Week 6 - Sequence Labelling

Question 1

What is the task of sequence labelling defined as

Answer 1

the ask of
-assigning a label yi to each token xi in an input token sequence X
-the output sequence Y has the same length as X

Question 2

What are the 8 main word classes for POS tagging

Answer 2

Nouns
pronouns
verbs
adjectives
adverbs

determiners
conjunctions
prepositions

Question 3

what are determiners

Answer 3

a, the, an
used to specify nouns

Question 4

what are prepositions

Answer 4

in, of, from
denote spatial information

Question 5

What are word classes defined base on:

Answer 5

(1) Their grammatical relationship with neighbouring words
eg I went for a walk/ I will walk to work

(2) morphological properties (eg of suffixes)
dance (VB), danced (VBD), dancing (VBG)
ie VBD - past tense

Question 6

What are POS tags broadly categorised into

Answer 6

closed class v open class

Question 7

what is a closed class

Answer 7

members are fixed; unlikely that new words are added

Question 8

what is an open class

Answer 8

new words likely to be added/coined over time

Question 9

POS: Task, Input, Output

Answer 9

Task: assign a POS tag to each word in a sequence
Input: sequence x1,…,xn of words and a tagset
Output: sequence y1,..,yn of tags where each tag corresponds to input token
eg Janet/NOUN will/AUX …bill/NOUN

Question 10

Why is POS tagging difficult

Answer 10

words are syntactically ambiguous
“back” -> noun, adjective

Question 11

How do we measure POS accuracy

Answer 11

proportion of POS tags that match gold standard POS tags

Question 12

What is Semantic Role Labelling (SRL)

Answer 12

identifying predicate-argument structures
Answers the question:
“who did what to whom where and when?”

Question 13

What is predicate (SRL)

Answer 13

word(s) expressing the event
ie the what

Question 14

What is argument (SRL)

Answer 14

the participants in the events
ie the who, whom, where, when

Question 15

What is the semantic role (SRL)

Answer 15

the role that each argument (of a predicate) takes

Question 16

What is the Task of SRL

Answer 16

• automatically find the semnatic role of each argument of each predicate (every argument and predicate)
• in principle: the predicate is pre-identified in the input
• in practice: most SRL models detect the predicate

Question 17

What is the Propositional Bank scheme (SRL)

Answer 17

roles are specified to a verb and named with numbers
“the waiter spilled the soup”
ARG0: agent (initiator of an action) “the waitor”
ARG1: patient (entity undergoing the effect of an action) “the soup”
ARG2 and so on: depends on the verb/action

Question 18

What is the FrameNet scheme (SRL)

Answer 18

Frame: a set of related concepts that together comprise background knowledge on some events

Question 19

what are the main roles that exist in the framenet scheme

Answer 19

item, attribute, initial_value, final_value, difference
not every sentence will have all of these elements

Question 20

What is Named Entity Recognition (NER)

Answer 20

named entity: anything that can be referred using a proper name (Person, organisation, location, geo-political entity)
But can also include expressions like dates, times, prices

Question 21

Why is NER hard

Answer 21

ambiguous
Washington - name, organisation, location adn GPE

Question 22

How can NER be applead to sequence labelling

Answer 22

Individual tokens are assigned named entity tags
BIO - beginning, inside, outside
IO - inside, outside
BIOES - beginning, inside, outside, end, single
single means consists of only one token

Question 23

What are Conditional Random Fields (CRFs)

Answer 23

Model that discriminates among all possible tag sequences
Y^ = argmax P(Y|X)
It assigns a probability to an entire sequence Y for every possible sequence in y, given the input sequence X
(selects the highest probability)

Question 24

What is a global feature (CRFs)

Answer 24

Fk
Property of the entire sequences X and Y, which is a sum of local features at each position i in Y

Question 25

What is a local feature (CRFs)

Answer 25

fk
makes use of current output token y, previous output token yi-1, any part of the input sequence X and current position i

Question 26

What is K and wk in CRFs

Answer 26

K = number of features
wk = feature weight

Question 27

What is Z(X) in CRFs

Answer 27

The normalisation factor: to make sure all probabilities equal to 1

Question 28

What is word shape features (CRF)

Answer 28

Abstract letter pattern of a given word
all lowercase letter mapped to ‘x’, all uppercase to ‘X’
all digits mapped to ‘d’, all punctuation retained

Question 29

What is short word shape features (CRF)

Answer 29

like word shape but consecutive character types removed
eg token = I.M.F
word shape = X.X.X, short word shape: X.X.X
eg2 token = DC10-30
word shape = XXdd-dd, short word shape: Xd-d

Question 30

What is affixes feature (CRFs)

Answer 30

prefixes and/or suffixes of size 1 to n

Question 31

What is gazetteer feature (CRFs)

Answer 31

presence of a word in a dictionary of entities (of interest)

Question 32

How are features binarised

Answer 32

every feature eg NNP pos tag is turned into a binary feature
so can be 0 or 1 depending on the token

Question 33

How is BERT finetuned for sequence labelling

Answer 33

On top of BERT, add a classifier (eg single feedforward layer)
-takes as input the output vector for each token
- produces a softmax distribution over all tags
- label with highest probability chosen as output

Question 34

Why is BERT a local approach

Answer 34

Does not take into account dependencies between tags
eg B-PER followed by B-PER
(should be very uncommon to happen)

Question 35

What is ## in BERT

Answer 35

represents a subword token
Crashes - > Crash, ##es

Question 36

How to we solve BERT as a local approach

Answer 36

Add a CRF layer on top of the classifier
- take the softmax output from the classifier
- pass it on to the CRF layer, which takes a global approach: takes into account the label of the previous token