Prediction and Part-Of-Speech Tagging

Question 1

Corpus

Answer 1

A body of text that has been collected for some purpose.

Question 2

Balanced Corpus

Answer 2

Contains texts which represent different genres.

Question 3

Prediction

Answer 3

Given a sequence of words, we want to determine what’s most likely to come next.

Question 4

N-gram Model

Answer 4

A type of Markove Chain where the sequence of the prior n -1 words i sused to predict the next.

Question 5

Trigram

Answer 5

Use preceding two words.

Question 6

Bigram

Answer 6

Models the preceding word

Question 7

Unigram

Answer 7

Use no context at all.

Question 8

Bigrams Model

Answer 8

Assigns a probabilitiy to a word based on the previous word alone.

Question 9

Viterbi Algorithm

Answer 9

A dynamic programming technique for efficiently applying n-grams in speech recognition and other applications to find the highest probability sequence. It is usually descibed in terms of an FSA.

Question 10

Smoothing

Answer 10

To allow for sparse data - we use smoothing. This means that we make some assumption about the probability of unseen or very infrequently seen events and distribute that probability appropriately.

Question 11

Add-one Smoothing

Answer 11

Add one to all counts - not sound theoretically but simple to implement.

Question 12

Backoff

Answer 12

Backing off to lower n-gram probabilities.

Question 13

Part of Speech Tagging

Answer 13

Associating words in a corpus with a tag indication some syntactic information that applies to that particular use of the word. POS tagging makes it easier to extract some types of information.

Question 14

Stochastic POS-tagging

Answer 14

Too complex to make flash card from look at pages 22-24 in the notes.

Question 15

Evaluation of POS tagging

Answer 15

POS tagging algorithms are evaluated in terms of percentage of correct tags. Success rates of 95% are misleading as baseline of choosing most common tag based on training set gives 90% accuracy.

Question 16

Training data and test data

Answer 16

The assumption in NLP is that a system should work on novel data. The test data must therefore be kept unseen.

Question 17

Baselines

Answer 17

Report evaluation with respect to a baseline, which is normally what could be achieved with a very basic approach, given the same training data.

Question 18

Ceiling

Answer 18

Ceiling for performance of an application. This is usually taken to be human performance on that task where ceiling is percentage agreement found between two annotators.

Question 19

Error Analysis

Answer 19

Error rate on a particular program will be distributed very unevenly. Some errors may also be more important than others e.g. treating an incoming order as junk is much worse than converse.

Question 20

Reproducibility

Answer 20

Evaluation should be done on a generally available corpus so that other researches can replicate the experiments.