Twitter is an ideal channel for finding out what users think of a company and the services it offers. Hence the need for fast, automated ways of collecting opinion data.
Online reputation management is starting to be standard practice for brand names, which have everything to gain by using the information collected about the brand on sites such as Twitter in order to improve their products. With this in mind, a team at AT&T Labs, headed by Narendra K. Gupta, has developed a method of autonomous filtering which is able to identify tweets that are talking about a particular brand, and then extract and classify the data. Dr Gupta believes that problems concerning his company’s products come to light much more quickly on the micro-blogging platform than in official incident reports. As far as minor problems are concerned, it seems that these only show up on the social network. So he has tried to transpose studies carried out on natural disasters to a more commercial field which could be of use to companies.
Identifying a customer’s opinion or concerns
The information collected covered users’ opinions, problems they have encountered, questions asked and some simply descriptive messages. The method analyses the tweet content by automatically removing the hashtags and correcting punctuation excesses and abbreviations. It even takes into account the emoticons expressing feeling such as happiness, sadness or anger. Then, using a database comprising frequently-used key words or those corresponding to the brand (here we’re talking about a telecommunications company), plus verbs and expressions conveying emotion, the system can identify what the tweet is talking about, classify it - putting it into a category such as information, questions or problems - and finally indicate whether the user’s opinion is positive or negative.
A success rate of 70%, under certain conditions
Using this method of correction and dynamic word fields, the system is capable of automatically identifying what the tweet is talking about, and the customer’s opinion as conveyed by the tweet, with a success rate of 70%. The study nevertheless makes it clear that the last step, where the tweet is automatically put into a category, is still a work in progress. Dr Gupta also stresses that since language fashion and usage trends change very quickly on this kind of social network, it’s essential to keep updating the database containing the key words and the corrections; otherwise the exercise is doomed to failure.