Implicit discourse relationship identification method, system and readable storage medium
A technology of relation recognition and discourse, applied in neural learning methods, instruments, natural language translation, etc., can solve the problems of not effectively using the semantic hierarchy, recognition task information sharing barriers, etc.
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0086] In order to solve the above-mentioned technical problems, the present invention proposes an implicit discourse relationship identification method, please refer to figure 2 and image 3 , for the implicit discourse relationship identification method proposed in the first embodiment of the present invention, the method includes the following steps:
[0087] S101. Receive the global semantic relationship vector and the local semantic relationship vector sent by the encoder, and use the global semantic relationship vector as the zeroth hidden state of the GRU network.
[0088] In the present invention, the decoder takes the global semantic relationship vector output by the encoder and local semantic relation vector As input, a sequence of multilevel implicit discourse relations is generated.
[0089] In this step, the global semantic relation vector As the initial state of the GRU network (the zeroth hidden state). Understandably, in this step, it is equivalent t...
Embodiment 2
[0131] It can be understood that before the decoder receives the global semantic relationship vector sent by the encoder, the encoder needs to encode the input sentence first. In this example, we focus on introducing an encoder based on Bi-LSTM (bidirectional long short-term memory network) and bidirectional attention mechanism.
[0132] In this example, see Figure 4 , the specific coding rules include the following steps:
[0133] S201, calculate the word-pair correlation matrix between the input first sentence and the second sentence, and according to the word-pair correlation matrix, perform normalization processing from two directions of row and column respectively to obtain the first weight matrix and Second weight matrix.
[0134] It should be pointed out here that Bi-LSTM (bidirectional long-short-term memory network) is a neural network structure commonly used to learn the semantic representation of sentences, which can encode contextual information into the vector ...
Embodiment 3
[0156] For the encoding of the input sentence, the third embodiment of the present invention also proposes an encoding method based on a Transformer encoder, and its specific implementation is as follows:
[0157] Firstly, the first sentence and the second sentence in the implicit discourse relation instance are organized into a sequence such as “[CLS]+first sentence+[SEP]+second sentence+[SEP]”. Among them, [CLS] is added as a special mark at the beginning of the first sentence, and it is expected that the global semantic information between the first sentence and the second sentence can be gathered here. [SEP] is used for the split marker between the first sentence and the second sentence.
[0158] To further distinguish the first statement from the second statement, the first statement uses vector identifier, the second statement uses Vector logotype. To take advantage of the word order information in a sentence, position vectors are used identification, where m, n...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More