Online reviews for products are majorly influencing purchasers buying decisions. 70% of people now consult online reviews before purchasing products. With this being said information overload for reviews is becoming a significant problem for online retailers. The user is presented with a bewildering number of reviews for each product. To compare two items the user must read a large number of reviews to get a fair and overall view of a product.
Amazon has tried to solve this by presenting the most helpful and unhelpful reviews. The problem with this scenario is that if the user only reads these two reviews, there is no guarantee that these reviews give a fair representation of the overall product. Amazon also provides a section for technical details/features. I believe reviews and features can be combined in a graphical way, to provide the user with information that may significantly influence their buying decisions. The method I have devised is to combine both the technical details/features with the user-generated reviews through sentiment analysis, opinion mining and natural language processing. The final goal is to present the results in a honeycomb structure.
The technique I have applied involves analyzing every user review for a product. The features mentioned in each review are then extracted along with the positive, negative or neutral sentiment relating to that feature. The image below shows an example of the extraction.
The extraction of sentiment is done by first splitting the text into sentences. Text chunking and part-of-speech (POC) tagging using maximum entropy have also performed. After the extracting of the features and positive/negative/neutral sentiment, a variety of processing techniques are applied to ensure the correct sentiment count for each feature have been obtained. The following image is a sample table of the sentiment for the top 10 most talked about features for a given camera. The table shown is generated from a subset of real user reviews from Amazon. Neutral sentiment indicates no positive or negative sentiment words appeared in the same sentence as the feature.
Once this table of information has been generated, I represented each feature as a hexagon. I varied the size of the hexagon depending on the word count of that feature. I also altered the color depending on the ratio of positive and negative sentiment. The equations I devised are shown in the following picture.
You then output the data in the sentiment table through the above equations. Then you order the hexagons from negative to positive. The honeycomb-like structure is shown below:
The following is another design I toyed with for some time: