0
$\begingroup$

I have a problem that I need to solve. It involves articles about football. I have to determine who is the main protagonist in the article. I already have a solution that I have implemented. Its good enough. But I need to improve it further by using latest NLP solutions.

The current solution is, use coreference resolution to replace the pronouns with their actual coreferents. Then the output article/text is then passed to NER model to get the entities extracted. Then I simply count for either PER or ORG. Then take the entity with the maximum counts.

Any more ideas?

$\endgroup$

1 Answer 1

0
$\begingroup$

Do you have an example of an article where this approach is not good enough? How are you evaluating different approaches?

If you have a bunch of articles where you have labels for who the main protagonist is, you could frame this as a supervised ML problem. You'd want to featurize the articles appropriately, eg. into a vector of the top-K most commonly-occurring candidate entities in each article (each represented as an embedding vector from some pretrained model along with an extra dimension that includes the count of the entity in the article, and perhaps another extra dimension that quantifies where in the article this entity is mostly mentioned). Then this could be a standard multi-class classification task with K-classes (which of the top-K entities is the main protagonist).

$\endgroup$
3
  • $\begingroup$ Yes, I already have a list of labels for each article. I use that to compare the results. The accuracy (TP) is good so far. I have tested on 100 articles. An example of 'not good enough' is here: "Erling Haaland is back... but he looks a little different. Ahead of a first return to the city of his birth, Leeds, the Manchester City striker was pictured with a Viking-inspired hairstyle. Let's hope Leeds are ready for the battle!" Prediction: Leeds Ground truth: ['MANCHESTER_CITY', 'ERLING_HAALAND', 'PREMIER_LEAGUE', 'LEEDS'] $\endgroup$ Jan 10 at 9:48
  • $\begingroup$ I could frame this as supervised problem as I have about 1500 articles. But the catch is that there is a humongous size of labels. There are as many labels as there are football clubs, players, managers etc. $\endgroup$ Jan 10 at 9:50
  • $\begingroup$ Right that's why I suggest you just frame it as a k-way classification where the k options are the top-1,..., top-K most commonly occurring entities in this article. The classifier should learn to mostly predict class 1, except for when it sees certain features. For instance if one of your features is where in the article (near beginning vs near middle/end) the mentions are most concentrated, then the classifier could learn to correctly classify the example you shared where the most-common heuristic doesn't work well $\endgroup$ Jan 10 at 23:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.