Sentiment Analysis

IntuScan™ performs deep sentiment analysis to identify not only the general sentiment of the text towards entities, but also: the specific nuance of sentiment towards each entity; the “aggregated sentiment” towards the entity, as deduced from different expressions of sentiments, and the inferred sentiment towards entities who are not mentioned explicitly in the text but alluded to through the affiliation of explicitly named entities with them.

The IntuScan methodology is based on the following stages:
  1. Differentiation between different language registers, domain “jargons” and lexicons of topics in expression of sentiment. For example, the language register of British youngsters will refer to something positive as “wicked”; in the United States, this appellative would be negative.
  2. Ontology-based analysis to apply extra-linguistic language-dependent and culture-dependent information (some contextual within a text but some from totally extraneous knowledge) For example, if a mother says to her ten-year-old son "Mommy's little boy", it is an endearment. If his classmates say it to him, it is negative. Another example: appellatives such as "nationalist" may be positive when used by a person identifies with a “right wing”, but negative when used by left-wing individuals. In Arabic, “Shiite” and "Iranian" (e.g. "Iranian Hamas") may be used as a derogatory epithet by Arab Sunnis whereas when used by Shiites, it would be simply a statement of sect or nationality. Languages that are rich in ways of expressing negative sentiment also employ oblique references. For example: Hizballat instead of Hizballah or Nassrallat instead of Nassrallah (replacing the part of the name that is a reference to Allah with Allat - the ancient pre-Islamic Goddess whose statues in the Kaaba Muhammad destroyed.
  3. Identification of the semantic affiliation between the ontological concepts. For example, the negative indicators “untrustworthy”, “unreliable”, “deceitful”, and “devious” all belong to a common “parent concept” (for example “untruthfulness”) with different nuances of intensity. An entity that is described in a number of these terms is tagged with the parent concept.
  4. Aggregation of information regarding sentiment towards specific entities to a general aggregated sentiment towards their “parent” entities. For example, negative sentiment towards entities that belong to a parent entity (e.g. senior officials of a country who all have in common their affinity to the country or a number of products that all belong to a certain company) may indicate a general negative sentiment towards the parent entity.
  5. Integration of the above algorithms with statistical methods that analyze an input text by comparison of its ontological digest with statistical models of sentiment, based on tagged corpora of texts.
The advantages of the IntuScan technology are, therefore:
  1. Identification of sentiment not on the basis of clusters of "positive" and "negative" indicators but through identification of nuanced ontological concepts of sentiment provides in-depth sentiment analysis.
  2. Aggregation of sentiments towards a specific entity to identify a general “parent” sentiment provides additional insights into the sentiment towards those entities.
  3. Identification of sentiment towards “parent” entities that do not occur explicitly in the text enables extraction of implicit or alluded sentiment.
  4. Domain-specific and “register” specific analysis of sentiment enhances precision and recall of sentiment analysis.
  5. Cross-linguistic analysis of sentiment based on the ontological (as opposed to the purely lexical) approach allows normalization of sentiment towards entities in different languages.