Please enable JavaScript.

Coggle requires JavaScript to display documents.

Trio knowledge base (Knowledge embedding (:heavy_plus_sign: is able to…

- - - - the text model
        
        Relational Concurrence Assumption: If two
        words w and v concur in some context, e.g., a window
        of text, then there is a relation rwv between
        the two words. That is, we can state the triplet of
        (w, rwv, v) is a fact.
        
        The goal is to maximize the
        likelihood of the concurrences of pairs of words in
        text windows
        
        text model is almost equivalent to Skip-Gram
      - the alignment model
        
        2 mechanisms
        
        Using names of entities
        
        :heavy_plus_sign: straightforward, does not rely on additional data sources, the number of triplets generated is large and can significantly change the results
        
        :heavy_minus_sign: risky, the name of an entity is ambiguous (different entities may have the same name) => the name graph may contaminate the knowledge embedding
        
        if in a fact triplet head, tail or both exist and belong to the word vocabulary, a new triplet will be generated and added to the graph
        
        Using Wikipedia anchors
        
        based on the connection between Wikipedia
        and Freebase: for most Wikipedia (English) pages,
        there is an unique corresponding entity in Freebase
        
        :heavy_plus_sign: the quality of the data is very high, no ambiguity/completeness issues
        
        :heavy_minus_sign: relies on the additional data source, the number of matched anchors (40M) is relatively small => contribution is limited
        
        aligns entity embedding and word embedding spaces
      - the knowledge model
        
        The goal: to maximize the conditional likelihoods of existing fact triplets in the knowledge graph
        
        Likelihood of observing a fact triplet = the sum of logs of conditional probabilities Pr(h|r,t), Pr(r|h,t) and Pr(t|h,r)
    - - difference
        
        in the text model we assume
        any pair of words h and t that concur in some
        text windows are of certain relation r, but r is a hidden
        variable, and the goal is to fit the concurring
        pairs of words
        
        the knowledge model the relation
        r is explicitly supervised and the goal is to
        fit the fact triplets