Closed yuanxiaosc closed 5 years ago
Yes, all the predictions are at the entity level rather than the mention level. As you note, an entity can appear in many sentences, each of these is referred to as a single mention. When we are making predictions about the entities, we are really asking if one or more of the mentions express the relation.
First of all, thank you for your answer. I still have a question and hope to get your answer. In the paper 2.4 Entity Level Predication section, the relation scores formula mentions "For each entity pair $(p^{head}, p^{ptail})$, let $P^{head}$, the set of indices of mentions of the entity $p^{head}$ ... ", how do you determine the indices of these related entities?
The model assumes that we already have the entity linking decisions. By that I mean, we know which mentions are referring to the same entities. In the CDR data, this comes directly from the entity annotations and for the CTD data we use an external entity linking model (taken from pubtator).
Thank you for your patience and this outstanding paper, I can understand the paper.
Although entities may appear in many sentences, there is still only one pair of relationships in each paragraph that needs to be judged?