Subevent relation extraction

1 . A method comprising:

obtaining a document comprising a first event mention and a second event mention;

generating a dependency tree based on the document;

identifying a dependency path (DP) between the first event mention and the second event mention based on the dependency tree;

computing an alignment function by performing optimal transport between a set of words in the dependency path and a set of words outside the dependency path;

identifying an irrelevant word based on the alignment function;

pruning the dependency tree by removing the irrelevant word to obtain a pruned dependency tree; and

generating subevent relation information for the first event mention and the second event mention based on the pruned dependency tree.

2 . The method of claim 1 , further comprising:

generating a word representation vector for each word of the document, wherein the subevent relation information is generated based on the word representation vector.

3 . The method of claim 1 , wherein generating the dependency tree comprises:

applying a natural language processing (NLP) model to a word representation vector corresponding to each word of the document.

4 . The method of claim 1 , further comprising:

adding edges between consecutive sentences in the document, wherein the dependency tree is based on the added edges.

5 . The method of claim 1 , further comprising:

computing a syntactic importance score for the set of words in the dependency path and the set of words outside the dependency path, wherein the alignment function is based on the syntactic importance score.

6 . The method of claim 1 , further comprising:

computing a base irrelevant word based on the set of words outside the dependency path; and

comparing a plurality of words in the dependency tree to the base irrelevant word to identify the irrelevant word.

7 . The method of claim 1 , wherein:

the first event mention is located in a first sentence of the document and the second event mention is located in a second sentence of the document.

8 . The method of claim 1 , further comprising:

generating a modified representation vector for each word of the pruned dependency tree using a graph convolutional network (GCN), wherein the subevent relation information is generated based on the modified representation vector.

9 . The method of claim 8 , further comprising:

applying a classification network to the modified representation vector to obtain the subevent relation information.

10 . The method of claim 1 , wherein:

the subevent relation information indicates a parent-child relation, a child-parent relation, or a null relation between the first event mention and the second event mention.

11 . The method of claim 1 , further comprising:

receiving a query; and

generating a response to the query based on the subevent relation information.

12 . A method comprising:

obtaining training data including ground-truth relation information for a first event mention and a second event mention in a document;

initializing a subevent relation extraction (SRE) machine learning model;

generating a dependency tree based on the document;

identifying a dependency path (DP) between the first event mention and the second event mention based on the dependency tree;

computing an alignment function by performing optimal transport between a set of words in the dependency path and a set of words outside the dependency path;

identifying an irrelevant word based on the alignment function;

pruning the dependency tree by removing the irrelevant word to obtain a pruned dependency tree;

generating predicted subevent relation information using the initialized SRE machine learning model based on the pruned dependency tree; and

training the SRE machine learning model based on the predicted subevent relation information and the ground-truth relation information.

13 . The method of claim 12 , further comprising:

computing a negative log-likelihood function based on the predicted subevent relation information and the ground-truth relation information, wherein the SRE machine learning model is trained based on the negative log-likelihood function.

14 . The method of claim 12 , wherein training the SRE machine learning model comprises:

jointly training a graph convolutional network (GCN) and a classification network.

15 . An apparatus comprising:

at least one processor;

at least one memory including instructions executable by the at least one processor;

a natural language processing (NLP) model configured to generate a dependency tree based on a document comprising a first event mention and a second event mention;

a pruning component configured to prune the dependency tree by identifying a dependency path (DP) between the first event mention and the second event mention based on the dependency tree, computing an alignment function by performing optimal transport between a set of words in the dependency path and a set of words outside the dependency path, identifying an irrelevant word based on the alignment function, and removing the irrelevant word to obtain a pruned dependency tree; and

a subevent relation extraction (SRE) machine learning model comprising parameters stored in the at least one memory and trained to generate subevent relation information for the first event mention and the second event mention based on the pruned dependency tree.

16 . The apparatus of claim 15 , wherein:

the SRE machine learning model comprises a graph convolutional network (GCN) and a classification network.

17 . The apparatus of claim 16 , wherein:

the classification network comprises a feed-forward network.

18 . The apparatus of claim 15 , further comprising:

a text encoder configured to generate a word representation vector for each word of the document.

19 . The apparatus of claim 18 , wherein:

the text encoder comprises a bi-directional encoder representations from transformers (BERT).