Verified by 
I recently bought an exercise bicycle over the Internet. Since then, the bike has been used frequently and it's been great, so I feel obliged to post a positive review. However, posting a review can take time, because some sites insist on verbal descriptions as well as summary ratings. It’s much easier just to give one to five stars on a 5-star rating scale. Isn't this single summary rating enough?
Internet reviews may be the most frequent and widely used way in which we express opinions in a public forum. Internet reviews help us to say, in a few words even, whether an item is good or bad. But is that all we can learn from the reviews? If so, why not just use the summary ratings alone?
A recent article by two social psychologists shows how Internet reviews can be used to reliably measure important aspects of attitude that are not necessarily captured in the summary rating. Researchers at the Ohio State University, Matthew Rocklage and Russell Fazio derived a method for scaling the degree to which different evaluative adjectives are positive (vs negative), extreme (versus neutral), and emotional (vs non-emotional). For instance, we can say that the exercise bike is “magnificent” or, if we don’t like it, we can say that it is “unsafe.” Rocklage and Fazio found that “magnificent” is high in positivity, extremity, and emotionality, whereas the “unsafe” is low in all three dimensions.
The researchers were able to code 96 adjectives along these three dimensions, creating an “Evaluative Lexicon”. Crucially, these scientists were also able to find that the three dimensions were distinct; an adjective could be high in one dimension while being relatively low or high on the others. For example, “useless” and “repulsive” are both negative and extreme, but “repulsive” is more emotional. Similarly, “perfect” and “magnificent” are both positive and extreme, but “magnificent” is more emotional.
Rocklage and Fazio (2015) applied their adjective scaling to 4.2 million reviews of items sold on Amazon.com. They were able to examine the associations between each of these three adjective dimensions and the summary rating that shoppers provided. The first important result was that the adjectives that people used were highly predictive of their summary ratings.
But does the summary rating say it all? One of several useful aspects of the verbal descriptions is that they are able to reveal ambivalence. Ambivalence is an important property of attitudes. We are ambivalent when we feel both positive and negative about something at the same time. If we just select three stars (out of five) on the summary rating scale, the reader might wonder whether we were simply neutral about the item or were feeling very positive and negative at the same time.
Looking at the adjectives people provided in their written summary, the scientists were able to see ambivalence in a fifth of the reviews. Ambivalent people were more likely to select only three stars, although they often also selected more or less stars. More interesting, the direction in which they leaned (i.e., toward slightly bad or good) was most strongly predictable from whether they used more emotional adjectives for one particular valence (i.e., positive vs negative). If the reviewers used emotional positive adjectives and relatively non-emotional negative adjectives, they tended to give more stars. In contrast, if the reviewers used non-emotional positive adjectives and emotional negative adjectives, they tended to give fewer stars. In other words, emotion was the tie-breaker when people felt ambivalent.
Why is this important? This research opens the door to a much broader spectrum of research on attitudes. Internet shopping reviews could be just one of many domains where the Evaluative Lexicon could be used. Imagine applying it to comments on social media (e.g., Facebook, Twitter), political debates, news articles, and even recorded conversations about a topic. Anytime people use words to evaluate an object, we might be able to measure attitude without ever asking participants to complete a summary rating. Previously, we needed people to complete rating scales or some other type of task (e.g., computer response time measures) to infer an attitude. Now, we can infer attitudes as they are described in people’s own words.
