Let's assume a system works by extracting the list of opinions which it considers occur dominantly in the product reviews. Following is a sample list of opinions for the given product:
Next, our sample system identifies which of the given reviews contain which opinions. This is represented in the form of an opinion matrix. Each row corresponds to a review, and each column corresponds to an opinion. A cell in this opinion matrix at the intersection of row r and column c will be marked 1 if the rth review expressed the cth opinion, and 0 otherwise.
Lastly, before submitting results to the organisers of the shared task, remove the column names, i.e. the specific opinion labels.
MOTIVATION: Specific labels do not hold much value as they're expressed in ambiguous and unverifiable natural language. We will instead use more complicated metrics to evaluate whether or not opinions were mined correctly or not.