Is there such thing as dataset imrovement?

Question

I know that we can use explained machine learning to find why a model chose a certain classification.

I wonder if there is a way I can find which features are going to improve my current model.

I will explain what I mean by this.

Case: NLP classification of sports, there is a paragraph talking about Ronaldo scores against Uruguay...

Is there a method that can ask which Ronaldo you mean (Ronaldo de Lime the Brazilian player or Cristiano Ronaldo the Portuguese)?

so the model can get a higher accuracy result to classify the paragraph about Brazilian Team or about Portugal Team?

Answer 1

The problem that you describe is the same as word sense disambiguation (WSD): given a known ambiguous word (or ambiguous entity, in your example) with the difference possible senses also known, the model is trained to recognize which sense of the word is used given an input sentence/context.

Note that normally the WSD model is specific to a single word or entity, so one needs a more complex system to solve ambiguity in general.

There is also the task of Word Sense Induction (WSI), which is similar but the senses are not known in advance, so the task is unsupervised and the model has to find the senses (and their number) itself, basically making it a clustering task.