Semantic search

RELEVANCE FEEDBACK

• RELEVANCE FEEDBACK

• Initially the given query by user is fired
• Some results are retrieved
• Analyze whether or not those results are relevant
• Perform a new query and then produce the final search results by firing this modified query.

• TYPES OF RELEVANCE FEEDBACK

• Explicit Feedback :
• Process of taking Feedback Taken By users for assessing a given output(Set of Documents).
• Eg: After a document is viewed, ask “Was this document helpful?”

• ANALYSIS:

• ADVANTAGE:
• It is able to depict the actual requirement and expectations of the user
• DISADVANTAGE:
• Large fraction of user may not be interested to participate in surveys and Feedbacks.
• These surveys may be biased based on personal choices of users.
• e.g. : When searched about inferno, most of the people may rank the pages of musical band named inferno over that of inferno OS

• IMPLICIT FEEDBACK:

• Feedback which is inferred by the actions of user on output documents.
• Factors:
• Number of times document is visited
• Duration of visit on particular URL
• Depth and number of links from visited

• ANALYSIS:

• ADVANTAGE :
• The interaction time with user is eliminated as the system takes the feedback of the user implicitly.
• DISADVANTAGE:
• Number of Hits on Url: Users may tend to always click on the initial document received. Thus if the search was initially not upto the mark, it may continue performing poor.
• Time Spent on URL: Sometimes the time taken to reject a document may be substantial enough for the algorithm to believe that it is relevant.
• Number and Depth of links visited: This will definitely rank a relevant document as relevant. But this will fail to rank a good document without links as relevant.

• PSEUDO RELEVANCE FEEDBACK OR BLIND FEEDBACK :

• Takes a query as an input.
• From some top k ranked results on that query, some keywords (as per their weights) are selected and augmented to the query which results in further search process.

• ANALYSIS:

• ADVANTAGE :
• It is a completely automated process. Hence totally free from human biasness.
• DISADVANTAGE:
• The efficiency heavily depends on the ranking algorithm used. If the top documents retrieved by the initial query are not very relevant then the final result will also not be very impressive.
• The type of term associations obtained for QE is restricted to co-occurrence based relationships in the feedback documents, and thus other types of term associations such as lexical and semantic relations (morphological variants, synonyms) are not explicitly captured .

• MULTI LINGUAL PRF

•   Given a query in a language, we take the help of another language to ameliorate the well known problems of PRF.
• The steps are:
• Translation: L1 -> L2
• PRF performed in L2.
• Result back-translation: L2 -> L1
• Combination of feedback models of L1,L2. 
• Fetch a new ranked list of documents.   

• ANALYSIS OF MULTILINGUAL PRF

• Good Feedback from Assisting Language: If the feedback model in the assisting language contains good terms, then the back-translation process will introduce the corresponding feedback terms in the source language, thus leading to improved performance.
• Finding Synonyms/Morphological Variations: Another situation in which MultiPRF leads to large improvements is when it finds semantically/lexically related terms to the query terms which the original feedback model was unable to.
• Abundance of documents in the assisting language in the web compared to the base language.

•