Word Senses and WordNet

Question 1

What is a word sense?

Answer 1

It is a discrete representation for one meaning of a word.

E.g. a mouse can be a mouse controlling a computer system, or it can be the animal

Question 2

How is a sense defined?

Answer 2

They are defined through its relationship to other senses

Question 3

What is a synonym?

Answer 3

A synonym is two different words that have a near identical meaning

Question 4

What is an antonym?

Answer 4

An antonym is two words that have almost completely opposite meanings

e.g. long and short, big and little, up and down

Question 5

What is a hyponym?

Answer 5

A hyponym is a more specific lemma (subclass)

e.g a car is a vehicle, a dog is an animal, a mango is a fruit

Question 6

What is a hypernym?

Answer 6

It is a more general lemma (superclass or superordinate)

e.g. A vehicle can be a car, an animal can be a dog, a fruit can be a mango

Question 7

What is a meroynm?

Answer 7

It is a lemma which is part of something (part-whole or part of)

e.g. a wheel is part of a car, a leg is part of a chair

Question 8

What is a holonym?

Answer 8

A holonym is a lemma which something can belong to

e.g. a car has wheels, a chair has legs

Question 9

What is WordNet 3.0?

Answer 9

It is a lexical database containing nearly 120,000 nouns, 11,000 verbs, 22,000 adjectives and 4,000 adverbs.

On average, there are 1.23 senses per noun and 2.16 senses per verb

Question 10

What is a synset?

Answer 10

A synset is a set of synonyms

Question 11

What does supersense mean?

Answer 11

Supersense is a semantic category.

There are 26 categories for noun, 15 for verb, 2 for adjective and 1 for adverb

Question 12

What can synsets have?

Answer 12

They can have relationships to each other, as seen in the image

Question 13

What is Word Sense Disambiguation?

Answer 13

It is a way of, given a sentence, selecting the correct word sense for all the words

Question 14

What training data is used with Word Sense Disambiguation?

Answer 14

Datasets with word sense annotations such as SensEval and SemEval that use WordNet annotations

Question 15

What type of problem is word sense disambiguation defined as?

Answer 15

It is defined as a sequence classification problem

Question 16

What is the input and output for the word sense disambiguation problem?

Answer 16

The input is a sentence, a sequence of words

The output is the word sense labels, which are a sequence of tags that are WordNet sense annotations

Question 17

What does the image show?

Answer 17

It shows how Word Sense Disambiguation works in practice. We have an input sentence, and then the classifier chooses the correct word sense based on the context of the rest of the sentence.

Question 18

What baselines are used in WordNet?

Answer 18

One is to take the statistically most frequent sense in the corpus. as we are picking the most frequent one. With this, we can use one sense per discourse (remember the previous sense and use that) - on average this works well but remember, this is just the baseline.

Question 19

What embeddings are used with a WordSense Disambiguation classifier?

Answer 19

Pre-trained contextual BERT embeddings

Question 20

Explain what the image shows

Answer 20

We have our input sentence, which is then encoded using the encoder (BERT embeddings). The decoder then creates word embeddings for each of the words and the word embeddings are aggregated together to form sense embeddings. For any given word, we want to plot where that word falls in the latent space and find the closest sense to that.

Question 21

What happens during the training phase of WSD?

Answer 21

For each word sense we get, we average across the entire corpus each token embedding (context) for the particular word to get a sense embedding.

Question 22

What happens during the test phase of WSD?

Answer 22

When we encounter unseen words, we look up the word embedding for BERT, plot the word embedding in the vector space and use a cosine distance measure to find the closest sense embedding to the new word.

Question 23

What happens when we encounter a sense that is not in the corpus of WordNet?

Answer 23

One option is to choose the statistically most frequent sense in the corpus (not good)

The other option is to impute the sense using WordNet taxonomy (averaging sense embeddings of children to compute embeddings for hypernyms)