search

UKPLab / starsem2018-entity-linking

Accompanying code for our *SEM 2018 @ NAACL 2018 paper "Mixing Context Granularities for Improved Entity Linking on Question Answering Data across Entity Categories"
Apache License 2.0
59 stars 16 forks source link

question regarding the experiment results in paper ? #8

Open WaNePr opened 4 years ago

WaNePr commented 4 years ago

In WebQSP dataset, the proposed method outperforms all other methods; However, S-MART gives a relatively higher recall (~10%) than the proposed method, Have you thought about why this is the case ?

daniilsorokin commented 4 years ago

Hi,

good question! S-MART does extract much more entities and achieves higher recall at the expense of big drops in precision. It also often does information extraction rather than the strict entity linking, which also produced more noise but also more sort of relevant entities, which again boosts recall. For example, it gives Stan Lee for a Peter Parker mention.

You can take a look at some S-MART output in the original repo: https://github.com/scottyih/STAGG/blob/master/webquestions.examples.train.e2e.top10.filter.tsv