[Corpora-List] EmoText - Software for opinion mining and lexical affect sensing
Ken Litkowski
ken at clres.com
Wed Dec 21 16:15:52 UTC 2011
As the purveyor of what I wish was a commercial-grade system, known as
Minnesota Contextual Content Analysis (MCCA), I can comment on Michal's
85% vs 82% dichotomy. Analyzing tweets or microblogs and sentiment has
become the research topic du jour. I was intrigued to examine how MCCA
might fare off the shelf against a very elaborate rule-based system
developed at Xerox-PARC for classifying tweets into questions and
not-questions. I've described this comparison on my blog
(http://www.clres.com/blog/?p=195), where the accuracy levels were 67%
vs. 62%. There are a couple of additional comments that seem relevant to
the points discussed in this thread.
The first is that context is very important, as Alexander pointed out,
so that having the discussion context does help in classification. (This
is a core component of MCCA.) MCCA uses a dictionary in which words are
classified into 116 categories and assesses their usage in 4 contextual
dimensions (hence, a 4x116 array). An important characteristic of the
lexicon is that it captures a distinctive level of granularity, above
the word level and below the genre level.
I was very encouraged by my results and intend to examine further the
possibilities of MCCA as a classification tool. However, I was brought
up short in a discussion about this with Ed Hovy, who has been
classifying microblog events from an IR perspective, and who was
dismissive of MCCA as anything new under the sun. I haven't yet
attempted the necessary work to determine the validity of Ed's point.
But, his perspective, and mine, suggests that there is a subtle
interplay between opinion mining and IR that is worth investigating.
Ken
On 12/21/2011 8:34 AM, Michal Ptaszynski wrote:
> Dear Taras, Iain
> Dear All,
>
> What Iain and taras say is one of the best things I've heard lately,
> mostly, because it confirms my findings too. However, your experience
> is probably based more on real world examples.
> If you could provide a proof of some kind, or a description of some
> examples, this would be a very useful hint.
>
> Just a word on "keeping classification cheap". I think this is not as
> much about the money, as it is about logic (and trying to find it).
> For example, it is not much of a research to just have, for example,
> 100 people write a lot of rules. Even if a system cerated this way
> would achieve high performance its not too interesting from the
> scientific point of view.
>
> What we, researchers, try to find is a kind of logical reasoning that
> could be represented computationally. So, for example, if Mr.X has a
> 1000-rule-system that gives him 85% accuracy, and Mr.Y has a
> 10-(general)-rule-system that gives him 82% accuracy, a researcher
> would rather be first interested in Mr.Y's system.
>
> I think this applies to all fields that have their commercial
> variations. For example, each year there is a number of papers on
> machine translation presenting high results, but the level of actual
> machine translation software available on the market is rather low (As
> a former translator I tried about 5 different ones).
>
> Best,
>
> Michal
>
>
> ---------------------
> Od: Taras <taras8055 at gmail.com>
> Do: corpora at uib.no
> Data: Wed, 21 Dec 2011 10:43:47 +0000
> Temat: Re: [Corpora-List] EmoText - Software for opinion mining and
> lexical affect sensing
>
>
> Hi
> I am a developer of one of commercial tools. And I think there are two
> majour problems that prevent them being more accurate:
> 1. They try to keep classification cheap. Cheap means generic. But the
> only way of getting a good sentiment accuracy is making classifiers
> specific. But this is expensive.
> 2. The other problem is the neutrality bias. In most cases texts are
> usually neutral or balanced, and it makes extraction of non-neutrals
> very difficult. The problem actually is not subjectivity or sentiment
> classification taken separately, but the combination of the two.
>
> Of course there are other problems: noisy language, various ways of
> expressing sentiment etc. But the two aforementioned are the most
> business-specific ones.
>
> Regards,
>
> Taras Zagibalov
>
> On 21/12/11 09:52, iain wrote:
> I've been following this thread with interest. I'm a commercial
> semi-lurker
> rather than an involved theorist, but my colleagues and I have done some
> work with some of the available commercial sentiment tools.
>
> Our experience is that they are really not very accurate.
>
> There are some issues with evaluating them. I'm not using a
> pre-marked gold
> standard to score them, but rather submitting text from web pages to them
> and comparing the output with the text, which makes our results far from
> scientific! And we've done dozens not thousands.
>
> What we tend to find is that we look at the output and then at the
> text and
> more often than not say, 'uh'. Pretty much as the reviewers of
> Alexander's
> test site have been doing! Which might make Alexander's work close to
> the
> commercial state of the art :-J ....
>
> Some of the reviews of commercial tools I've seen seem to indicate
> that if
> you take the 'neutral' sentiment articles out then the actual accuracy
> drops
> down from the claimed 70% quite considerably. In short, the tools are
> very
> good at detecting no sentiment but rather poorer at getting actual
> sentiment
> right.
>
> I was wondering if anyone on the list had experience with the commercial
> tools and what sort of results they found. Could they recommend one or
> another of the suppliers? I'd also be interested if any tool suppliers
> (also commercially semi-lurking ) might have some input to this -
> what is
> their real expectations of quality?
>
>
> Iain
> -----Original Message-----
> From: corpora-bounces at uib.no [mailto:corpora-bounces at uib.no] On Behalf Of
> Justin Washtell
> Sent: 20 December 2011 20:32
> To: Alexander Osherenko; ptaszynski at ieee.org
> Cc: corpora at uib.no; corpora-request at uib.no
> Subject: Re: [Corpora-List] EmoText - Software for opinion mining and
> lexical affect sensing
>
> Michal and Alexander,
>
>
>
> I thoroughly agree with Michal (and Graham) that these kinds of demo
> are a
> good thing, and despite my - ongoing - criticisms, I'd like to take my
> hat
> off to Alexander for sharing this work. There are already countless
> papers
> describing technical approaches to this-and-that, and showing
> impressive-looking results achieved upon [perhaps sometimes carefully
> selected or tuned-to] test datasets. But I suspect that there's
> presently no
> better way to get a feel for where the state-of-the-art really is (and to
> shed some qualitative light on matters) than by complementing these works
> with some inquisitive and unrestrained hands-on tinkering.
>
>
>
> I tried a couple of reviews from Amazon. Among different feature sets
> from
> 1 to 6, always one is close to the amazon's ranking, but unfortunately
> its
> never one feature set in particular, but rather randomly one from the
> six.
> Besides the closest method, all other are usually reversed (e.g., if the
> closest method gives 5 star, all other give 1). However, this might have
> just happen for those couple examples I tried (Reviews of Kindle on
> Amazon).
>
>
>
> Isn't that more-or-less what one would expect from random output?
>
>
>
> Social aspects. You have to consider that the reviews from Amazon are
> composed by different authors that have their own style of writing.
> Moreover, you have to consider different cultural background, for
> example,
> Americans and Englishmen use different words to express same things.
> Goethe
> used other words than a truck driver does.
>
>
>
> As a human, and an Englishman, I expect I can understand and fairly judge
> the sentiment of most reviews written by, say, an American truck driver,
> without undue reprogramming. Is this really an unrealistic goal for our
> algorithms? And I wonder, is mastering a highly restricted style or
> register
> a necessary step in that direction... or is it in fact a detour.
>
>
>
> How can a classifier calculate a weight of a lexical feature if this
> lexical feature is not present in the analyzed text?
>
>
>
> By inferring from similarities between that feature and those that *are*
> present (e.g. through semi-supervised learning/bootstrapping of
> unannotated
> data)? That's at least one method about which a fair amount has been
> written
> already. I'm not saying its a solved problem mind you, but perhaps you're
> not up against a brick wall yet?
>
>
>
> Justin Washtell
> University of Leeds
>
> _______________________________________________
> UNSUBSCRIBE from this page: http://mailman.uib.no/options/corpora
> Corpora mailing list
> Corpora at uib.no
> http://mailman.uib.no/listinfo/corpora
>
>
> _______________________________________________
> UNSUBSCRIBE from this page: http://mailman.uib.no/options/corpora
> Corpora mailing list
> Corpora at uib.no
> http://mailman.uib.no/listinfo/corpora
>
>
> _______________________________________________
> UNSUBSCRIBE from this page: http://mailman.uib.no/options/corpora
> Corpora mailing list
> Corpora at uib.no
> http://mailman.uib.no/listinfo/corpora
>
--
Ken Litkowski TEL.: 301-482-0237
CL Research EMAIL: ken at clres.com
9208 Gue Road Home Page: http://www.clres.com
Damascus, MD 20872-1025 USA Blog: http://www.clres.com/blog
_______________________________________________
UNSUBSCRIBE from this page: http://mailman.uib.no/options/corpora
Corpora mailing list
Corpora at uib.no
http://mailman.uib.no/listinfo/corpora
More information about the Corpora
mailing list