formal process (algorithm) to determine a pronoun's antecedent. However, determining a pronoun's antecedent usually involves examining the pronoun and contextual information to determine what it refers to. People generally receive formal education about pronouns and antecedents, but determining a pronoun's antecedent is something we often do unconsciously. We can even reliably determine a pronoun's antecedent in the face of ambiguity and incorrect pronoun usage.

http://en.wikipedia.org/wiki/Antecedent_(grammar)

[Bragi]

Hi Vicco,
you might be interested in my system. It can handle 'it' up to a certain degree. Basically, I store every part in the previous sentences, so I can query them.
I'm currently working on a mobile app, like siri (with less features of course, I'm just a single person, not an entire team). In fact, it's funny you should mention wordnet, I am looking into ways of using that dataset again: I got it imported, now I only need to massage it a little more, so it becomes useful. I looked at it a long time ago, and used it for stress testing (large dataset and all), but then dropped it again cause of the missing links. Can't beat the large dataset though.

My main question is this: what would it look like technologically

Think resonance.

[Vicco]

Thanks for replying. I'm not sure what the proper etiquette is for asking people about the details of their work, but when you say you store every part of the previous sentences (analogous to human short term memory) to be queried, does that mean your app parses the sentences, stores them for a while and when it encounters a pronoun, or "this" or "that," queries the stored parsed sentences to guess at the antecedent?

Since I'm writing a stage play for general consumption and most of it will concern relationships between the characters, I'll be glossing over details, but it means a lot to me to make the technology in it plausible. Would the following scenario be plausible? Someone developing a "Siri-killer" mobile app to have antecedent-detection and conversations that seem more context aware and semantically grounded suddenly loses access to a cloud-based system and its API that his app needs in order to function. That's the crux of the dramatic problem, so now I'm looking for what technology could hazily be imagined in the mobile app and that fictional cloud-based service.

Does that make sense? Thanks for your help!

[Bragi]

I'm not sure what the proper etiquette

Don't exactly know about the etiquette,...

when you say you store every part of the previous sentences (analogous to human short term memory) to be queried, does that mean your app parses the sentences, stores them for a while and when it encounters a pronoun, or "this" or "that," queries the stored parsed sentences to guess at the antecedent?

That's basically it yes. Though I don't really 'parse' the sentences, but I do a form of pattern matching combined with some logic for each pattern to extract the interesting bits.

so now I'm looking for what technology could hazily be imagined in the mobile app and that fictional cloud-based service.

What exactly do you mean? Are you looking for some 'plausible' implementation that such an app could have? Perhaps it could have something to do with knowledge querying (like IBM's watson does). This requires large amounts of 'factual' data, which could possibly be stored in the cloud.

[Vicco]

Well I will try to explain what I mean exactly. My difficulty finding the right words might have to do with an underlying conceptual muddiness to my question. Is there a hypothetical Siri-like mobile app one could imagine whose ability to "remember" the context of conversations (guess a pronoun's antecedent, etc) depends on querying some kind of grammar-representing... "database"? I probably need to do some more thinking and research on this before I can articulate it more clearly.

The easy thing would be for me to make it just about an app developer losing access to the equivalent of WolframAlpha, and then being unable to answer general-knowledge questions.

[Merlin]

Anaphora (linguistics) gives further background on the topic.
http://en.wikipedia.org/wiki/Anaphora_(linguistics)

A Siri-Killer could do all of this in the cell phone. Disconnection from the cloud would be no big deal. Siri is cloud only. If you are disconnected, it is like someone hung-up on a telephone conversation. In the future, logic of the Intelligent Agent may be split across the client and cloud. Currently, my bot can use Google Chrome's cloud based speech recognition but the AI resides in the cell phone. If disconnected, it could not recognize speech anymore but text entry would work (Siri doesn't currently have text entry).

Disconnection from the cloud might be more like a stroke, or Alzheimers. Short term emeory might be OK but not long term memory or other cloud based services (like search or facebook). Maybe it is like being in prison (no communication with the outside world). I sometimes wonder if kids could socially survive anymore without texting and facebook.

[Bragi]

Yes, I'd agree with Merlin.
It could be possible that, by loosing the cloud connection, the app can no longer properly resolve 'it', 'that', 'those,'....  But more then likely, it can't do anything anymore, except returning a standard reply. Something like 'no data', 'error',...

[Vicco]

Makes sense! I'm leaning toward keeping the technical glitch in the play something simple like being unable to make general-knowledge queries or some particular module stored elsewhere, rather than something integral to the AI that's split across client and cloud. Thanks!