Sentiment Analysis/Appraisal Theory

By Eric Cruet

Opinions are like elbows; everyone has two of them for every topic.  Scientifically, however, opinions are very difficult to examine.  As of late, the computational linguistic community has recognized value in extracting, mining, and analyzing opinions from bulk text found in SMSs (Social Media Sites).  Sentiment Analysis is the task of having computers use machine learning algorithms to automatically perform such tasks, and attempt the classification of the opinions into “emotions”.

Computational approaches to sentiment analysis focus on extracting the affective content of the text from the detection of expressions of sentiment.  These expressions are assigned a positive or negative value representing the corresponding positive, negative or neutral sentiment towards a specific issue.  For example, using information retrieval, text mining, and computational linguistics, one can calculate opinions using the Support Vector Machines classification algorithm with a “bag of sentiment words”.  This technique was very popular for movie review classifications.  In a bag of words technique, the classifier identifies single word opinion clues and weights them according to their ability to help classify reviews as positive or negative (the number of times they appear).   So the word “sucked” (as in the movie sucked) would have a higher weight than the word “ok” (as in the movie was ok).


It is obvious that there are many opinion scenarios that this classification technique will not address.  For instance, it cannot account for the effect of the word “not” which will turn a review of “good” into “not good”, thereby reversing a positive sentiment into a negative one.   It also cannot account for more complicated sentiments i.e. “I wish the movie was in 3D.”   In order to incorporate more complicated sentiment tasks, it is important to further structure the approach in order to capture these elements.

The tasks described above were part of sentiment classification. In order to incorporate more complicated sentiment tasks, a more appropriate technique is structured opinion extraction.

The goal of structured opinion extraction is not only to extract individual opinions from text, but to also break down those opinions and parts so that those subcomponents can be used by sentiment analysis applications.  This is defined by identifying product features and opinions about those product features.

One way to accomplish this is using an appraisal expression. An appraisal expression is a basic grammatical structure expressing a single evaluation, based on linguistic analysis of evaluative language, to correctly capture the full complexity of opinion expressions.  Most existing work and corpora in sentiment analysis have considered only three parts of an appraisal expression: evaluator, target and attitude.  However, Hunston’s and Sinclair’s [3] local grammar of evaluation demonstrated the existence of other parts of an appraisal expression that can also provide useful information about the opinion when they are identified. These parts include superordinates, aspects, processes, and expressors.

Evaluator I attitude love it target when she walks to me and smiles.


Target He is Attitude one mean superordinate bastardevaluator said the employee.


Extracting appraisal expressions is an essential subtask in Sentiment Analysis because it provides sentiment words that can help define the features used by many higher-level applications.  As stated in Cognitive Appraisal theory, we decide what to feel after interpreting or explaining what has just happened. Two things are important in this: whether we interpret the event as positive or negative and what we believe is the cause of the event.  The resulting classification of the appraisal expression allow for a finer granularity in the application of quantitative methods so the results more closely represent what is being measured.





Google Prediction API


[1] Asher, N., Benamara, F., & Mathieu, Y. Y. (2009). Appraisal of opinion expressions in discourse. Lingvisticæ Investigationes32(2), 279-292.

[2] Hunston, S., & Sinclair, J. (2000). A local grammar of evaluation. Evaluation in Text: Authorial stance and the construction of discourse, 74-101.

[3] Jackendoff, R. (2002). Foundations of language: Brain, meaning, grammar, evolution. Oxford University Press, USA. (2), 279-292.

[4] Taboada, M., Brooke, J., Tofiloski, M., Voll, K., & Stede, M. (2011). Lexicon-based methods for sentiment analysis. Computational linguistics37(2), 267-307.