Generation of sentence representation
地域：
NY
NY Armonk
摘要
Embodiments of the present disclosure relate to generation of sentence representation. In an embodiment, a method is disclosed. According to the method, a sentence graph is generated from a sentence containing words, the sentence graph comprising nodes representing the words and edges connecting the nodes to indicate relationships between the words. Word representations for the plurality of words are determined based on the sentence graph by applying a graph convolution operation on respective sets of neighbor nodes for respective ones of the nodes, a set of neighbor nodes for a node having edges connected with the node. A sentence representation for the sentence is determined based on the word representations for use in a natural language processing task related to the sentence. In other embodiments, a system and a computer program product are disclosed.
說(shuō)明書(shū)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
where hi(k) is an output at the k-th layer for the i-th node, which may be in form of a feature vector, aij(k) is a weight in the first set specific for the type of the relationship rij at the k-th layer. In the directed sentence graph 512, aij(k) is a weight specific to the relationship from the word vj represented by the j-th node to the word vi represented by the i-th node. The weight aij(k) may be learned during the training process. The outputs for all the nodes at the k-th layer may be written as:
H(k)=MLP(k)(A(k)H(k-1))??Equation (4)
Alternatively, or in addition, the weights may include a second set of weights each determined based on the numbers of nodes having edges connected with respective neighbor nodes in the set of neighbor nodes and the number of nodes in the set of neighbor nodes. In a graph convolution based on the second set of weights (referred to as a second graph convolution), the word representation for the given node may further based on the given node in addition to the weighted summation of the set of neighbor nodes.
The second graph convolution based on the second set of weights performed at each layer 710, . . . , 720 may be represented as follows:
意見(jiàn)反饋
使用該功能遇到問(wèn)題
該功能需要專業(yè)版企業(yè)版VIP權(quán)限，您可以：
您也可以聯(lián)系官方QQ: 2157717237 電話: 13264338900