Natural Language Processing

8 flashcards · Computer Science · by humancontext

  1. [POS] Compare different tagging algorithm.

    Rule-based tagging: - costly hand-crafted rules + requires no training data - but does require data to validate rules Transformation-based: - requires tagged data + machine learning - no clear statistical interpretation HMM (to be discussed next time) - requires tagged data + machine (supervised) learning + also: (unsupervised) learning with untagged data

  2. [POS] Two assumptions of HMM

    1. Probability of word depends only on its POS, or: 2. (Markov assumption:) Probability of tag only depends on previous tag, or:

  3. [POS] Describe the procedure of HMM for POS tagging

    • nondeterministically walking through an automaton, • while scanning the sentence left-to-right, • guessing the next tag at each step, and • accumulating the product of many probabilities • taking best choice of all of this

  4. [POS] What is golden standard of tagging

    By gold standard we mean a corpus annotated by hand and carefully checked Tags obtained automatically are compared to the gold standard to evaluate the accuracy. Very simple models achieve about 91% accuracy The state of the art is about 96-97%

  5. [POS] Difference between Hidden Markov Model and Markov Chain

    HMMs can be seen as an extension of Markov chains Like HMM, Markov chain has states and transition probabilities But lacks emission probabilities

  6. [POS] Forward algorithm

    Computes sum of probabilities of all paths for given sequence of observable symbols, Very similar to Viterbi’s algorithm, but using summation instead of maximisation

  7. [POS] What is trellis of forward algorithm

    Introduce one node for the initial state and one for the final state For other states, introduce one node per symbol in w Introduce arcs between nodes labelled by transition and emission probabilities, as sketched below We compute one value for each node Value at initial node is 1 Valueatanyothernodeiscomputedbysummingasequenceofvalues,oneforeachincoming edge For each incoming edge, this value is obtained by multiplying edge value and value of previous node Value at final node is desired probab…

  8. [POS] backward probability