Please enable JavaScript.

Coggle requires JavaScript to display documents.

Cognitive Systems - Coggle Diagram

- - - - Focusing on verbs with semantic slot/roles
        
        Slots/Roles - IRS-CG-Day1-Slide50
- - - - Parse input to output, allows for finite states transitions
        
        Rule based tagging
      - IRS-CG-Day1-Slide84
    - - IRS-CG-Day1-Slide87
      - Input > Features (e.g., bag of words) > ML classifier > Intent class > intent-specific-parameter + slot detection
        
        Linear-chain conditional random field (CRF)
        
        Use info from previous labels, gradually increasing in info to lastly have the full info to make prediction
      - Deep learning (same as ML except no features required)
        
        LSTM with CRF
        
        https://medium.com/@utkarsh.kumar2407/named-entity-recognition-using-bidirectional-lstm-crf-9f4942746b3c
      - Intent and slot detection
        
        Intent detection is a classification problem,
        slot detection is Name Entity Recognition needs CRF (with LSTM)
        
        Info Extraction (IE) for Slot detection
        
        Main components of IE system
        
        Co-reference resolution (good to have)
        
        Dialog state tracking (tracking that the date slot is "then", which refers to the date in previous frame
        
        Relationship between entities
        
        e.g., entities with the same referant e.g., Anaphora
        
        Definite descriptions e.g., Usain Bolt and fastest man in the world
        
        Tokenisation (must have)
        
        POS tagger (good to have)
        
        Grammatical tagging of a term
        
        Challenge is POS disambiguation i.e., a book vs to book a room
        
        Rule based (tag all the most frequent tag first, then use contextual rules to correct any errors)
        
        Stochastic taggers using Hidden Markov Models or n-gram
        
        Shallow Parsing (good to have)
        
        Identify phrases such as "About six and a half hours", "Mr Armstrong"
        
        Name Entity (lowest level of recognition)
        
        Concept: Rule to create an abstraction
        
        Information (Words/named entity/phrase that fulfills a need)
        
        Name Entity Recognition
        
        Recognise noun like person, locations etc
        
        Rule based
        
        2 more items...
        
        Statistics based
        
        3 more items...
        
        Type of IE system: Rule based
        
        Hand-coded rules
        
        Machine learning
        
        With an annotated corpus, derive rules
        
        Hybrid
        
        Type of IE system: Statistics-based System
        
        Instead of deriving if-then rules, derive statistical rules to create a model
        
        Advantanges over rule-based - language independent, no linguistics/domain knowledge
        
        Issues: must have a lot of examples, well-annotated corpus
        
        Intent detection
        
        Similarity based intent detection
        
        Compare current input as a vector to past inputs as vector and respond with the associated responses
        
        Cosine similarity
        
        IRS-CG-Day2AM-Slide19
      - Pre-processing
        
        IRS-CG-Day2AM-Slide 8
        
        Tokenisation
        
        Find delimiters (like space, newline) and punctuation marks, then split the sentence
        
        Issues
        
        comma in 12,345
        
        period in Dr.
        
        apostrophe in My friend's or isn't
        
        Case normalisation, change all to lower cast
        
        Lemmization/stemming
        
        Change words to root form e.g., classes to class, ran to run
        
        Punctuation removal
        
        Stopword removal (remove common words like the, a, of)
        
        Vector frequency
        
        tf-idf - IRS-CG-Day2AM-Slide16
  - - - Ranking, find responses that best match the utterance
- - - - IRS-CG-Day2PM-Slide29