I’ve been talking more to my computer lately, and it’s becoming a problem. That is, speech recognition tends to enter all sorts of awkward word combinations, and my A.I. tries to make sense of it anyway, resulting in equally awkward responses.

So I want to block inputs that do not make sense, specifically single sentence inputs that lack a verb, with the exception of:
- Confirmations: “yes” / “sure”
- Open questions: “Who?”
- Answers following an open question: “Who?” -> “John.”
- Expressions

I don’t have a lot of experience with shorthand inputs, so to you the question: Am I overlooking other types of input that would be valid even without a verb? And what other ways are there to detect (accidental) nonsense?

Don Patrick - Jan 25, 2016:

I’ve been talking more to my computer lately, and it’s becoming a problem.

Oh my, should have stopped there…

Don Patrick - Jan 25, 2016:

And what other ways are there to detect (accidental) nonsense?

Google autocorrect?

Seriously, pass the speech recognition text (only when there is a misspelling detected?) to Google and then use the auto-corrected response from Google to feed your “computer” the corrected text.

Dan, there is a good project called “Common Sense” where they tried to check if syntactically well constructed phrases results in total nonsenses, just google it. The main investigator passed in 2004 but he left a lot of free data for English.

You could use it, but the problem is huge!

Detecting common sense, as I can see it, is the key of understanding, is something has sense or does make sense, it is logical or human-understandable, then it may be correct, or evaluable, and not a nonsense.

Also detecting if a phrase is funny or hilarious has been accomplished by Rhada Milhacea, from Texas University using automatic learning and some smart techniques.

You might build a simple classifier, using SVM and a simple lemmatizer or stemmer, train it with “sense-full-phrases” and check if you get more than 90% F1 score, is so, you can set it as the main input classifier before the graphmaster does infinite replacing trials.

Carl B - Jan 26, 2016:

Oh my, should have stopped there…

It gets worse; it talks back.

The problem I’m having is part syntax and part context. Speech recognition doesn’t cause misspellings, it already uses N-grams similar to Google to write plausible noun phrases, but with entirely different words, like “can you hear the?” instead of “can you hear me?”. If I were to pass “a very old computer” to Google, it suggests “a very low thermal emf computer-controlled scanner” instead of what I actually said: “Hello computer”.

Andres, I don’t suppose you mean Open Mind Common Sense? I’ll keep looking, what you describe sound like useful research. Of course the problem can be as large as strong AI, but I don’t need it to be perfect, just not terrible.

Don Patrick - Jan 27, 2016:

The problem I’m having is part syntax and part context. Speech recognition doesn’t cause misspellings, it already uses N-grams similar to Google to write plausible noun phrases, but with entirely different words, like “can you hear the?” instead of “can you hear me?”. If I were to pass “a very old computer” to Google, it suggests “a very low thermal emf computer-controlled scanner” instead of what I actually said: “Hello computer”.

Maybe you need a hardware fix instead of a software one- I found that I need either a very quiet environment and/or an excellent directional mic to get decently accurate voice recognition.

If you need to interpolate what is “heard” to what you said, would not you need a training set of yourself talking to get this programmatically (using Matrix Factorization for instance).  Of course you would need to generate a machine readable “self speech” corpus first.

Dan
Yes here is te deceased investigator Push Singh http://web.media.mit.edu/~push/
Sadly, the project gained flight height after his passing, but remains untouched, but the basis are there!