Sentiment Analysis/Opinion Mining
Sentiment analysis (or opinion mining) focuses on analysing web documents, especially user-generated content such as product reviews, to identify opinionated documents, sentences and opinion holders. Compared to traditional document classification, sentiment analysis and polarity classification are significantly harder. One of the main applications of sentiment analysis is in voice of the customer systems.
One of the sentiment analysis systems we developed recently uses supervised machine learning methods, trained on human-annotated data, co-occurrence statistics, and lexicons of positive and negative words, in order to identify problems with products and company services reported on blogs. We also have experience with lightly-supervised and unsupervised approaches.
Unlike other sentiment analysis work, we also consider richer user-sensitive models of opinions to reflect the fact that the same opinion could be positive for one user group and negative for another (e.g. a camera can be rated as good for a photographer, but too complex for the casual user). Similarly, we model the fact that negative opinions on certain features can be more important than others (e.g., limited choice of colours vs. poor battery life).
Where applicable, we also combine sentiment analysis with author authority models to account for the fact that personal authority often influences whether the reader will be influenced significantly by a given review or not (e.g., an endorsement from an unknown person does not carry the same weight as that of a respected expert).
Another important dimension of sentiment analysis which is often overlooked is the temporal element. We not only identify opinions but also identify when they were stated, which makes it possible to detect shifts in attitudes over time. For instance, initially most users might have reported predominantly positive experiences with a given product, but later on faults started being reported. The temporal dimension is particularly important in applications such as monitoring the impact of marketing campaigns or containing damage to brands and companies through quick response to problems as soon as a significant number of users start reporting them online.