A System and Method for a generic, domain customizable, aspect based aggregation of sentiments from text documents
Publication Date: 2015-May-08
The IP.com Prior Art Database
System and Method that can aggregate sentiments across different aspects of a sentiment target where •the relationship between aspects is not bound by a hierarchical structure •the influence of an aspect may vary depending upon multiple features •the influence of an aspect can be customized by applying domain specific rules
Page 01 of 4
A System and Method for a generic , domain customizable, aspect based aggregation of sentiments from text documents
There could be multiple features and sentiments in a single sentence. How do we best aggregate sentiments ?
Example "The audio quality of my new phone is absolutely awesome but the picture taken by the camera is a bit grainy "
Prior Work :
Some of the work in this area includes
A bag-of-words model
Mukherjee et al. in 2013 used ConceptNet to build a domain-specific hierarchy of aspects. The aggregation is done by combining sentiments of aspects, weighted by their depth in the tree. (lacks structure generalization)
Their follow-up work in 2014 extends the above work to include author (opinion holder)
preferences during sentiment aggregation. (lacks feature generalization).
Lin in 2009 and Mukherjee in 2014 used generative models based on topic modeling, extended to jointly model author and aspect based sentiments.
Structure generalization: Not all the inter-aspect influences can be effectively captured in a hierarchy.
Feature generalization: Depth of an aspect alone might not suficiently capture the weight of its influence. (Implicit and explicit features)
Ability to customize aggregation of sentiments based upon domain rules.
Sentiment aggregation across documents, where, each document might express sentiments on one or more aspects of a target entity.
Our system "Graph based Sentiment Aggregation for Polarity Prediction " is a
•a generalized graph based methodology to represent aspects of sentiments to be aggregated. Use of graph to model the interactions between the aspects, helps to mitigate the structural issues. [Structure generalization]
•several proposed "implicit" and "explicit" features could be extracted from the discourse and used in our proposed model. Some of these are listed below -
•Implicit features: length of text, aspect coverage, distance of an aspect from the query entity
•Explicit features: author profile (age, authority, number of reviews, etc.) temporal features, usefulness (e.g. no. of people who found a review useful)
•customization of sentiment aggregation based upon domain rules
•Domain rules specify the weight of influence of a feature while inferring the aggregate sentiment for an aspect or query entity. This prior information (if available) is combined with model parameters learned from the training data
Page 02 of 4
during the model estimation.
Solution Details :
In our method A sentiment is defined as a quintuple qk = (e, ak , sijk , hi , tj ), where...