Once More Without Feeling (Sentiment Analysis, Part 2)

Last time, we discussed the problems text analytics for marketing has (and always will have) with sentiment analysis.

This comes off the back of our recent work with text analytics for one of the leading UK retail companies, which with the sole and ironic exception of sentiment analysis, has been very positive.

You may recall that we raised the ugly spectre of “pragmatics”. Pragmatics is a problematic area even for linguists looking at spoken conversations. Its analysis depends on an intimate knowledge of the psychologies of participants in a conversation. It also depends on a deep understanding of cultural associations. If you don’t know what they were talking about in the first place, you don’t know what they’re talking about at all.

We also said we have a workable alternative, which we didn’t mention for the purposes of narrative suspense. You’re reading this now, so it obviously worked.

A much more reliable approach to gauging sentiment is to approach the problem indirectly. Work out what people are talking about. Work out how those topics relate to each other. Although you’ll never see a sentiment score, you’ll get a much better idea of how people feel about things.

For example, if we investigate the concepts being discussed in conjunction with “clothes”, we might see associations with “availability” “sizes” and “range”. By looking at these relationships and how they are expressed, we can naturally draw robust conclusions about the sentiment which is attached to the comments.

This works because it replaces some of the wider pragmatic context that is lost by examining the qualities of the words in a verbatim in isolation. And if you have semi-structured data, you’ll likely already have an NPS or similar numerical metric for sentiment. If they’ve already given you a 9, that’s really all the sentiment you need.

But creating these links, either through a pre-determined taxonomy or through a more intuitive system, is only half the battle.

Since withholding information made you come back once, we’ll tell you all about how you can use linguists, linguistics, and creative writing to draw useful information from the conclusions of text analytics next time.

If you have any questions about text analytics for marketing and how it can provide you with real consumer insights, or you just want to express some sentiment of you own, please email us or leave a comment below.