Please enable JavaScript.

Coggle requires JavaScript to display documents.

Compare Questions (Articles (Learning a Concept-based Document Similarity…

- - - - http://scholarworks.umass.edu/cgi/viewcontent.cgi?article=1137&context=cs_faculty_pubs
      - similarity measures
        
        cosine similarity (affected by length og Question)
        
        negative KL divergence between the language models (higher
        negative KL divergence than the other pair does not necessarily
        have stronger semantic connections)
        
        Okapi model
      - IBM Model Algorithm for translation probabilities
  - - - Transform question and use BLEU to score the result question is, to
        
        Get topic
        
        Get type of question
        
        Feature generator
        
        KMean
  - - - RNN +LSTM 86-67 F1 score
  - - - Elasticsearch
        
        Indexing (tokenization, lower case, stemming, n-grams(shingles in Elasticsearch))
  - - - Learning a Concept-based Document
        Similarity Measure
        
        “Concepts” based
        
        less redundant
        
        U.S. and United States
        
        disambiguate terms such as apple and jaguar t
        
        semantic relations between concepts
        
        endangered species may relate to one on environmental
        pollution even though they may have no words in common
        
        25 feature in feature vector
        
        Dataset
        
        HE50
        
        F1 = 0.8
    - - Sentence Similarity Based on Semantic Nets
        and Corpus Statistics
        
        Dataset
        
        f1 = 0.816
        
        semantic+ word order similarity
    - - The Evaluation of Sentence Similarity Measures
        
        Measures
        
        Word Overlap( score based on a number of words shared by two sentences)
        
        Jaccard Similarity Coefficient
        
        Simple Word Overlap and IDF Overlap Measures
        
        Phrasal Overlap Measure
        
        TF-IDF Measures
        
        TF-IDF Vector Similarity
        
        Novelty Detection and Identity Measure
        
        Linguistic Measures
        
        Sentence Semantic Similarity Measures
        
        Word Order Similarity
        
        The Combined Semantic and Syntactic Measures
        
        Datasets
        
        TREC9
        
        0.94
        
        MSRP
        
        0.8
        
        (RTE3)
        
        0.65
    - - Incorporating Word Correlation Knowledge into Topic Modeling
        
        MRF-LDA : coherence measure (%) on NIPS 85.8
    - - topic-based semantic similarity computing approach
        
        Measures
        
        lexical(BoW method)
        
        Jaccard similarity coefficient
        
        document frequency (IDF) overlap
        
        phrase overlap method
        
        no word sense
        
        no word order
        
        no syntactic information
        
        syntactic
        
        calculate the common sub-trees between two questions
        
        no semantic and topic level similarity between two questions
        
        semantic
        
        Models
        
        vector space model
        
        language model and translation model
        
        can bridge the lexical gap between the query and questions
        
        can only capture the lexical level similarity
        
        okapi
        
        approach
        
        Latent Dirichlet Allocation (LDA) model to identifying the question topics
        
        Assumption A.
        Each question can be represented as a topic vector, in which its elements and their values indicate the topic distributions and the importance of the corresponding topics respectively.
        
        topic clustering based approach to finding similar questions
        
        features
        
        lexical information
        
        topic vectors
        
        weight of the topic
        
        dataset
        
        Yahoo! Answers Q&A repository
        
        0.67
    - - more about Jaccard similarity, and plagiarism
    - - Retrieval Models for Question and Answer Archives
        
        translation-based approaches(word-to-word translation probabilities)
        
        IBM
        translation model (performs bad)
    - - https://aclweb.org/anthology/K15-1013
    - - https://github.com/jiyfeng/tfkld
    - - прикольно на .NEt
  - - - Question Classification
      - Relation Detection
      - generate potential answer.
      - Watson uses more than 50 scoring components