What does the expression Word Sense Disambiguation mean? How does it affect NLP? And also how does the proposed method solve this problem?
"Word Sense Disambiguation" refers to the idea that words can have different meanings in different contexts. Here are some examples
- "I went to the river river bank" vs "I deposited my check at the bank"
- "He's mad good at that game" vs "I am so mad at you"
How it effects NLP, comes down to the way we process text. This generally includes the steps of tokenization and embedding them into some form of vector space. These embeddings in many cases are trained either through some self supervised task on some corpora (examples include Word2Vec or Glove) or doing it from scratch under whichever task/data-set is being used.
Now regarding that paper it does not solve the problem but it does extend a new methodology that assists in helping achieve a better learned generalizable representation for this task. The way I interpreted it is that they don't just use a sense-label (which would be a one-hot encoding / what they call discrete) but instead use a continuous sense representation and do the comparison there. This difference allows for words with similar but different senses to not be equidistant from words with completely different senses.
Word Sense Disambiguation (WSD) is the task of associating meanings or senses (from an existing collection of meanings) with words, given the context of the words. (The word sense is a synonym for meaning.)
For example, consider the noun "tie" in the following two sentences
- He wore a vest and a tie.
- Their record was 3 wins, 6 losses, and one tie.
In these two sentences, the meaning of the word tie is different. In sentence 1, the word tie refers to a necktie, which is a piece of cloth. In sentence 2, the word tie is a synonym for a draw, so it refers to a situation of a game. Therefore, we could associate the meaning (or sense) "neckwear consisting of a long narrow piece of material" to the word tie in the first sentence and the meaning "the finish of a contest in which the winner is undecided" to the same word in the second sentence.
The goal of WSD is thus to predict the appropriate sense or meaning of a word, given the context of the word.
Why is WSD important in NLP? Of course, there are many words that change meaning depending on the context, so WSD is important because you expect NLP algorithms and models to be able to correctly give meanings to words, given their context.