Method and apparatus for generating semantic representation model, and storage medium

1. A method for generating a semantic representation model, comprising:

obtaining an original text set comprising original texts;

performing recognition and segmentation on the original text to obtain knowledge units and non-knowledge units in the original text;

performing knowledge unit-level disorder processing on the knowledge units and the non-knowledge units in the original text to obtain a disorder text;

generating a character attribute of each character in the disorder text, the character attribute comprising an original position of the character in the original text and a current position of the character;

generating a training text set based on the character attribute of each character in the disorder text; and

training an initial semantic representation model by employing the training text set to generate the semantic representation model;

wherein training the initial semantic representation model by employing the training text set comprises:

obtaining a present character of the disorder text in the training text set;

determining a first character based on the character attribute of each character in the disorder text, the first character being a character at each position in front of the present position of the present character in the disorder text;

inputting the first character, a character attribute of the first character and the character attribute of the present character into the semantic representation model to obtain a character prediction result of the semantic representation model for the current position of the present character; and

calculating a loss function value and adjusting a coefficient of the semantic representation model based on the current character and the character prediction result, to generate the semantic representation model.

2. The method of claim 1 , wherein performing the recognition and segmentation on the original text comprises:

performing first-level recognition and segmentation on the original text to obtain first-level knowledge units and first-level non-knowledge units; and

performing second-level recognition and segmentation on the first-level knowledge units to obtain second-level knowledge units and second-level non-knowledge units.

3. The method of claim 2 , wherein the first level is a sentence level, and the second level is a word level.

4. The method of claim 1 , after performing the knowledge unit-level disorder processing on the knowledge units and the non-knowledge units in the original text to obtain the disorder text, further comprising:

performing character-level disorder processing on a plurality of characters comprised in the knowledge units.

5. The method of claim 1 , after training the initial semantic representation model by employing the training text set to generate the semantic representation model, further comprising:

determining a task scene of the semantic representation model;

obtaining training data in the task scene, the training data comprising a training text and a task target; and

training the semantic representation model by employing the training text and the task target to generate the semantic representation model in the task scene.

6. An apparatus for generating a semantic representation model, comprising:

one or more processors;

a memory storing instructions executable by the one or more processors;

wherein the one or more processors are configured to:

obtain an original text set comprising original texts;

perform recognition and segmentation on the original text to obtain knowledge units and non-knowledge units in the original text;

perform knowledge unit-level disorder processing on the knowledge units and the non-knowledge units in the original text to obtain a disorder text;

generate a character attribute of each character in the disorder text, the character attribute comprising an original position of the character in the original text and a current position of the character;

generate a training text set based on the character attribute of each character in the disorder text; and

train an initial semantic representation model by employing the training text set to generate the semantic representation model;

wherein the one or more processors are configured to:

obtain a present character of the disorder text in the training text set;

determine a first character based on the character attribute of each character in the disorder text, the first character being a character at each position in front of the current present of the present character in the disorder text;

input the first character, a character attribute of the first character and the character attribute of the present character into the semantic representation model to obtain a character prediction result of the semantic representation model for the present position of the current character; and

calculate a loss function value and adjusting a coefficient of the semantic representation model based on the current character and the character prediction result, to generate the semantic representation model.

7. The apparatus of claim 6 , wherein the one or more processors are configured to:

perform first-level recognition and segmentation on the original text to obtain first-level knowledge units and first-level non-knowledge units; and

perform second-level recognition and segmentation on the first-level knowledge units to obtain second-level knowledge units and second-level non-knowledge units.

8. The apparatus of claim 7 , wherein the first level is a sentence level, and the second level is a word level.

9. The apparatus of claim 6 , wherein the one or more processors are configured to:

perform character-level disorder processing on a plurality of characters comprised in the knowledge units.

10. The apparatus of claim 6 , wherein the one or more processors are configured to:

determine a task scene of the semantic representation model;

obtain training data in the task scene, the training data comprising a training text and a task target; and

train the semantic representation model by employing the training text and the task target to generate the semantic representation model in the task scene.

11. A non-transitory computer readable storage medium having computer instructions stored thereon, wherein the computer instructions are configured to cause a computer to execute a method for generating a semantic representation model, and the method comprises:

obtaining an original text set comprising original texts;

performing recognition and segmentation on the original text to obtain knowledge units and non-knowledge units in the original text;

performing knowledge unit-level disorder processing on the knowledge units and the non-knowledge units in the original text to obtain a disorder text;

generating a character attribute of each character in the disorder text, the character attribute comprising an original position of the character in the original text and a current position of the character;

generating a training text set based on the character attribute of each character in the disorder text; and

training an initial semantic representation model by employing the training text set to generate the semantic representation model;

wherein training the initial semantic representation model by employing the training text set comprises:

obtaining a present character of the disorder text in the training text set;

determining a first character based on the character attribute of each character in the disorder text, the first character being a character at each position in front of the present position of the present character in the disorder text;

inputting the first character, a character attribute of the first character and the character attribute of the present character into the semantic representation model to obtain a character prediction result of the semantic representation model for the current position of the present character; and

calculating a loss function value and adjusting a coefficient of the semantic representation model based on the current character and the character prediction result, to generate the semantic representation model.