Systems and methods for semi-supervised relationship extraction
View Patent ↗Systems and methods are disclosed to perform relation extraction in text by applying a convolution strategy to determine a kernel between sentences; applying one or more semi-supervised strategies to the kernel to encode syntactic and semantic information to recover a relational pattern of interest; and applying a classifier to the kernel to identify the relational pattern of interest in the text in response to a query.
1. A method to perform relation extraction in text, comprising:
applying a convolution strategy to determine a kernel between sentences;
deriving an unweighted undirected graph G D (S) for a sentence S from a set of dependency relations supplemented by a linear-order structure, where the set is denoted by D(S) and V(S) is the set of vertices, with each v i εV(S) representing a certain word
G D ( S )=( V ( S ), E ( S ));
determining a single path p from a dependency graph G D (S) composed from a sequence of words and their associated dependencies
p =( w i ,d i,j ,w j , . . . ,w p ,d p,q ,w q )
where word w i and w j are connected by the dependency edge d i,j ;
determining a convolution kernel K G as a sum of kernels on paths (K p ):
K
G
(
G
D
(
S
)
,
G
D
(
S
′
)
)
=
∑
p
∈
P
n
(
G
D
(
S
)
)
∑
p
′
∈
P
″
(
G
D
(
S
′
)
)
K
p
(
p
,
p
′
)
Pr
(
p
G
D
(
S
)
)
Pr
(
p
′
G
D
(
S
′
)
)
where Pr(p|G D (S)) is a probability that single path p happens in the graph G D (S) and calculated as a ratio of path count over sum of path counts;
applying one or more semi-supervised strategies to the kernel to encode syntactic and semantic information to recover a relational pattern of interest; and
applying a classifier to the kernel to identify the relational pattern of interest in the text in response to a query.
2. The method of claim 1 , comprising applying the convolution strategy on one or more sequence kernels to extract relationship.
3. The method of claim 1 , comprising applying a convolution strategy on a graph kernel or a string kernel to extract relationship.
4. The method of claim 1 , wherein one of the semi-supervised strategies relates to add-ons to a word sequence to represent natural English sentences.
5. The method of claim 1 , wherein one of the semi-supervised strategies is added on a graph kernel or a string kernel to consider semantics of natural English words using word embedding representations.
6. The method of claim 1 , wherein one of the semi-supervised strategies are added on to sequence kernels to encode semantic evidence inside English text.
7. The method of claim 1 , comprising generating a graph kernel with semi-supervision added to one of: nodes, edges and whole graphs.
8. The method of claim 1 , comprising generating a graph convolution kernel to extract relationship.
9. The method of claim 1 , comprising applying a string kernel to solve relation extraction.
10. The method of claim 1 , comprising a graph convolution kernel with a sentence representation to simultaneously encode semantic and syntactic information.
11. The method of claim 1 , comprising applying a semi-supervised string kernel in a unified framework for solving relationship extraction at multiple levels.
12. The method of claim 11 , wherein the levels include an article level, a sentence level, and a relation level.
13. The method of claim 1 , wherein the semi-supervised strategies include strategies on word embedding, dependency similarity and pseudo positive sentences.
14. The method of claim 1 , comprising generating a convolution graph kernel to recover relation patterns between entities.
15. The method of claim 1 , comprising generating a convolution graph kernel by determining single paths from graphs and comparing path similarities.
16. The method of claim 1 , comprising
determining single paths by finding short paths first and extending short paths to longer ones;
saving path similarities for subsequent reuse on longer paths that involve corresponding shorter paths.
17. The method of claim 1 , comprising performing semi-supervised learning on sentence level.
18. The method of claim 17 , comprising constructing pseudo training sentences utilizing unlabeled sentences to expand a training set.
19. The method of claim 1 , comprising:
representing dependency relations between words as a graph of sentences;
representing each vertex on the graph with word embedding contextual semantics;
describing each graph edge with one or more semantic categories; and
applying semi-supervised analysis to unlabeled text sets.
20. The method of claim 1 , comprising applying a semi-supervised string kernel to protein sequence classification.
21. The method of claim 20 , comprising training an embedding for each amino acid based on unlabeled protein sequence data set.
22. The method of claim 1 , comprising adding semi-supervised strategy on string kernel.
23. The method of claim 22 , comprising grouping similar words to generate one or more abstract entities (abstraction).
24. The method of claim 23 , wherein the grouping of similar words relies on semi-supervised embeddings of words from unlabeled data.
25. The method of claim 22 , comprising using original text sequence and learned abstracted sequence jointly.