AI Zone Admin Forum Add your forum

NEWS: survey on 3000 US and UK consumers shows it is time for chatbot integration in customer service!read more..

Rating a blob of text

Hello everyone,

For the past six months or so I have been sitting on a project that used a neural network to assign a 4-dimensional rating on an object, based on a text description. I imagine a lot of people will groan at my use of a neural network for a linguistic task (but who knows? Maybe chatbots are all neural networks under the covers?)

I did so because four dimensions is pretty simple and my text source - wikipedia smile - has a whole lot of meta data that I am perfectly content to rely on. I am not after purity, I am trying to make a fun imaginitive game and wikipedia’s licenses allow doing so. If I am trying to match a human interpretation of something, and the program uses reflections of large-scale human biases down on things like how many words have been contributed, links people decided add, etc in order to somehow infer how humans may feel about the object, so be it.

But anyways I have millions (trillions? months) of CPU cycles that evolved it into…well, not very good.

But all of this babbling aside, if I am trying to take a wikipedia article and determine from the text things including how good the person in the article would be in a fight, what tools exist? What models should I use? How slow would the calculations be? Is this idea totally absurd? The entire project hinges off of determining these things well enough to be funny when wrong, no better but certainly no worse. smile


  [ # 1 ]


I have found that the deeper you go into *meaning* (the meaning of meaning) the murkier it gets.  Semantics, that’s really the $64,000 Question, isn’t it?  My understanding of neural networks is that it’s not very well understood.  Learning, the extent of my foray into machine learning for chatbots has been classifier APIs; see my Quora answer to: “Are there any online Bayesian SaaS text classifiers apart from” (above link).

Triples have been a recent revelation for me, so simple yet so obscure.  Triples are basically simple sentences, with subject, predicate and object.  Triples represent relations, meanings beyond the words.  Thus, if your text is annotated with semantic metadata, you should be able to apply a semantic reasoner, or OWL reasoner.  I don’t know of any available dialog system that is build primarily upon semantic metadata, or semantic reasoner.

Then there is the statistical approach.  I’m not convinced that n-grams are adequate.  For myself, I prefer skip-grams (s-grams), or “gappy” n-grams.  However, in the end, what even qualifies as hit or miss in terms of answers?  I’m not a fan of the Turing test anyway, and believe it to be a red-herring.  Who needs machines that can fool people anyway?  I want to skip forward directly to machines that are way better than people, at certain things, and therefore be used as better tools….


  [ # 2 ]

I’m using neural nets, but a little different compared to traditional 3 layered nets. You can check out my blog and/or homepage for more info.
I recently made an insult detector using this system for a kaggle competition.


  login or register to react
‹‹ Chatbot API      Will Dockery A.I. updates ››