3
$\begingroup$

In Sentiment Analysis, I've seen the sentiment of an entire string of words (tweets, paragraphs, sentences, etc.) computed by summing the sentiment of all of the terms in the string.

I know there is a monotonic relationship between summing and averaging. So, if sentences are only ever compared to other sentences, then of course it wouldn't make a difference.

So, why is it that the average is not used instead of the sum? Especially if there are cases where strings could contain only one word, or many words, it seems to me that an average sentiment would be the way to go. Is that not true?

$\endgroup$
1
  • 1
    $\begingroup$ There are a lot of nuances in choosing an aggregation method, such as whether (corpus) normalization has been performed and how. So much so, that it will be difficult for people to answer this without a specific reference to where you have seen summing employed. Could you please add a reference to where you are seeing this so we can see the entire featurization, normalization, and aggregation process? Thanks! $\endgroup$ – AN6U5 Sep 24 '15 at 16:20
5
$\begingroup$

Consider the following two sentences:

My awesome girlfriend bought me a delicious popsicle at the store.
0--1-------0----------0------0--0-1---------0--------0--0---0-----:2:11

My awesome girlfriend, Joyce, drove to the grocery store to buy me a delicious Dole popsicle.
0---1------0-----------0------0-----0--0---0-------0-----0--0---0--0-1---------0----0-------:2:16

See how you can sum the values and get 2 for both, but average the sentiments and get 2/11 vs. 2/16. There are cases where people have seen better results measuring the total number of good sentiment words rather than letting zero value words cloud the picture. In this case, I would argue that the sentences have the same sentiment and averaging artificially pulls down the sentiment of the second sentence. In certain methods like TFIDF, the word vectors are normalized against are relevant corpus, so this may change things. In other methods, the stop words may have already been removed, so this has less effect.

In twitter examples, users are limited to 140 characters, so one might consider this to be an inherent denominator which makes averaging really averages of averages. Its very dependent on the specific application and methodology.

Without a more specific reference, it is difficult to be more specific in my answer.

Hope this helps!

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.