Word Sense Disambiguation (WSD)

A sense is one meaning of a word. Its like saying what I sense of this word.

bank  sense 1: a financial institution
bank  sense 2: sloping land beside a river

WSD is the task of choosing the correct sense of a word, with reference to the context around the word.

2 Terms to know

Term	Meaning	Example
Homonymy	Two unrelated meanings share a spelling	bank (money), bank (river)
Polysemy	Several related meanings of one word	mouth of a person, mouth of a river
Gloss	Simply a short definition of a sense	-

3 Approaches to WSD

Approach	Needs labelled data?	Resource used	How it decides
Dictionary-based (also called knowledge-based)	No	Dictionary or WordNet glosses	Counts overlapping words between the gloss and the context
Supervised	Yes, a sense-annotated corpus	SemCor, Senseval, SemEval	Trains a classifier for each ambiguous word
Unsupervised (word sense induction)	No	Raw text only	Clusters similar contexts, then treats each cluster as a sense
Semi-supervised	A few seed examples	Seeds plus raw text	Bootstraps from the seeds using heuristics

The Lesk Algorithm

This is a dictionary based method that Michael Lesk published in 1986. The main idea is to choose a sense whose gloss shares the most words with the context.

There are some variants of Lesk:

• Simplified Lesk compares the gloss against the words in the sentence
• Original Lesk compares the glosses of 2 ambiguous words against each other
• Extended Lesk adds the glosses of related synsets, such as hypernyms and examples, to make the gloss longer

1. Steps

1. Take the ambiguous word and the sentence around it
2. Remove stopwords from the sentence to get the context set
3. Look up every sense of the word in WordNet
4. For each sense, take its gloss, and count how many words it shares with the context set. This count is the overlap score
5. Return the sense with the highest overlap

2. Worked Example of Simplified Lesk

1. Given: I went to the bank to deposit my money.
2. Context set after stopword removal: went, deposit, `money
3. Look up every sense of this word bank

Sense	Gloss (Meaning)	Shared words	Score
bank #1 financial institution	"a financial institution that accepts deposits and channels the money into lending activities"	deposit, money	2
bank #2 sloping land	"sloping land beside a body of water"	none	0

In this case, sense #1 is chosen because of a higher matching rate.

3. Weakness of Lesk:

• glosses are short, so the overlap is often zero for every sense
• the result depends on the exact wording of the dictionary
• accuracy sits around 50-60%

Code

# ---------- 1. setup ----------
import nltk
nltk.download('wordnet')
nltk.download('omw-1.4')
nltk.download('punkt_tab')
nltk.download('stopwords')
 
from nltk.corpus import wordnet as wn
from nltk.tokenize import word_tokenize
from nltk.wsd import lesk
 
# ---------- 2. list the senses of a word ----------
senses = wn.synsets('bank')
print(len(senses), senses)
 
# ---------- 3. read one sense ----------
s = wn.synset('bank.n.01')
print(s.name(), s.pos())
print(s.definition())
print(s.examples())
 
# ---------- 4. lemmas: synonyms and antonyms ----------
print(s.lemma_names())
print(wn.synset('good.a.01').lemmas()[0].antonyms())
 
# ---------- 5. the hierarchy ----------
dog = wn.synset('dog.n.01')
print(dog.hypernyms())
print(dog.hyponyms())
 
# ---------- 6. run WSD ----------
sent = word_tokenize("I went to the bank to deposit my money")
sense = lesk(sent, 'bank', 'n')
print(sense, '|', sense.definition())
 
# ---------- 7. Lesk written by hand ----------
def my_lesk(context_sentence, word, pos=None):
    context = set(context_sentence)
    synsets = wn.synsets(word, pos) if pos else wn.synsets(word)
    if not synsets:
        return None
    best, best_score = None, -1
    for ss in synsets:
        score = len(context.intersection(ss.definition().split()))
        if score > best_score:
            best, best_score = ss, score
    return best

Lesk needs the word class before it can compare glosses, so it runs after Part-of-Speech (POS) Tagging.