Week 1 - Intro to NLU and NLU tasks

Question 1

What is NLP

Answer 1

Natural language processing
Converts unstructured data into a structured (posssibly machine-readable) form

Question 2

What is NLU

Answer 2

Natural language understanding
A specification of NLP
Determines the intended meaning of natural language expressions; focuses on the comprehension of human language by machines
makes agents more intelligent

Question 3

What is NLG

Answer 3

Natural language generation
A specification of NLP
produces natural language expressions

Question 4

What are the 4 main NLU task categories

Answer 4

1. Sequence classification
2. Pairwise sequence classification
3. Sequence labelling
4. Span-based operations

Question 5

What are single-problem applications of NLU tasks

Answer 5

Focus on addressing a specific task or problem within natural language understanding

Question 6

What are multi-problem applications of NLU tasks

Answer 6

Involves addressing multiple NLU tasks within a single application or system
Relate complex applications to underlying NLU tasks, by decomposing them into subtasks

Question 7

What is sequence classification and what is its output

Answer 7

Takes a series of words (tokens) (eg sentence, tweet, document)
Output: classification category

Question 8

What is pairwise sequence classification and what is its output

Answer 8

Classify relationship between 2 input sequencies
output: Neutral, contradicts, entails

Question 9

What is sequence labelling and what is its output

Answer 9

Classification at the level of the individual tokens in the sequence
output: eg noun/verb
Can also use subsequence of tokens if the context if relevant
eg john smith = person

Question 10

What is a Unit

Answer 10

A subsequence of tokens

Question 11

What is the BIO scheme

Answer 11

B: beginning of a subsequence of interest

I: Inside the subsequence of interest
- Will also be used for last token in subsequence

O: outside of a subsequence of interest

Question 12

What are span-based operations

Answer 12

Analysis of a span - not necessarily a full sentence nor full document

Takes a sequence, finds all possible spans

Question 13

What is a span

Answer 13

A contiguous sequence of tokens

Question 14

What are the total number of spans found in a sequence with max span = T

Answer 14

Total = (T(T+1)) / 2

Question 15

What are the 3 subtasks of span operations

Answer 15

Identification
Classification
Relation Classification

Question 16

What is Identification: span-based

Answer 16

Identifying spans of interest as binary classification
Input: sequence and question (eg find the keyphrases)
“her best groundstroke is her two-handed backhand”
output: “groundstroke” “two-handed backhand”

Question 17

What is classification: span-based

Answer 17

classifying spans according to a set of labels
Eg “United Airlines” : ORG

Question 18

What is an embedded entity

Answer 18

Eg ORG “United Airlines Holdings”,
ORG “United Airlines”
Here we have an ORG within an ORG

Question 19

Why can span-based classification be better than sequence labeling

Answer 19

It identifies embedded entitieis

Question 20

What is relation classification: span based

Answer 20

Classifying relations between spans
eg Output
EMPLOYEE-OF(”Jane Vickers”, “United Airlines Holdings”)

Question 21

Underlying NLU task: Sentiment Analysis

Answer 21

Identifying overall meaning: -ve, +ve, neutral
Sequence classification

Question 22

Underlying NLU task: Emotion Recognition

Answer 22

Identify the emotions in the input text (eg: sad, lonely)
Sequence classification

Question 23

Underlying NLU task: Hate Speech Detection

Answer 23

Determine whether text contains hate speech
Some also aim to define the type: race/gender/etc
Sequence classification

Question 24

Underlying NLU task: NLI

Answer 24

Hypothesis and Premise
true (entails) or false (contradicts) or neutral
Pairwise sequence classification

Question 25

Underlying NLU task: Paraphrase Identification

Answer 25

Determine whether one is a paraphrase of another (degree of similarity)
eg plagiarism
Pairwise sequence classification

Question 26

Underlying NLU task: NER

Answer 26

Identify subsequence corresponding to categories
Sequence labelling or span based classification

Question 27

Underlying NLU task: Entity Linking

Answer 27

Link subsequence to its standard form in vocabulary
Pairwise sequence classification or span based classification(find span from document)

eg linking a name in wikipedia article to its own wikipedia page

Question 28

Underlying NLU task: Semantic Role Labelling (SRL)

Answer 28

Identify predicate-argument structures: who did what to whom; which bit is linked grammatically (labelled as Pred (verb) or Arg (noun))
Sequence labelling or span-based classification

Question 29

Underlying NLU task: Relation extraction

Answer 29

Identify relation type that holds between two spans
EG Bill was born on April 13th in Seattle
Bill, Seattle: BORN-IN relationship

Span based relation classification

Question 30

Underlying NLU task: Coreference Resolution

Answer 30

Determine if the spans refer to the same real-world entity or concept
EG If “Tom” and “he” refer to the same real-world man

Span based relation classification

Question 31

Multi-underlying NLU task: Aspect-based sentiment Analysis

Answer 31

Identify the target and then aspect category
Eg “Horrible services. The room was dirty and unpleasant.”
Target: Room
Aspects: price, location, comfort, cleanliness → cleanliness
span-based classification(target)
span-based classification(aspect)
span-based relation classification

Question 32

Multi-underlying NLU task: Fact Verification

Answer 32

Determine whether information is supported by facts or not
involves 3 tasks:

1. Claim identification
   determine whether piece of text is worth fact checking
   sequence classification or span based classification
2. Evidence Retrieval
   find (within a pre-existing support corpus) pieces of text which are relevant to a given claim
   pairwise sequence classification
3. Automated Verification
   determine if a piece of text contains information that is supported or refuted by provided pieces of evidence
   pairwise sequence classification

Question 33

Multi-underlying NLU task: Argument mining

Answer 33

Identify argumentative structures
involves 2 tasks:

1. Argument component identificatoin
   identify claims and premises
   span-based classification
2. Argument relation classification
   classify whether a premise supports a claim, i.e., whether the relationship between them is support or oppose
   span-based relation classification

Question 34

Multi-underlying NLU task: Question answering (Extractive)

Answer 34

Given two pieces of text; a passage(context) and a question
To identify the span of text that answers the question

(combination) pairwise, span-based identification
(identifying whether span is of interest in relation to the question)

Question 35

Multi-underlying NLU task: Event extraction

Answer 35

Given a sequence and list of named entities
To identify events, i.e., the event trigger and event participants
has 2 subtasks:

1. Event Trigger Detection
   identify the word that denotes the event and its type
   span-based classification or
   sequence labelling
2. Event Participant Identification (which named entitiesa are actually involved in the event)
   to determine the relationship that holds between a named entity and the event trigger
   span-based relation classification

eg given: Tom was hit by Harry, Tom = person, Harry = person
want to identify event trigger (“hit”) that indicates conflict
and who (tom and harry, attacker and victim) is involved

Question 36

What is the difference between Pairwise sequence classification and span-based relation identification

Answer 36

Pairwise sequence are usually taken from two different sources (the entire source)

span-based is a shortened sample - usually from the same source