- April
Posted By : Adrian Chan
Sentimetrix… Sentimentrix… Sentimistaken?

Sentimetrix, which I have to admit I haven’t used, makes an interesting claim viz its sentiment algos for social media sentiment analysis. The first para claims that sentiment can be measured semantically. Interesting. That would assign intensity of sentiment to the word. Makes sense, from a text analytical perspective. But misses the critical difference between the two — the speaker’s expressive intentions. Which might be to rant and rave, swoon, or exclaim…

Are some words more enthusiastic than others? Is zeal a property of the speech or the speech act?

Which is more enthusiastic?

a) Last night’s Lost rocked
b) Lost last night OMG! WTF!!!

Then there’s the matter of for whom… intensity of opinion can involve how it is addressed (and to whom)…

c) OMG did you see Lost last night? Ben rules!

Intensity of sentiment is more than a matter of factual expressions vs opinions.

d) Ben is the new leader

And self-referentiality, using “I” “my” “mine” etc is a valuable qualifier…

e) Ben is my favorite new character

Text analysis may not yet be up to parsing sentiment expressed in conversation, esp in micro-blogging and feed apps. Stylistic variations are huge, and context can be gleaned in many cases only over a series of posts (as noted below also). But the description below mistakenly places sentiment in semantic meanings, and strikes me as a misreading of expressive statements.


Measuring the Intensity of Opinion the Way People Express It
People express intensity of their opinions in two ways: by choosing words and by either expressing facts or describing actions.

For example, the word “terrific” expresses a stronger positive sentiment than “good”. Words, therefore, are a giveaway, although some cases are not as straightforward as others. We believe that once the types of words that can express opinion have been identified, it is possible to assign to each such word a measure of sentiment that an average person would associate with it. We have done exactly this, by taking a large body of texts, giving them to real people to grade, and using a mathematical model to assign to each word, a measure of opinion which would ensure the highest possible similarity between human and machine grading of these and similar texts.

Additionally, expression of facts can also play a role in shaping the opinion of a reader. Or instance, the sentence “The product didn’t work” does not express an opinion, but still plays a role in shaping opinion. Likewise, “President Castro underwent dialysis three times yesterday” does not express an opinion, but certainly plays a role in shaping our perception of Mr. Castro’s health.

Our technology allows both types of sentiment shaping sentences to be evaluated.


Leave a Reply