4 ms·

I was at a company that tried "sentiment analysis" 10 or 15 years ago. It was impossible to get right. The results were useless. It's funny how nothing seems

by ken 7y ago

I was at a company that tried "sentiment analysis" 10 or 15 years ago. It was impossible to get right. The results were useless. It's funny how nothing seems to have changed. There's always some executive who's heard of "sentiment" and thinks they can spend a week shoving a bunch of text into a database, and then have it tell you whether the words are perceived as good or bad. - I typed "Chevrolet" and it says "mpg" and "venerable" have negative sentiment, while "underachieving", "supposedly", "kilometers", and "cars" have positive sentiment. "Stunning" and "awkward" are both neutral. - I tried "Mazda" and it was only slightly better. "Praise" is neutral, "regret" is positive, and "touring" is negative. - I tried "Petzl" and only shows 16 words, all neutral, and that includes stop words like "had" and "and" and "etc". Those are at least companies with unique names. For companies with less common names, you might be lucky to get keywords related to the right company at all. I think there can be good uses for word clouds, but they are few and far between, and this isn't one. Just make 3 lists, side by side, and title them "positive", "neutral", and "negative". Instead of font size, put the more common words higher on their respective list. The only reason I can see to use a word cloud here is to hide how bad the analysis is.

jonathanbgn 7y ago
Let me elaborate a bit more on how the app computes sentiment. For a particular word, its sentiment is the average of the sentences' sentiments which contain both the word and the brandname (in order to identify the sentiment targeted at the brand, not just the overall sentiment). For example, in the case of Mazda where you say that "regret" is classified as positive, if you look into which message it comes from you can see the original sentence: "Buy a Mazda, you won't regret it :)" I agree with you that the word cloud is not useful on its own, and this is why you can click on a word to see the actual messages. Think of the word cloud as merely an entry point into a more detailed analysis by a human. Thanks for the feedback.

by ken 7y ago

jonathanbgn 7y ago

Let me elaborate a bit more on how the app computes sentiment. For a particular word, its sentiment is the average of the sentences' sentiments which contain both the word and the brandname (in order to identify the sentiment targeted at the brand, not just the overall sentiment). For example, in the case of Mazda where you say that "regret" is classified as positive, if you look into which message it comes from you can see the original sentence: "Buy a Mazda, you won't regret it :)" I agree with you that the word cloud is not useful on its own, and this is why you can click on a word to see the actual messages. Think of the word cloud as merely an entry point into a more detailed analysis by a human. Thanks for the feedback.