CN115080693A - Text processing methods, electronic devices, dialogue systems and automobiles - Google Patents

Abstract

The invention relates to a text processing method, electronic equipment, a dialogue system and an automobile. The method according to one embodiment of the invention comprises the following steps: performing word embedding on a text to obtain a word embedding vector; obtaining a task identifier; inputting the word embedding vector and the task identifier into a trained conversion mark model to obtain a task type and a conversion mark, the conversion mark comprising a first conversion mark representing a part needing to be converted in the text and a second conversion mark representing a part not needing to be converted in the text; fusing the task type and the text part corresponding to the first conversion mark to obtain fusion information; and inputting the fusion information into a trained text conversion model, and performing corresponding conversion on the to-be-converted text part according to the task type to obtain a conversion result. According to the invention, the accuracy of two tasks of text standardization and inverse text standardization and inverse text standardization can be synchronously improved.

Description

Text processing methods, electronic devices, dialogue systems and automobiles

technical field

The invention relates to the field of natural language processing, in particular to a text processing method, an electronic device, a dialogue system and a car.

Background technique

The intelligent data cockpit dialogue system in autonomous driving includes modules such as speech recognition, natural language processing, and speech synthesis. The text used by these modules often needs to undergo some processing to facilitate modeling and improve legibility. Among them, text normalization (TN) and Inverse Text Normalization (ITN) plays an important role. The function of text standardization is to realize the conversion from the handwritten format of the text to the spoken language format, such as converting the Arabic numeral information in a sentence of Chinese into Chinese character information, which is mainly used in the preprocessing process of speech synthesis. Inverse text standardization is its inverse process. After the recognized text is obtained, in order to facilitate the user to read, the Chinese character information is converted into a form that is easier for the user to read. It is mainly used in the processing stage after speech recognition. For example, after speech recognition The text of the Chinese character is converted into the corresponding digital information through the mapping rule comparison table.

There are relatively mature solutions for text standardization and inverse text standardization, but their goal is only to solve one of these two tasks, and it is impossible to solve both tasks at the same time, which increases the overall complexity of the dialogue system and the system's performance. maintenance costs.

SUMMARY OF THE INVENTION

In order to solve the above technical problems, the present invention proposes a text processing method, electronic device, dialogue system, storage medium and automobile, so as to realize the integrated processing of inverse text standardization and text standardization for text information.

In a first aspect, the present invention provides a text processing method, comprising:

Embedding the text to get the word embedding vector;

get task identifier;

Inputting the word embedding vector and the task identifier into a trained conversion tagging model to obtain a task type and conversion tag, wherein the conversion tag includes a first conversion tag that characterizes the part of the text that needs to be converted and a conversion tag that characterizes the A second conversion token for the part of the text that does not require conversion;

Fusing the task type and the text portion corresponding to the first conversion mark to obtain fusion information;

Inputting the fusion information into a trained text conversion model, and correspondingly converting the to-be-converted text portion according to the task type to obtain a conversion result.

In a specific embodiment, the obtaining the task identifier includes:

Obtaining the scene to be applied to the text, wherein the scene to be applied includes a first scene corresponding to text standardization and a second scene corresponding to inverse text standardization;

The task identifier is determined according to the scenario to be applied, wherein the task identifier includes a first identifier corresponding to the first scenario and a second identifier corresponding to the second scenario.

In a specific embodiment, inputting the word embedding vector and the task identifier into a trained conversion tagging model to obtain the task type and conversion tag includes:

Inputting the word embedding vector and the task identifier into a trained conversion token model to obtain a token sequence, wherein the token sequence includes a task type bit and a transition token bit;

Wherein, the task type on the task type bit includes a text normalization task corresponding to the first identifier and an inverse text normalization task corresponding to the second identifier;

Wherein, the conversion flag on the conversion flag bit includes the first conversion flag and the second conversion flag.

In a specific embodiment, the task type and the text part corresponding to the first conversion mark are fused to obtain fusion information, including:

The task type, the text portion corresponding to the first conversion mark, and the context of the text corresponding to the first conversion mark are fused to obtain fusion information.

The fusion of the task type, the text portion corresponding to the first conversion mark, and the context of the text corresponding to the first conversion mark includes:

The task type, the text portion corresponding to the first conversion mark, and the context of the text corresponding to the first conversion mark are spliced together.

In a specific embodiment, inputting the fusion information into a trained text conversion model, and performing corresponding conversion on the part of the text to be converted according to the task type to obtain a conversion result, including:

If the task type is a text normalization task, the text conversion model performs text normalization processing on the to-be-converted text portion to obtain a text normalization conversion result;

If the task type is an inverse text normalization task, the text conversion model performs inverse text normalization processing on the to-be-converted text portion to obtain an inverse text normalization conversion result.

In a specific embodiment, the method further includes:

The complete text is obtained by splicing the conversion result with the part of the text that does not need to be converted.

In a specific embodiment, the trained conversion labeling model is composed of M layers of bidirectional LSTM networks; the trained text conversion model is composed of N Transformer networks, where M and N are natural numbers.

In a specific embodiment, the method further includes the step of training a Norm network composed of the conversion labeling model and the text conversion model.

In a specific embodiment,

进行训练，其中：Use the loss function as

to train, where:

The output conditional probability of the Norm network is:

in,

Norm(.) represents the output conditional probability, sl represents the I-th element of the sequence, θ represents the parameters of the network, and D represents the number of input samples per batch in training.

In a second aspect, the present invention provides an electronic device comprising a processor and a memory, the memory having a program stored therein, the program implementing the method according to the first aspect when executed by the processor.

In a third aspect, the present invention provides a storage medium, in which a program is stored in the memory, and when the program is executed by the processor, the method according to the first aspect is implemented.

In a fourth aspect, the present invention provides an in-vehicle dialogue system including the electronic device according to the second aspect and an input and output device.

In a fifth aspect, the present invention provides an automobile including the vehicle-mounted dialogue system according to the fourth aspect.

The above-mentioned one or more technical solutions of the present invention have at least one or more of the following beneficial effects:

The present invention introduces the task identifier when processing the text, obtains the task type and the text part that needs to be converted obtained according to the task type by converting the mark model, and fuses them as the input of the text conversion model. The text conversion model is based on the task type. The corresponding text conversion can support both text standardization and inverse text standardization, which not only simplifies the complexity of the overall intelligent dialogue system, but also can achieve accurate text standardization and inverse text standardization due to the use of the same text conversion model. Synchronized increase.

Description of drawings

The disclosure of the present invention will become more easily understood with reference to the accompanying drawings. It can be easily understood by those skilled in the art that these drawings are only for the purpose of illustration, and are not intended to limit the protection scope of the present invention. In addition, like numerals in the figures are used to designate like parts, where:

1 is a schematic flowchart of a text processing method according to an embodiment of the present invention;

2 is a schematic structural diagram of a conversion markup model according to an embodiment of the present invention;

3 is a schematic structural diagram of a text conversion model according to an embodiment of the present invention;

FIG. 4 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.

Detailed ways

Some embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principle of the present invention, and are not intended to limit the protection scope of the present invention.

In the description of the present invention, "module" and "processor" may include hardware, software or a combination of both. A module may include hardware circuits, various suitable sensors, communication ports, memory, and may also include software parts, such as program codes, or a combination of software and hardware. The processor may be a central processing unit, a microprocessor, an image processor, a digital signal processor, or any other suitable processor. The processor has data and/or signal processing functions. The processor may be implemented in software, hardware, or a combination of the two. Non-transitory computer-readable storage media include any suitable media that can store program code, such as magnetic disks, hard disks, optical disks, flash memory, read-only memory, random-access memory, and the like.

Referring to FIG. 1, FIG. 1 is a schematic flowchart of a text processing method according to an embodiment of the present invention, including the following steps:

S10: Perform word embedding on the text to obtain a word embedding vector.

Perform word embedding on the text to obtain an embedding vector X=(x ₁ , . . . , x _L ), where L represents the sentence length. In a specific embodiment, for example, the text obtained from the speech recognition module of the intelligent cockpit dialogue system is "call 1385526696", then L=15. Further, assuming that the embedding dimension n is 100, each xi (1≤i≤L) in the embedding vector is 100-dimensional, and finally a 17*100-dimensional matrix is formed. In another specific embodiment, for example, the text obtained from the speech recognition module of the intelligent cockpit dialogue system is "this journey is about 215.6km", then L=12.

Of course, those skilled in the art can understand that the text used in the present invention is not limited to being obtained from the intelligent cockpit dialogue system, which is just an example.

S20: Obtain the task identifier.

In a specific embodiment, for example, a task identifier T of 0 represents a text normalization task, and a task identifier T of 1 represents an inverse text normalization task.

Those skilled in the art can understand that, in an alternative solution, 0 can be used to represent the inverse text normalization task, and 1 can be used to represent the text normalization task, which is not limited in the present invention.

In a specific embodiment, step S20 includes the following steps:

Obtaining the scene to be applied to the text, wherein the scene to be applied includes a first scene corresponding to text standardization and a second scene corresponding to inverse text standardization;

Specifically, if the text processing using the method of the present invention is used for speech recognition post-processing scenarios, for example, for converting text and presenting it on the human-machine interface of intelligent driving, the scenario corresponds to inverse text standardization, such as the above example The text "call 1385526696" in the text is converted to "call 13855026696" and is displayed on the human-computer interface for the convenience of users to read. For another example, if the text processing using the method of the present invention is used for a speech synthesis preprocessing scenario, that is, the speech is broadcast after the text is converted, then the text corresponding to the scenario is standardized, for example, the text in the above example "This trip is about 215.6km", the conversion Afterwards, "this trip is about 215.6 kilometers" to facilitate conversion into voice for broadcast.

The task identifier is determined according to the scenario to be applied, wherein the task identifier includes a first identifier corresponding to the first scenario and a second identifier corresponding to the second scenario.

In the above two examples, the task identifier corresponding to the first scenario is 1, and the task identifier corresponding to the second scenario is 0.

S30: Input the word embedding vector and the task identifier into a trained conversion tagging model to obtain a task type task and a conversion tag sig, wherein the conversion tag includes a first conversion tag representing the part of the text that needs to be converted and a second conversion token characterizing the portion of the text that does not require conversion.

In a specific embodiment, in order to facilitate the operation, the task identifier is converted into a vector of the same dimension of each word in the word embedding vector, for example, 100 dimensions in the above example.

The dimension-aligned task identifiers and embedding vectors X ₌ (x ₁ , . For example, in the above example where the text is "call 1385526696", the task identifiers of the embedding vector X=(x1,...,x17) and T=1 are formed into a matrix and input to the trained Convert the markup model.

In a specific embodiment, the conversion markup model adopts the LSTM model commonly used in the field of natural language processing. LSTM models are particularly suitable for handling categorical data where the input sequence is a temporal problem. More preferably, the present invention adopts the bidirectional LSTM model as shown in Fig. 2, and the bidirectional LSTM models trained on the input sequence are two LSTMs instead of one. The first in the input sequence is the original sample and the second is the reversed sample of the input sequence. As can be seen from the figure, the forward layer and the reverse layer are jointly connected to the output layer, which contains 6 shared weights w1-w6, which can provide additional context for the network and can learn the problem faster and more comprehensively. More preferably, in order to learn the problem more comprehensively, the transformation labeling model of the present invention can adopt M-layer (eg, 2-layer) bidirectional LSTM network, and the output layer of the bidirectional LSTM network of the first layer is connected to the output layer of the bidirectional LSTM network of the second layer. input, and so on.

In a specific embodiment, the transition token model outputs a token sequence, wherein the token sequence includes task type bits and transition token bits. Wherein, the task type on the task type bit includes a text normalization task corresponding to the first identifier and an inverse text normalization task corresponding to the second identifier; the conversion mark on the conversion mark bit The first conversion mark and the second conversion mark are included.

For example, for the example with the text "Call 1385526696" and T=1, the token sequence is obtained: "ITNSSSSBEEEEEEEEEE", where the start of the token sequence is the task type bit, which in this example is the corresponding For T=1 task type ITN. However, those skilled in the art can understand that whether to locate the task type bit at the beginning of the marker sequence or at other positions can be set as required, which is not limited in the present invention. For the ITN task, the part of the text that needs to be converted is "1385526696", the conversion marker is B as the start marker, E as the end marker, and the middle part is represented by E. However, those skilled in the art can understand that other letters or forms can be used to represent the conversion mark, which is not limited in the present invention. On the other hand, for the ITN task, the part of the text that does not need to be converted is "make a call", and the conversion mark is represented by S. However, those skilled in the art can express the conversion mark in other letters or forms with other solutions, which is not limited in the present invention.

In another example, such as for the example where the text is "This trip is about 215.6km" and T=0, a sequence of tokens is obtained: "TNSSSSSBEEEEEE", where the sequence of tokens begins with a task type bit, which in this example corresponds to T A task type TN of =0. However, those skilled in the art can understand that whether to locate the task type bit at the beginning of the marker sequence or other positions can be set as required, which is not limited in the present invention. For the TN task, the part of the text that needs to be converted is "215.6km", the conversion marker is B as the start marker, E as the end marker, and the middle part is represented by E. However, those skilled in the art can understand that other letters or forms can be used to represent the conversion mark, which is not limited in the present invention. On the other hand, for the TN task, the part of the text that does not need to be converted is "this trip is about", and the conversion mark is denoted by S. However, those skilled in the art can understand that other letters or forms can be used to represent the conversion mark, which is not limited in the present invention.

Those skilled in the art can understand that for the text normalization task and the inverse text normalization task, which parts of the text need to be converted can be preset as rules, and the method of the present invention identifies according to the rules.

S40: Fusion of the task type and the text part corresponding to the first conversion mark to obtain fusion information.

According to the first conversion mark, the corresponding text part is extracted and fused with the task type. For example, for the example where the text is "call 1385526696" and T=1, the vector corresponding to the text part "1385526696" that needs to be converted needs to be summed The vectors corresponding to the ITN are fused (eg, concatenated).

Preferably, in order to achieve a better conversion effect, the present invention also concatenates the task type, the text part to be converted, and the context of the text part to be converted. In a specific example, for example, n words before and after the text part to be converted are fused together as context, and n can be set as required, for example, 2. For example, for the text "Call 1385526696", as the text part "1385526696" that needs to be converted, the 2 preceding texts are "Telephone", and then If there is no text, only the preceding text is taken as the context in this example.

S50: Input the fusion information into a trained text conversion model, and perform corresponding conversion on the text part to be converted according to the task type to obtain a conversion result.

If the task type is a text normalization task, the text conversion model performs text normalization processing on the to-be-converted text portion to obtain a text normalization conversion result.

If the task type is an inverse text normalization task, the text conversion model performs inverse text normalization processing on the to-be-converted text portion to obtain an inverse text normalization conversion result.

For example, for the example with the text "Call 13855206696" and T=1, "13855206696" is output after conversion. For another example, for the example in which the text is "this trip is about 215.6 km" and T=0, after conversion, "215.6 km" is output.

In a specific embodiment, the Transformer network adopted by the text conversion model of the present invention, as shown in FIG. 3 , the Transformer network is composed of an encoder (Encoder)-decoder (Decoder), and each block includes an attention mechanism. More preferably, in order to improve the effect of transformation, multiple Transformer networks can be used, and the output of the previous network is used as the input of the latter network.

In a specific embodiment, for final application, a complete text is obtained by final splicing the conversion result and the part that does not need to be converted in the text, such as "call 13855026696", "this trip is about 215.6" km".

Those skilled in the art can understand that in order to apply the above-mentioned conversion labeling model and text conversion model, training needs to be performed. In a specific embodiment of the present invention, the conversion labeling model and the text conversion model constitute a Norm network, which is trained using training samples.

进行训练，In a specific embodiment, training uses the following loss function as

to train,

The output conditional probability of the Norm network is:

in,

Norm(.) represents the output conditional probability, sl represents the I-th element of the sequence, θ represents the parameters of the network, and D represents the number of input samples per batch in training.

For example, the total training samples are 1000 texts, the number of input samples in each batch is 100, and the model is trained in 10 inputs.

Specifically, a backpropagation method can be used, and the training ends when the loss function reaches a predetermined threshold. After that, test with test samples.

The present invention introduces the task identifier when processing the text, obtains the task type and the text part that needs to be converted obtained according to the task type by converting the mark model, and fuses them as the input of the text conversion model. The text conversion model is based on the task type. The corresponding text conversion can support both text standardization and inverse text standardization, which not only simplifies the complexity of the overall intelligent dialogue system, but also can achieve accurate text standardization and inverse text standardization due to the use of the same text conversion model. Synchronized increase.

It should be pointed out that, although the steps in the above embodiments are described in a specific sequence, those skilled in the art can understand that in order to achieve the effect of the present invention, different steps do not necessarily need to be executed in such an order. It may be performed simultaneously (in parallel) or in other sequences, and these variations are within the scope of the present invention.

In particular, the processes described above with reference to the flowcharts may be implemented as computer software programs according to embodiments of the present invention. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer readable medium, the computer program containing program code for performing the method illustrated in the flowchart. In such an embodiment, the computer program may be downloaded and installed from a server, or installed from a storage device. When the computer program is executed, the above-mentioned functions defined in the methods of the embodiments of the present invention are realized.

The text processing method according to an embodiment of the present invention can be applied to the above-mentioned dialogue system of the intelligent data cockpit, for example, as a function of a natural language processing module.

To this end, the present invention also provides an electronic device, as shown in FIG. 4 , which may include a processor (such as a central processing unit, a graphics processor, etc.) and a storage medium. A program is stored in the storage medium, and when the program is executed by the processor, implements the text processing method according to an embodiment of the present invention.

Wherein, the program includes computer program code, and the program code may be in the form of source code, object code, executable file or some intermediate form, and the like.

Further, when the program is executed by the processor, the functions of speech recognition and speech synthesis can also be realized, that is to say, the electronic device includes software function modules such as a speech recognition module, a natural language processing module and a speech synthesis module.

To this end, the present invention also provides an in-vehicle dialogue system, including the above electronic equipment and an input/output device, and the processor and the storage medium can be connected to each other and to the input/output device through a bus.

In a specific embodiment, the input/output device of the in-vehicle dialogue system of the present invention includes a touch screen, a microphone, a speaker, etc., which are used to form the hardware and software required for the intelligent data cockpit dialogue system in automatic driving.

The present invention also provides a computer-readable storage medium, in which a program is stored, and when the program is executed by a processor, the text processing method according to an embodiment of the present invention is implemented.

It should be noted that the above-mentioned computer-readable storage medium of the present invention may be a computer-readable signal medium or a computer-readable storage medium, or any combination of the above two. The computer readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared or semiconductor system, apparatus or device, or a combination of any of the above. More specific examples of computer readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable Programmable read only memory (EPROM or flash memory), fiber optics, portable compact disk read only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. In the present invention, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, however, a computer readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer readable program code therein. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium that can transmit, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device . Program code embodied on a computer readable medium may be transmitted using any suitable medium including, but not limited to, electrical wire, optical fiber cable, RF (radio frequency), etc., or any suitable combination of the foregoing.

Further, it should be understood that since the setting of each module is only for describing the functional units of the apparatus of the present invention, the physical device corresponding to these modules may be the processor itself, or a part of software in the processor, a part of hardware, or Part of the combination of software and hardware. Therefore, the numbers of the various modules in the figures are merely schematic.

So far, the technical solutions of the present invention have been described with reference to the preferred embodiments shown in the accompanying drawings, however, those skilled in the art can easily understand that the protection scope of the present invention is obviously not limited to these specific embodiments. On the premise of not departing from the principle of the present invention, those skilled in the art can make equivalent changes or replacements to the relevant technical features, and the technical solutions after these changes or replacements will fall within the protection scope of the present invention.

Claims (10)

1. a text processing method, is characterized in that, comprises: Embedding the text to get the word embedding vector; get task identifier; Inputting the word embedding vector and the task identifier into a trained conversion tagging model to obtain a task type and conversion tag, wherein the conversion tag includes a first conversion tag that characterizes the part of the text that needs to be converted and a conversion tag that characterizes the A second conversion token for the part of the text that does not require conversion; Fusing the task type and the text portion corresponding to the first conversion mark to obtain fusion information; Inputting the fusion information into a trained text conversion model, and correspondingly converting the to-be-converted text portion according to the task type to obtain a conversion result. 2. The method according to claim 1, wherein the obtaining the task identifier comprises: Obtaining the scene to be applied to the text, wherein the scene to be applied includes a first scene corresponding to text standardization and a second scene corresponding to inverse text standardization; The task identifier is determined according to the scenario to be applied, wherein the task identifier includes a first identifier corresponding to the first scenario and a second identifier corresponding to the second scenario. 3. The method according to claim 2, wherein, inputting the word embedding vector and the task identifier into a trained conversion labeling model to obtain a task type and a conversion label, comprising: Inputting the word embedding vector and the task identifier into a trained conversion token model to obtain a token sequence, wherein the token sequence includes a task type bit and a transition token bit; wherein the task type on the task type bit includes a text normalization task corresponding to the first identifier and an inverse text normalization task corresponding to the second identifier; and Wherein, the conversion flag on the conversion flag bit includes the first conversion flag and the second conversion flag. 4. The method according to claim 1, wherein the merging of the text portion corresponding to the task type and the first conversion mark to obtain fusion information, comprising: The task type, the text portion corresponding to the first conversion mark, and the context of the text corresponding to the first conversion mark are fused to obtain fusion information. 5. The method according to claim 4, wherein the merging the task type, the text part corresponding to the first conversion mark, and the context of the text corresponding to the first conversion mark, comprises: The task type, the text portion corresponding to the first conversion mark, and the context of the text corresponding to the first conversion mark are spliced together. 6. The method according to claim 1, characterized in that, inputting the fusion information into a trained text conversion model, and performing corresponding conversion on the part of the text to be converted according to the task type to obtain a conversion result, comprising: : If the task type is a text normalization task, the text conversion model performs text normalization processing on the to-be-converted text portion to obtain a text normalization conversion result; If the task type is an inverse text normalization task, the text conversion model performs inverse text normalization processing on the to-be-converted text portion to obtain an inverse text normalization conversion result. 7. The method according to claim 6, wherein the method further comprises: The complete text is obtained by splicing the conversion result with the part of the text that does not need to be converted. 8. An electronic device comprising a processor and a memory, wherein a program is stored in the memory, and when the program is executed by the processor, the method according to any one of claims 1-7 is implemented . 9. A vehicle-mounted dialogue system, characterized in that, comprising: The electronic device according to claim 8; Input and output device. 10. An automobile, characterized by comprising the vehicle-mounted dialogue system according to claim 9.

Images