Two more points..
First, the AIs will be like those in Iain M Banks' Culture novels, they'll have
their own personalities which will be co-evolved with ours, but utterly
different (like our cats are now). These will not be emotions as we understand
them (which are dominated by true evolution and the hormones it creates), but
will arise spontaneously from the complexity of their mental lives.
Second, and on the technical points: CEPT mainly uses Wikipedia as a source,
and, as Wikipedia tries to suppress polemic and over-opinionated text, the
corpus will be very dry emotionally. Also, as pointed out in that article, the
sequencing of words is critical to inferring the emotional intent. CEPT at
present does not include this element of ordering (that's why I'm involved).
I'm surprised that these guys didn't use the ratings associated with each
comment on Rotten Tomatoes to associate the sentiment with the text. This would
have saved them a fortune, and the same technique coild be used with other
systems like Amazon.
Regards,
Fergal Byrne
—
Sent from Mailbox for iPhone
On Sun, Dec 8, 2013 at 6:41 PM, Fergal Byrne <[email protected]>
wrote:
> Hi /Mar(e?)k/,
> Thanks for the link. This is a hard problem, which is why they resort to
> getting Mechanical Turk people to label the phrases. That's effectively
> programming the system, so it's yet another apparent step forward, which is
> nothing of the kind.
> The "emotional content" of text is all about how the current input echoes
> other contexts in which the words were used, and a memory of how we felt
> then. This is something which no AI system can truly learn, since it has no
> genuine intrinsic emotional "sense".
> Jeff has no illusions about this. He's said many times that the goal of
> creating an emotional AI is flawed. It's a hubristic notion rooted in our
> wish to dominate: why create an emotionally sentient intelligence only to
> have to choose between granting it rights or enslaving it?
> By all means the systems we will build should have an awareness of our
> emotions, but this will be treated just like any variable in their thinking,
> rather than being an intrinsic governor of their minds, as we have.
> Regards,
> Fergal Byrne
> —
> Sent from Mailbox for iPhone
> On Sun, Dec 8, 2013 at 6:03 PM, Marek Otahal <[email protected]> wrote:
>> Hi all,
>> I just stumbled upon an article about natural language processing and
>> sentiment analysis, a topic I'm interested in.
>> http://gigaom.com/2013/10/03/stanford-researchers-to-open-source-model-they-say-has-nailed-sentiment-analysis/
>> The idea they used seems to work for them well and is simple in its nature.
>> When I've seen it, I immediately thought CLA (for its sequence learning
>> ability) and CEPT (for the "ontological matrix" view on words). This
>> combined should give interesting results and has a good (commercial)
>> usecase.
>> Tricky part would be training the classifier, getting some people to label
>> the phrases (or some smart ways around this).
>> Regards, Mark
>> --
>> Marek Otahal :o)
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
