WO2021082570A1 - 基于人工智能的语义识别方法、装置和语义识别设备 - Google Patents

基于人工智能的语义识别方法、装置和语义识别设备 Download PDF

Info

Publication number
WO2021082570A1
WO2021082570A1 PCT/CN2020/105908 CN2020105908W WO2021082570A1 WO 2021082570 A1 WO2021082570 A1 WO 2021082570A1 CN 2020105908 W CN2020105908 W CN 2020105908W WO 2021082570 A1 WO2021082570 A1 WO 2021082570A1
Authority
WO
WIPO (PCT)
Prior art keywords
corpus
negative
training
value
negative corpus
Prior art date
Application number
PCT/CN2020/105908
Other languages
English (en)
French (fr)
Inventor
张晴
刘畅
杨瑞东
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Priority to EP20881934.2A priority Critical patent/EP4030335A4/en
Priority to US17/771,577 priority patent/US20220414340A1/en
Publication of WO2021082570A1 publication Critical patent/WO2021082570A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis
    • G06F40/35—Discourse or dialogue representation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/332—Query formulation
    • G06F16/3329—Natural language query formulation or dialogue systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3343—Query execution using phonetics
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3344—Query execution using natural language analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • G06F16/5846—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using extracted text

Definitions

  • the embodiments of the present application relate to the technical field of semantic recognition in artificial intelligence, and particularly relate to artificial intelligence-based semantic recognition methods, devices, and semantic recognition equipment.
  • the human-machine dialogue system is a new generation of human-computer interaction interface.
  • the Bot platform is a model training platform.
  • the Bot platform provides developers with a platform for rapid capacity building, supporting such as three-party business parties to quickly build dialogue skills for their Interactive triggering of business functions.
  • the Bot platform provides developers with a one-click trigger button to automatically train the skill model configured by the developer.
  • the retrained model may be inconsistent with the model obtained by the previous training, resulting in large fluctuations in confidence, which can be expressed as the last recognizable corpus cannot be recognized under the retrained model, or
  • the corpus that cannot be recognized by the model obtained by previous training can be recognized on the model obtained by retraining.
  • the carefully adjusted corpus, accuracy and/or recall rate and other indicators will fluctuate greatly and do not meet expectations.
  • the instability caused by this retraining model will eventually affect the developer's experience.
  • the embodiment of the application provides a semantic recognition method, device, and semantic recognition device based on artificial intelligence.
  • the embodiment of the application also provides a computer-readable storage medium to realize that the training corpus is not added, deleted, or modified. , It can keep the model obtained by two or multiple trainings without obvious difference, so that the test corpus of the developer has almost the same confidence in the model obtained by multiple training (difference ⁇ 0.01), thereby reducing the fluctuation of the accuracy rate. Thereby improving the developer experience.
  • an embodiment of the present application provides a semantic recognition method based on artificial intelligence, including: obtaining a query sentence input by a user;
  • the query sentence is recognized through a pre-trained semantic recognition model to obtain the intention of the query sentence; wherein the pre-trained semantic recognition model is trained using training corpus and negative corpus, and the negative corpus is The coding value of the training corpus is mapped to the negative corpus and extracted; wherein, the above training corpus is configured by the developer on the model training platform, for example: the Bot platform, and the above negative corpus is provided by the above model training platform of;
  • the response corresponding to the query sentence is displayed.
  • the negative corpus is extracted through the mapping relationship according to the coding value of the training corpus.
  • the training corpus is not added, deleted or modified, the coding value of the training corpus remains unchanged, and the mapping relationship is not Will change, so the negative corpus extracted through the mapping relationship according to the coding value of the training corpus will not change.
  • the training corpus and the extracted negative corpus remain unchanged, the stability of the model obtained by training with the training corpus and the negative corpus Higher, it can be achieved that in the case of no addition, deletion or modification of the training corpus, the model obtained by two or multiple trainings has no obvious difference, so that the test corpus of the developer has almost the same in the model obtained by multiple training Confidence (difference ⁇ 0.01), which in turn makes the accuracy fluctuation smaller, thereby improving the developer's experience.
  • the obtaining the query sentence input by the user includes: obtaining the query sentence input by the user through text; or,
  • the user can input the above query sentence in the form of text, voice or picture.
  • the training process of the semantic recognition model includes: grouping the training corpus according to the number of negative corpus to be extracted; coding each group of training corpus to obtain the coding of each group of training corpus Value; extract the first type of negative corpus and the second type of negative corpus according to the coding value, where the first type of negative corpus can be a chat negative corpus, and the second type of negative corpus can be High-frequency positive vocabulary negative corpus; using the training corpus, the first type of negative corpus, and the second type of negative corpus for training to obtain the semantic recognition model.
  • each group of training corpus can be coded separately, so that each group of training corpus can have a unique coding value, and the coding method can include: hash value or simHash value, etc.;
  • the above-mentioned training corpus can also be sorted. Specifically, the following sorting methods can be used to sort the above-mentioned training corpus: according to the string, the hash (Hash) value of the training corpus, or the simHash value of the training corpus, etc., of course, other sorting methods can also be used to sort the training corpus. Sorting is not limited in this embodiment. In this embodiment, the training corpus is sorted, so that when the training corpus is completely the same, the coded value after the grouping does not change due to the change of the corpus order, so as to ensure that the grouping of the training corpus does not change.
  • each group of training corpus is coded, and the first type of negative corpus and the second type of negative corpus are extracted according to the above-mentioned coding value, and then the above-mentioned training corpus and the above-mentioned first
  • the type of negative corpus and the above-mentioned second type of negative corpus are trained to obtain a semantic recognition model, thereby achieving a unique extraction of negative corpus based on the coding value of the training corpus, and turning the randomization method of negative corpus generation into a stable
  • the model obtained by two or more trainings can be kept without obvious difference, so that the test corpus of the developer has almost the same in the model obtained by multiple training Confidence (difference ⁇ 0.01), which in turn makes the accuracy fluctuation smaller, thereby improving the developer's experience.
  • the extracting the negative corpus of the first type according to the code value includes: obtaining a first quantity of the negative corpus of the first type included in the first negative corpus set; wherein , The foregoing first negative corpus set may be a small chat negative corpus, and the first quantity is the total number of the first type of negative corpus included in the first negative corpus set;
  • the coding value of each group of training corpus and the first quantity obtain the first sampling value of the negative corpus of the first type; specifically, according to the coding value of each group of training corpus and the first quantity Obtaining the first sample value of the negative corpus of the first type may be: using the coding value of each set of training corpus to take the remainder of the first quantity, and the remainder operation is used as the mapping relationship, and the remainder is used as the first sample value;
  • Extracting a first negative corpus of a first type from the first negative corpus set according to the first sample value Specifically, a search may be performed in the first negative corpus set according to the first sample value, and the first negative corpus whose identifier (or index) matches the first sample value can be extracted.
  • the first negative corpus is extracted through the mapping relationship according to the coding value of the training corpus.
  • the training corpus is not added, deleted, or modified, the coding value of the training corpus remains unchanged, and the mapping relationship does not change. Therefore, the first negative corpus extracted through the mapping relationship according to the coding value of the training corpus will not change. Since the training corpus and the extracted negative corpus remain unchanged, the model obtained by training with the training corpus and the negative corpus is more stable.
  • the method further includes:
  • the training corpus includes all the positive corpus, that is, all the positive training corpus configured by the model training platform;
  • first similarity is less than the first similarity threshold, it is determined that the first negative corpus is successfully sampled, and the first negative corpus is added to the sampled corpus.
  • the method further includes:
  • the second sampling value is obtained according to the first sampling value; in specific implementation, the first sampling value may be added to the preset value , To obtain the second sampling value;
  • the second similarity is less than the first similarity threshold, it is determined that the second negative corpus is successfully sampled, and the second negative corpus is added to the sampled corpus.
  • the method further includes:
  • the number of repetitions is greater than the preset repetition number threshold, if the similarity between the negative corpus obtained by the current sampling and the training corpus is less than the second similarity threshold, it is determined that the negative corpus obtained by the current sampling is successfully sampled, Add the negative corpus obtained by the current sampling to the sampled corpus; if the similarity between the negative corpus obtained by the current sampling and the training corpus is greater than or equal to the second similarity threshold, then the negative corpus that was successfully sampled last time Add the sampled corpus again.
  • this embodiment removes the sampled corpus with higher similarity to the training corpus, so as to avoid the influence on the positive intent.
  • the extracting the second type of negative corpus according to the coded value includes: sequentially obtaining every M coded values from the coded value;
  • the second type of negative corpus is extracted from the second negative corpus set according to the third number of coded values; wherein, M and N are positive integers, and M ⁇ N.
  • the second negative corpus is extracted according to the coding value of the training corpus through the mapping relationship.
  • the training corpus is not added, deleted or modified, the coding value of the training corpus remains unchanged and the mapping relationship does not change. Therefore, the second negative corpus extracted through the mapping relationship according to the coding value of the training corpus will not change. Since the training corpus and the extracted negative corpus remain unchanged, the model obtained by training with the training corpus and the negative corpus is more stable.
  • the extracting the negative corpus of the first type and the negative corpus of the second type according to the coding value includes: according to the coding value of each group of training corpus and the pre-learned mapping relationship , Obtaining the third sampling value of the negative corpus of the first type and the fourth sampling value of the negative corpus of the second type;
  • the third sampling value of the first type of negative corpus and the second type of negative are obtained according to the coding value of each set of training corpus and the pre-learned mapping relationship Before the fourth sample value of the corpus, it also includes:
  • the training sample pair including the coding value of the training corpus and the sampling value of the corresponding negative corpus; wherein the distance between the sampling values of the negative corpus corresponding to the training corpus satisfies a preset constraint distance;
  • the training sample pair is used to learn the mapping relationship, and the mapping relationship includes the mapping relationship between the coding value of the training corpus and the sampling value of the corresponding negative corpus.
  • the negative corpus is extracted through the mapping relationship based on the coding value of the training corpus.
  • the training corpus is not added, deleted, or modified, the coding value of the training corpus remains unchanged, and the mapping relationship does not change, so according to The coding value of the training corpus will not change.
  • the negative corpus extracted through the mapping relationship will not change. Since the training corpus and the extracted negative corpus remain unchanged, the model obtained by training with the training corpus and the negative corpus has high stability and can be realized.
  • an artificial intelligence-based semantic recognition device including:
  • the obtaining module is used to obtain the query statement input by the user
  • the recognition module is used to recognize the query sentence through a pre-trained semantic recognition model to obtain the intent of the query sentence; wherein the pre-trained semantic recognition model is trained using training corpus and negative corpus, so The negative corpus is extracted according to the coding value of the training corpus mapped to the negative corpus;
  • a query module configured to obtain a response corresponding to the query sentence according to the query sentence acquired by the acquisition module and the intent of the query sentence recognized by the recognition module;
  • the display module is used to display the response corresponding to the query sentence.
  • the acquisition module is specifically used to acquire the query sentence input by the user through text; or, to acquire the query sentence input by the user through voice; or, to acquire the picture input by the user, and perform processing on the picture. Identify and obtain the query sentence included in the picture.
  • the device further includes:
  • the grouping module is used to group the training corpus according to the number of negative corpus to be extracted
  • the encoding module is used to encode each group of training corpus to obtain the coding value of each group of training corpus;
  • the extraction module is used to extract the negative corpus of the first type and the negative corpus of the second type according to the coding value obtained by the coding module;
  • the training module is configured to use the training corpus, the first type of negative corpus and the second type of negative corpus for training to obtain the semantic recognition model.
  • the extraction module includes:
  • the quantity acquisition sub-module is used to acquire the first quantity of the negative corpus of the first type included in the first negative corpus set;
  • the sampling value obtaining submodule is configured to obtain the first sampling value of the negative corpus of the first type according to the coding value of each group of training corpus and the first quantity;
  • the corpus extraction submodule is configured to extract the first negative corpus of the first type from the first negative corpus set according to the first sample value obtained by the sample value obtaining submodule.
  • the extraction module further includes:
  • the similarity calculation submodule is used to calculate the first similarity between the first negative corpus and the training corpus after the first negative corpus of the first type is extracted by the corpus extraction submodule;
  • the corpus extraction submodule is further configured to determine that the first negative corpus is successfully sampled when the first similarity is less than the first similarity threshold, and add the first negative corpus to the sampled corpus.
  • the sampling value obtaining submodule is further configured to, after the similarity calculation submodule calculates the first similarity, if the first similarity is greater than or equal to the first similarity threshold , Then obtain a second sample value according to the first sample value;
  • the corpus extraction sub-module is further configured to obtain the second sample value obtained by the sub-module according to the sample value, and extract the second negative corpus of the first type from the first negative corpus set;
  • the similarity calculation sub-module is also used to calculate a second similarity between the second negative corpus and the training corpus;
  • the corpus extraction submodule is further configured to determine that the second negative corpus is successfully sampled when the second similarity is less than the first similarity threshold, and add the second negative corpus to the sampled corpus .
  • the sampling value obtaining sub-module is further configured to, after the similarity calculation sub-module calculates the second similarity, if the second similarity is greater than or equal to the first similarity threshold , Repeating the step of obtaining a second sample value according to the first sample value and subsequent steps;
  • the corpus extraction submodule is also used to determine if the similarity between the negative corpus obtained by the current sampling and the training corpus is less than the second similarity threshold when the number of repetitions is greater than the preset repetition number threshold If the negative corpus obtained by the current sampling is successfully sampled, the negative corpus obtained by the current sampling is added to the sampled corpus; if the similarity between the negative corpus obtained by the current sampling and the training corpus is greater than or equal to the second similarity threshold, Then, the negative corpus that was successfully sampled last time is added to the sampled corpus again.
  • the extraction module includes:
  • An encoding value acquisition sub-module for acquiring every M encoding values in order from the encoding values; and selecting a second number of encoding values from every M encoding values obtained;
  • the corpus extraction sub-module is used to extract the second type of negative corpus from the second negative corpus set according to the second number of coded values;
  • Coding value sorting sub-module used for sorting the coding value
  • the coded value obtaining submodule is further configured to obtain every N coded values in order from the sorted coded values; and select a third number of coded values from every N obtained coded values;
  • the corpus extraction submodule is also used to extract the second type of negative corpus from the second negative corpus set according to the third number of coded values; where M, N are positive integers, and M ⁇ N.
  • the extraction module includes:
  • the sampling value obtaining sub-module is used to obtain the third sampling value of the negative corpus of the first type and the first sampling value of the negative corpus of the second type according to the coding value of each group of training corpus and the pre-learned mapping relationship.
  • the corpus extraction sub-module is used to extract the negative corpus of the first type from the first negative corpus according to the third sample value obtained by the sample value obtaining sub-module, and obtain the negative corpus of the first type according to the fourth sample value. Extract the second type of negative corpus from the second negative corpus set.
  • the extraction module further includes:
  • the sample pair acquisition submodule is used to acquire the training sample pair before the sample value acquisition submodule acquires the third sample value of the negative corpus of the first type and the fourth sample value of the negative corpus of the second type, so
  • the training sample pair includes the coding value of the training corpus and the sampling value of the corresponding negative corpus; wherein the distance between the sampling values of the negative corpus corresponding to the training corpus satisfies a preset constraint distance;
  • the mapping relationship learning sub-module is configured to use the training sample pair to learn the mapping relationship, and the mapping relationship includes the mapping relationship between the coding value of the training corpus and the sampling value of the corresponding negative corpus.
  • an embodiment of the present application provides an artificial intelligence-based semantic recognition device, including: a display screen; one or more processors; a memory; a plurality of application programs; and one or more computer programs, wherein the one One or more computer programs are stored in the memory, and the one or more computer programs include instructions, and when the instructions are executed by the device, the device executes the following steps:
  • the query sentence is recognized through a pre-trained semantic recognition model to obtain the intention of the query sentence; wherein the pre-trained semantic recognition model is trained using training corpus and negative corpus, and the negative corpus is Mapping the coding value of the training corpus to the negative corpus extracted from the set;
  • the response corresponding to the query sentence is displayed.
  • the device when the instruction is executed by the device, the device specifically executes the following steps:
  • the device when the instruction is executed by the device, the device specifically executes the following steps:
  • Training is performed using the training corpus, the negative corpus of the first type, and the negative corpus of the second type to obtain the semantic recognition model.
  • embodiments of the present application provide a computer-readable storage medium in which a computer program is stored, and when it runs on a computer, the computer executes the method described in the first aspect.
  • an embodiment of the present application provides a computer program, which is used to execute the method described in the first aspect when the computer program is executed by a computer.
  • the program in the fifth aspect may be stored in whole or in part on a storage medium packaged with the processor, or may be stored in part or in a memory not packaged with the processor.
  • FIG. 1 is a schematic diagram of problems in the model generated by the Bot platform in the related art
  • Fig. 2 is a schematic diagram of the confidence level of the model obtained by training using the method provided in this application;
  • FIG. 3 is a flowchart of an embodiment of the training process of the semantic recognition model in the artificial intelligence-based semantic recognition method of this application;
  • FIG. 5 is a flowchart of another embodiment of the training process of the semantic recognition model in the artificial intelligence-based semantic recognition method of this application;
  • FIG. 6 is a flowchart of another embodiment of the training process of the semantic recognition model in the artificial intelligence-based semantic recognition method of this application;
  • FIG. 7 is a flowchart of an embodiment of a semantic recognition method based on artificial intelligence in this application.
  • FIG. 8 is a schematic structural diagram of an embodiment of a semantic recognition device based on artificial intelligence in this application.
  • FIG. 9 is a schematic structural diagram of another embodiment of an artificial intelligence-based semantic recognition device according to the present application.
  • FIG. 10 is a schematic structural diagram of an embodiment of a semantic recognition device based on artificial intelligence in this application.
  • the machine learning intention recognition algorithm used by the Bot platform retrains the model, and the main reason for the model differentiation is that there are two random factors.
  • a random factor is to randomly oppose the generation of negative corpus.
  • the method of generating this negative corpus is random extraction, for example: randomly select "Hello” and other small chat corpora; another random factor is to automatically counter the generation of negative corpus .
  • This kind of negative corpus is generated in a generative way, for example, according to "card”, negative corpus such as "bank card” or “mobile phone card” is generated.
  • the above two methods of randomly generating negative corpus will make the model unstable.
  • this application proposes a de-randomized stability sampling method, that is, the two methods of generating negative corpus mentioned above become stable, and at the same time, it has the effect of randomized extraction of negative corpus, so that training The model remains stable.
  • the two training models on the Bot platform are consistent, so that the same corpus has the same prediction effect under the model obtained by multiple training.
  • the Bot platform is only an example of a model training platform, and the above-mentioned model training platform may also be other platforms, which is not limited in this application.
  • Fig. 1 is a schematic diagram of problems in the model generated by the Bot platform in the related art.
  • the random generation method of negative corpus has instability, which leads to the model obtained by retraining the Bot platform.
  • the technical problem to be solved by this application is to change the randomized generation method of negative corpus into a stable generation method.
  • the training corpus is not added, deleted or modified, it can be maintained twice or more.
  • the test corpus of the developers has almost the same confidence in the models obtained in multiple trainings (difference ⁇ 0.01), as shown in Figure 2, which in turn makes the accuracy fluctuations smaller, thus Improve the developer experience.
  • Fig. 2 is a schematic diagram of the confidence level of the model obtained by training using the method provided in the present application.
  • the semantic recognition model in this application can be obtained by training on a model training platform, such as a Bot platform.
  • the model training platform can be deployed on a cloud server.
  • the semantic recognition model mentioned above It can also be obtained by training on other devices, and this embodiment does not limit the execution subject for training the above semantic recognition model.
  • FIG. 3 is a flowchart of an embodiment of the training process of the semantic recognition model in the artificial intelligence-based semantic recognition method of this application. As shown in FIG. 3, it may include:
  • Step 301 Group the training corpus according to the number of negative corpus to be extracted.
  • the number of negative corpora that needs to be extracted can be set during specific implementation according to implementation requirements and/or system performance. This embodiment does not limit this, assuming that the number of negative corpora that needs to be extracted is Num , Then you need to divide the training corpus into Num groups, where Num is a positive integer.
  • the above-mentioned training corpus can also be sorted. Specifically, the following sorting methods can be used to sort the above-mentioned training corpus: according to the string, the hash (Hash) value of the training corpus, or the simHash value of the training corpus, etc., of course, other sorting methods can also be used to sort the training corpus. Sorting is not limited in this embodiment. In this embodiment, the training corpus is sorted, so that when the training corpus is completely the same, the coded value after the grouping does not change due to the change of the corpus order, so as to ensure that the grouping of the training corpus does not change.
  • Step 302 Coding each group of training corpus to obtain the coding value of each group of training corpus.
  • each group of training corpus can be coded separately, so that each group of training corpus can have a unique coding value.
  • the coding method can include: hash value or simHash
  • this embodiment can use N-grams for each group of training corpus, for example, unigram and bigram, and use the calculated simHash value as the coding value of each group of training corpus.
  • Step 303 Extract the negative corpus of the first type and the negative corpus of the second type according to the above-mentioned coded value.
  • the above-mentioned first type of negative corpus may be small talk negative corpus
  • the second type of negative corpus may be high-frequency positive vocabulary negative corpus.
  • the negative corpus of small talk can include small talk corpus such as "Hello";
  • the negative corpus of high-frequency positive vocabulary can include "bank card” or "good mobile phone card” generated from the high-frequency positive vocabulary "ka” Iso-negative corpus, wherein the above-mentioned high-frequency positive words include words that appear frequently in the above-mentioned training corpus.
  • Step 304 Use the aforementioned training corpus, the aforementioned first type of negative corpus, and the aforementioned second type of negative corpus for training to obtain the aforementioned semantic recognition model.
  • each group of training corpus is coded, and the first type of negative corpus and the second type of negative corpus are extracted according to the coding value, and then the training corpus and the first type are used.
  • the first type of negative corpus and the second type of negative corpus are trained to obtain a semantic recognition model, so that the negative corpus is uniquely extracted according to the coding value of the training corpus, and the randomization method of negative corpus generation is changed into Stable generation method, in the case that the training corpus is not added, deleted or modified, the model obtained by two or more trainings can be kept without obvious difference, so that the test corpus of the developer is almost the same in the model obtained by multiple training
  • the confidence level (difference ⁇ 0.01), which in turn makes the accuracy fluctuation smaller, thereby improving the developer’s experience.
  • FIG. 4 is a flowchart of another embodiment of the training process of the semantic recognition model in the artificial intelligence-based semantic recognition method of this application. As shown in FIG. 4, in the embodiment shown in FIG. 3 of this application, step 303 is based on the above coding
  • the first type of negative corpus for value extraction can include:
  • Step 401 Obtain a first number of negative corpora of the first type included in the first negative corpus set.
  • the above-mentioned first negative corpus set may be a small chat negative corpus, and the first quantity is the total number of the first type of negative corpus included in the first negative corpus set.
  • Step 402 Obtain the first sample value of the negative corpus of the first type according to the coding value of each group of training corpus and the first quantity.
  • obtaining the first sampling value of the negative corpus of the first type may be: using the coding value of each training corpus to take the first quantity For the remainder, the remainder operation is used as the mapping relationship, and the remainder is used as the first sample value.
  • the above is just one way to obtain the first sampling value of the negative corpus of the first type according to the coding value of each group of training corpus and the first quantity. It can also be based on the coding value of each group of training corpus and the above For the first quantity, other implementations are used to obtain the first sample value of the negative corpus of the first type, which is not limited in this embodiment.
  • Step 403 Extract the first negative corpus of the first type from the first negative corpus according to the above-mentioned first sample value.
  • a search may be performed in the first negative corpus set according to the first sample value, and the first negative corpus whose identifier (or index) matches the first sample value can be extracted.
  • step 403 it may also include:
  • Step 404 Calculate the first similarity between the first negative corpus and the above-mentioned training corpus. Then step 405 or step 406 is executed.
  • the training corpus here includes all the positive corpus, that is, the model training platform All configured forward training corpora.
  • the lucene algorithm may be used to calculate the first similarity between the first negative corpus and the above-mentioned training corpus.
  • Step 405 If the first similarity is less than the first similarity threshold, it is determined that the first negative corpus is successfully sampled, and the first negative corpus is added to the sampled corpus. This process ends.
  • the above-mentioned first similarity threshold may be set according to system performance and/or implementation requirements during specific implementation, and this embodiment does not limit the size of the above-mentioned first similarity threshold.
  • Step 406 If the first similarity is greater than or equal to the first similarity threshold, a second sampling value is obtained according to the first sampling value.
  • the above-mentioned first sampling value may be added to a preset value to obtain the second sampling value.
  • the size of the aforementioned preset value can be set by itself during specific implementation according to system performance and/or implementation requirements, etc.
  • the present embodiment does not limit the size of the aforementioned preset value.
  • Step 407 Extract the second negative corpus of the first type from the first negative corpus according to the above-mentioned second sampling value.
  • Step 408 Calculate the second similarity between the second negative corpus and the above-mentioned training corpus. Then step 409 or step 410 is executed.
  • the training corpus here includes all the positive corpus, that is, the model training platform All configured forward training corpora.
  • Step 409 If the second similarity is less than the first similarity threshold, it is determined that the second negative corpus is successfully sampled, and the second negative corpus is added to the sampled corpus. This process ends.
  • Step 410 if the second similarity is greater than or equal to the first similarity threshold, repeat steps 406 to 409; when the number of repeated executions is greater than the preset repetition threshold, if the current sampling obtains a negative If the similarity between the corpus and the training corpus is less than the second similarity threshold, it is determined that the negative corpus obtained by the current sampling is successfully sampled, and the negative corpus obtained by the current sampling is added to the sampling corpus; if the negative corpus obtained by the current sampling is If the similarity between the corpus and the training corpus is greater than or equal to the second similarity threshold, the negative corpus that was successfully sampled last time is added to the above-mentioned sampled corpus again.
  • the foregoing preset repetition count threshold can be set by itself during specific implementation according to system performance and/or implementation requirements. This embodiment does not limit the size of the foregoing preset repetition count.
  • the foregoing The preset number of repetitions can be 5.
  • the size of the second similarity threshold can be set during specific implementation according to system performance and/or implementation requirements. This embodiment does not limit the size of the second similarity threshold, as long as the second similarity threshold is greater than The first similarity threshold is sufficient.
  • this embodiment removes the sampled corpus with higher similarity to the training corpus, so as to avoid the influence on the positive intent.
  • Fig. 5 is a flowchart of another embodiment of the training process of the semantic recognition model in the artificial intelligence-based semantic recognition method of this application. As shown in Fig. 5, in the embodiment shown in Fig. 3 of this application, step 303 is based on the above coding
  • the second type of negative corpus for value extraction can include:
  • Step 501 Obtain every M coded values in order from the coded values.
  • Step 502 Select a second number of coded values from every M coded values obtained.
  • the second number can be set by itself during specific implementation, and this embodiment does not limit the size of the second number.
  • Step 503 Extract a second type of negative corpus from the second negative corpus set according to the above-mentioned second number of coded values.
  • the second negative corpus set may be a high-frequency positive vocabulary negative corpus, and the second type of negative corpus included in the second negative corpus set may be a high-frequency word.
  • Step 504 Sort the above-mentioned coded values.
  • Step 505 Obtain every N coded values in order from the above-mentioned sorted coded values.
  • Step 506 Select a third number of coded values from every N obtained coded values.
  • the third number can be set by itself during specific implementation, and this embodiment does not limit the size of the third number.
  • Step 507 Extract a second type of negative corpus from the second negative corpus set according to the third number of coded values.
  • M and N are positive integers, and M ⁇ N.
  • the size of M and N can be set during specific implementation according to system performance and/or implementation requirements. This embodiment does not limit the size of M and N.
  • M can be 2, and N can be Is 3.
  • each 2 coding values can be obtained in order from the above coding values, namely a1a2, a2a3 and a3a4, then, you can select the second number of encoding values from every two acquired encoding values (a1a2, a2a3, and a3a4).
  • the second number is 2, then you can select the two sets of encoding values of a1a2 and a2a3, of course
  • the two sets of coding values of a1a2 and a3a4 can also be selected, which is not limited in this embodiment.
  • the selection of the two sets of coding values of a1a2 and a2a3 is taken as an example for description. It should be noted that if the second number is 2, in the first selection, the two sets of coding values a1a2 and a2a3 are selected, then each subsequent model training, the two sets of coding values a1a2 and a2a3 still need to be selected .
  • the coding values of a1 and a2 are respectively mapped to the second negative corpus set.
  • the simplest mapping method here is that the coding values are included in the second negative corpus set. Take the remainder of the total number of negative corpus, and then concatenate the second negative corpus extracted according to the coding values of a1 and a2 to generate Bigram negative corpus, and use the generated Bigram negative corpus as the second type of negative corpus , And then add the generated Bigram negative corpus to the set of negative corpus required for training corpus. The same method can be followed to generate the second type of negative corpus corresponding to a2a3.
  • the third number of encoding values can be obtained from every three acquired encoding values (ie, a2a1a3 and a1a3a4).
  • the group code value is not limited in this embodiment.
  • the group code value a2a1a3 is selected as an example for description. It should be noted that if the third number is 1, the a2a1a3 group of coded values is selected in the first selection, then the group of coded values a2a1a3 still needs to be selected for each subsequent model training.
  • the coded values of a2, a1, and a3 can be respectively mapped to the second negative corpus set.
  • the simplest mapping method here is that the coded value is taken from the total number of negative corpus included in the second negative corpus set The remainder, then the second negative corpus extracted according to the coding values of a2, a1, and a3 are spliced to generate a Trigram negative corpus, and the generated Trigram negative corpus is used as the second type of negative corpus, and then the generated Trigram negative corpus is added to the set of negative corpus required for training corpus.
  • the purpose of reordering a1, a2, a3, and a4 is to make the generated Trigram negative corpus not include the generated Bigram negative corpus.
  • the second negative corpus is extracted according to the coding value of the training corpus through the mapping relationship.
  • the training corpus is not added, deleted or modified, the coding value of the training corpus remains unchanged and the mapping relationship does not change. Therefore, the second negative corpus extracted through the mapping relationship according to the coding value of the training corpus will not change. Since the training corpus and the extracted negative corpus remain unchanged, the model obtained by training with the training corpus and the negative corpus is more stable.
  • This embodiment does not limit the total number of negative corpora included in the first negative corpus set and the second negative corpus set.
  • a mapping method based on the coding value is proposed to map to the first negative corpus set and the second negative corpus set.
  • the step 303 is based on the above coding Value extraction of the first type of negative corpus and the second type of negative corpus can include:
  • Step 601 Obtain the third sample value of the negative corpus of the first type and the fourth sample value of the negative corpus of the second type according to the coding value of each group of training corpus and the pre-learned mapping relationship.
  • the foregoing pre-learned mapping relationship may include the remainder, which is not limited in this embodiment.
  • Step 602 Extract the negative corpus of the first type from the first negative corpus set according to the third sample value, and extract the negative corpus of the second type from the second negative corpus set according to the fourth sample value Corpus.
  • step 601 may further include:
  • Step 603 Obtain a training sample pair.
  • the training sample pair includes the coding value of the training corpus and the sampling value of the corresponding negative corpus; wherein the distance between the sampling values of the negative corpus corresponding to the training corpus satisfies a preset Constrain the distance.
  • the aforementioned pair of training samples can be obtained in the following manner:
  • the problem of measuring the probability distribution of different spaces is transformed into an equidistant constraint problem of the relative positions of training samples in different spaces, ensuring two The spatial probability distribution is consistent in the sense of distance.
  • MDS Multi-dimensional Scaling
  • the specific method can be based on the kernel (kernel) learning method, without explicitly defining the mapping, directly through the nearest neighbor samples, defining the relative mapping relationship , Construct a training sample pair.
  • Step 604 Use the training sample pair to learn a mapping relationship.
  • the mapping relationship includes the mapping relationship between the coding value of the training corpus and the sampling value of the corresponding negative corpus.
  • the first type of negative corpus is extracted from the first negative corpus to obtain a sampled corpus.
  • the distribution of k sampled corpora ⁇ k'i ⁇ in the sampled corpus set the same as the distribution of k sampled corpora ⁇ ki ⁇ randomly selected from the first negative corpus set, it can be equivalent to finding ⁇ ki ⁇
  • the KL divergence Kullback-Leibler divergence
  • the KL divergence can be calculated by the formula shown in formula (1).
  • P(i) is the distribution of ⁇ ki ⁇
  • Q(i) is the distribution of ⁇ k'i ⁇
  • Q) is the distribution of ⁇ ki ⁇ and the distribution of ⁇ k'i ⁇ KL divergence between distributions.
  • the hash method is selected as simhash, and the simplest method to embed in the first negative corpus is to take the remainder (as described in the embodiment shown in FIG. 4 of this application).
  • the training corpus is grouped, and then the simhash value of each training corpus is calculated as the coding value. To the greatest extent, the overlap of hash(x'i) can be intuitively minimized.
  • Simhash guarantees the similarity constraint (based on Hamming distance, generally 64-bit simhash, Hamming distance less than 3 can be considered dissimilar).
  • the category refers to the internal structure category of the negative corpus, such as the category label of the negative corpus itself, or no category label.
  • the internal structure category is determined by means such as clustering.
  • this embodiment proposes a distance constraint based on the projection to the first negative corpus: the encoding value x of the training corpus before mapping is solved in reverse, and this x is a "virtual sample" mapped by the hash function to be designed.
  • Mapping(x) gets the final sample value of negative corpus.
  • the simplest mapping relationship (Mapping) is to take the remainder, and the simplest hash is the hashcode of JDK String.
  • CDF Cumulative Distribution Function
  • simhash controls overlapping in advance, and embeds original prior information to maintain the mapping relationship
  • the negative corpus is extracted through the mapping relationship according to the coding value of the training corpus.
  • the training corpus is not added, deleted, or modified, the coding value of the training corpus remains unchanged, and the mapping relationship does not change, so according to The coding value of the training corpus will not change.
  • the negative corpus extracted through the mapping relationship will not change. Since the training corpus and the extracted negative corpus remain unchanged, the model obtained by training with the training corpus and the negative corpus has high stability and can be realized.
  • the embodiment shown in Fig. 3 to Fig. 5 of the present application uses the remainder of the coded value to map to the negative corpus set, and the embodiment shown in Fig. 6 of the present application learns the mapping relationship to extract the random method of negative corpus equivalent to
  • the extraction method of the stable method is not only suitable for extracting negative corpus in this application, but also suitable for other random factors that need to be consistent every time.
  • a trained semantic recognition model can be obtained, and then the above-mentioned trained semantic recognition model can be used to perform semantic recognition on the input query sentence.
  • the above-mentioned semantic recognition model can be mounted on a model training platform, for example, semantic recognition is performed on the Bot platform, as shown in Figure 2.
  • the box on the right side of Figure 2 shows that according to Figures 3 to 6 of this application
  • the above-mentioned trained semantic recognition model is used on the Bot platform to query the query sentence "How is the weather in Shenyang the day after tomorrow" input by the user to obtain the query sentence The intention of is "weather query”, and then the answer to the above query sentence is "Shenyang Houtian is sunny,", and an example of the above answer is displayed.
  • the aforementioned semantic recognition model can also be carried on other electronic devices, such as servers or terminals for semantic recognition.
  • the above-mentioned electronic devices may include: cloud servers, mobile terminals (mobile phones), smart screens, drones, intelligent connected vehicles (Intelligent Connected Vehicles; hereinafter referred to as ICVs), smart/intelligent cars (smart/intelligent cars) or in-vehicles Equipment and other equipment.
  • FIG. 7 is a flowchart of an embodiment of an artificial intelligence-based semantic recognition method according to this application. As shown in FIG. 7, the above artificial intelligence-based semantic recognition method may include:
  • Step 701 Obtain a query sentence input by the user.
  • obtaining the query sentence input by the user may include:
  • the picture input by the user is acquired, the picture is recognized, and the query sentence included in the picture is obtained.
  • the user can input the above query sentence in the form of text, voice or picture.
  • Step 702 Recognize the query sentence through a pre-trained semantic recognition model to obtain the intention of the query sentence; wherein, the pre-trained semantic recognition model is trained using training corpus and negative corpus, and the negative corpus is based on The coding values of the above training corpus are mapped to those extracted from the negative corpus set.
  • Step 703 Obtain a response corresponding to the query sentence according to the query sentence and the intent of the query sentence.
  • step 704 the response corresponding to the above query sentence is displayed.
  • the negative corpus is extracted through the mapping relationship according to the coding value of the training corpus.
  • the training corpus is not added, deleted or modified, the coding value of the training corpus remains unchanged, and the mapping relationship is not Will change, so the negative corpus extracted through the mapping relationship according to the coding value of the training corpus will not change.
  • the training corpus and the extracted negative corpus remain unchanged, the stability of the model obtained by training with the training corpus and the negative corpus Higher, it can be achieved that in the case of no addition, deletion or modification of the training corpus, the model obtained by two or multiple trainings has no obvious difference, so that the test corpus of the developer has almost the same in the model obtained by multiple training Confidence (difference ⁇ 0.01), which in turn makes the accuracy fluctuation smaller, thereby improving the developer's experience.
  • FIG. 8 is a schematic structural diagram of an embodiment of an artificial intelligence-based semantic recognition device of this application.
  • the aforementioned artificial intelligence-based semantic recognition device 80 may include: an acquisition module 81, a recognition module 82, a query module 83, and a display Module 84.
  • the artificial intelligence-based semantic recognition device 80 may correspond to the device 900 in FIG. 10.
  • the functions of the acquiring module 81, the identifying module 82, and the query module 83 may be implemented by the processor 910 in the device 900 in FIG. 10; the display module 84 may specifically correspond to the display unit 970 in the device 900 in FIG.
  • the obtaining module 81 is used to obtain the query sentence input by the user; in this embodiment, the obtaining module 81 is specifically used to obtain the query sentence input by the user through text; or, obtain the query sentence input by the user through voice; or, obtain The picture input by the user is identified, and the query sentence included in the picture is obtained.
  • the user can input the above query sentence in the form of text, voice or picture.
  • the recognition module 82 is used for recognizing the above-mentioned query sentence through a pre-trained semantic recognition model to obtain the intention of the above-mentioned query sentence; wherein, the above-mentioned pre-trained semantic recognition model is trained using training corpus and negative corpus. The corpus is extracted according to the coding value of the training corpus mapped to the negative corpus set;
  • the query module 83 is configured to obtain a response corresponding to the query statement according to the query statement obtained by the obtaining module 81 and the intention of the query statement recognized by the recognition module 82;
  • the display module 84 is used to display the response corresponding to the above query sentence.
  • the artificial intelligence-based semantic recognition device provided in the embodiment shown in FIG. 8 can be used to implement the technical solution of the method embodiment shown in FIG. 7 of the present application.
  • FIG. 9 is a schematic structural diagram of another embodiment of the artificial intelligence-based semantic recognition device of this application. Compared with the artificial intelligence-based semantic recognition device shown in FIG. 8, the difference is that the artificial intelligence-based semantic recognition device shown in FIG.
  • the semantic recognition device 90 may further include: a grouping module 85, an encoding module 86, an extraction module 87, and a training module 88; it should be understood that the semantic recognition device 90 based on artificial intelligence may correspond to the device 900 in FIG. 10.
  • the functions of the acquiring module 81, the identifying module 82, and the query module 83 can be implemented by the processor 910 in the device 900 of FIG. 10; the display module 84 may specifically correspond to the display unit 970 in the device 900 of FIG. 10; the grouping module 85
  • the functions of the encoding module 86, the extraction module 87, and the training module 88 can be implemented by the processor 910 in the device 900 in FIG. 10.
  • the grouping module 85 is used to group the training corpus according to the number of negative corpus to be extracted; wherein, the number of negative corpus to be extracted can be implemented according to the implementation requirements and/or system performance. Set it by yourself, this embodiment does not limit this. Assuming that the number of negative corpus to be extracted is Num, then the training corpus needs to be divided into Num groups, and Num is a positive integer.
  • the artificial intelligence-based semantic recognition device 80 may further include: a sorting module 89;
  • the sorting module 89 is used for sorting the training corpus before the grouping module 85 groups the training corpus according to the number of negative corpus to be extracted. Specifically, the sorting module 89 may use the following sorting method to sort the above-mentioned training corpus: sort according to the string, the hash (Hash) value of the training corpus, or the simHash value of the training corpus, etc. Of course, other sorting methods can also be used to sort the training corpus.
  • the training corpus is sorted, which is not limited in this embodiment. In this embodiment, the training corpus is sorted by the sorting module 89. When the training corpus is completely the same, the grouped coding value does not change due to the change of the corpus order, so as to ensure that the grouping of the training corpus does not change.
  • the encoding module 86 is used to encode each set of training corpus to obtain the encoding value of each set of training corpus; specifically, in this embodiment, after the grouping module 85 groups each set of training corpus, the encoding module 86 can separately Each group of training corpus is coded, so that each group of training corpus can have a unique coding value.
  • the coding method may include: hash value or simHash value, etc. Of course, other coding methods may also be used, which is not limited in this embodiment.
  • the encoding module 86 can use N-grams for each group of training corpus, such as unigram and bigram, to use the calculated simHash value as the encoding value of each group of training corpus.
  • the extraction module 87 is used to extract the negative corpus of the first type and the negative corpus of the second type according to the coding value obtained by the coding module 86; wherein, the negative corpus of the first type may be a chat negative corpus, and the second The type of negative corpus can be high-frequency positive vocabulary negative corpus.
  • the training module 88 is configured to use the above-mentioned training corpus, the above-mentioned first type of negative corpus and the above-mentioned second type of negative corpus for training to obtain the above-mentioned semantic recognition model.
  • the encoding module 86 encodes each group of training corpus
  • the extraction module 87 extracts the first type of negative corpus and the second type of negative corpus according to the aforementioned coding values
  • the training module 88 uses the above-mentioned training corpus, the above-mentioned first type of negative corpus and the above-mentioned second type of negative corpus for training to obtain a semantic recognition model, thereby achieving a unique extraction of negative corpus based on the coding value of the training corpus ,
  • the randomization method of negative corpus generation is turned into a stable generation method.
  • the model obtained by two or more trainings can be maintained without obvious difference, so that the developer's test
  • the corpus has almost the same confidence (difference ⁇ 0.01) in the models obtained through multiple trainings, which in turn reduces the fluctuation of the accuracy rate, thereby improving the developer's experience.
  • the extraction module 87 may include: a quantity acquisition sub-module 871, a sampling value acquisition sub-module 872, and a corpus extraction sub-module 873;
  • the quantity acquisition sub-module 871 is used to acquire the first quantity of the negative corpus of the first type included in the first negative corpus set; wherein, the first negative corpus set may be a small chat negative corpus, the first The quantity is the total quantity of the first type of negative corpus included in the first negative corpus set.
  • the sampling value obtaining sub-module 872 is configured to obtain the first sampling value of the negative corpus of the first type according to the coding value of each group of training corpus and the first quantity; specifically, according to the coding value of each group of training corpus In addition to the above first quantity, obtaining the first sample value of the negative corpus of the first type may be: the sample value obtaining submodule 872 uses the coding value of each set of training corpus to take the remainder of the above first quantity, and perform the remainder operation As the mapping relationship, the remainder is used as the first sample value. The above is just one way to obtain the first sampling value of the negative corpus of the first type according to the coding value of each set of training corpus and the first quantity.
  • the sampling value obtaining submodule 872 may also be based on the training of each set of the above.
  • the coding value of the corpus and the above-mentioned first quantity are obtained by using other implementation methods to obtain the first sampling value of the above-mentioned negative corpus of the first type, which is not limited in this embodiment.
  • the corpus extraction sub-module 873 is configured to extract the first negative corpus of the first type from the first negative corpus set according to the first sample value obtained by the sample value obtaining sub-module 872.
  • the corpus extraction submodule 873 may search in the first negative corpus set according to the first sample value, and extract the first negative corpus whose identifier (or index) matches the above-mentioned first sample value.
  • the aforementioned extraction module 87 may also include: a similarity calculation sub-module 874;
  • the similarity calculation submodule 874 is used to calculate the first similarity between the first negative corpus and the above training corpus after the corpus extraction submodule 873 extracts the first negative corpus of the first type; specifically, in the corpus extraction submodule 873 After the module 873 extracts the first negative corpus according to the first sample value, the similarity calculation sub-module 874 needs to calculate the first similarity between the first negative corpus and the above-mentioned training corpus.
  • the training corpus here includes all the positive corpus. That is, all the forward training corpus configured by the model training platform.
  • the corpus extraction submodule 873 is further configured to determine that the first negative corpus is successfully sampled when the first similarity is less than the first similarity threshold, and add the first negative corpus to the sampled corpus.
  • the above-mentioned first similarity threshold may be set according to system performance and/or implementation requirements during specific implementation, and this embodiment does not limit the size of the above-mentioned first similarity threshold.
  • the sampling value obtaining sub-module 872 is further configured to, after the similarity calculation sub-module 874 calculates the first similarity, if the first similarity is greater than or equal to the first similarity threshold, according to the first sampling Value to obtain the second sampled value; in specific implementation, the sampled value obtaining submodule 872 may add the above-mentioned first sampled value to a preset value to obtain the second sampled value.
  • the size of the aforementioned preset value can be set by itself during specific implementation according to system performance and/or implementation requirements, etc.
  • the present embodiment does not limit the size of the aforementioned preset value.
  • the corpus extraction sub-module 873 is further configured to obtain the second sample value obtained by the sub-module 872 according to the sample value, and extract the second negative corpus of the first type from the first negative corpus set; similarly, the corpus extraction sub-module 873 can Search in the first negative corpus according to the second sample value, and extract the second negative corpus whose identifier (or index) matches the above-mentioned second sample value.
  • the similarity calculation submodule 874 is also used to calculate the second similarity between the second negative corpus and the training corpus; specifically, after the corpus extraction submodule 873 extracts the second negative corpus according to the second sampling value, the similarity
  • the degree calculation sub-module 874 needs to calculate the first similarity between the second negative corpus and the above-mentioned training corpus.
  • the training corpus here includes all the positive corpus, that is, all the positive training corpus configured by the model training platform.
  • the corpus extraction submodule 873 is further configured to determine that the second negative corpus is successfully sampled when the second similarity is less than the first similarity threshold, and add the second negative corpus to the above-mentioned sampled corpus.
  • the sampling value obtaining sub-module 872 is further configured to, after the similarity calculation sub-module 874 calculates the second similarity, if the above-mentioned second similarity is greater than or equal to the first similarity threshold, repeat the execution of obtaining the first sample value according to the first sampling value. Steps to sample the value and subsequent steps;
  • the corpus extraction sub-module 873 is also used to determine the current sampling if the similarity between the negative corpus obtained by the current sampling and the training corpus is less than the second similarity threshold when the number of repeated executions is greater than the preset repetition number threshold.
  • the obtained negative corpus is successfully sampled, and the negative corpus obtained by the current sampling is added to the above-mentioned sampled corpus; if the similarity between the negative corpus obtained by the current sampling and the above-mentioned training corpus is greater than or equal to the second similarity threshold, the previous The negative corpus that is successfully sampled is added to the above-mentioned sampled corpus again.
  • the foregoing preset repetition count threshold can be set by itself during specific implementation according to system performance and/or implementation requirements. This embodiment does not limit the size of the foregoing preset repetition count.
  • the foregoing The preset number of repetitions can be 5.
  • the size of the second similarity threshold can be set during specific implementation according to system performance and/or implementation requirements. This embodiment does not limit the size of the second similarity threshold, as long as the second similarity threshold is greater than The first similarity threshold is sufficient.
  • this embodiment removes the sampled corpus with higher similarity to the training corpus, so as to avoid the influence on the positive intent.
  • the extraction module 87 may include: an encoding value acquisition submodule 875, a corpus extraction submodule 873, and an encoding value ranking submodule 876;
  • the code value obtaining sub-module 875 is configured to sequentially obtain every M code values from the above code values; and select the second number of code values from every M code values obtained;
  • the corpus extraction sub-module 873 is used to extract the second type of negative corpus from the second negative corpus set according to the second number of coded values; wherein the second negative corpus set can be a high-frequency positive vocabulary negative corpus Set, the second type of negative corpus included in the second negative corpus set may be high-frequency words.
  • the coded value sorting sub-module 876 is used to sort the aforementioned coded values
  • the code value obtaining submodule 875 is further configured to obtain every N code values in order from the above sorted code values; and select a third number of code values from every N code values obtained;
  • the corpus extraction submodule 873 is further configured to extract the second type of negative corpus from the second negative corpus set according to the third number of coded values; where M and N are positive integers, and M ⁇ N.
  • the size of M and N can be set during specific implementation according to system performance and/or implementation requirements. This embodiment does not limit the size of M and N.
  • M can be 2, and N can be Is 3.
  • each 2 coding values can be obtained in order from the above coding values, namely a1a2, a2a3, and a3a4, then, you can select the second number of encoding values from every two acquired encoding values (a1a2, a2a3, and a3a4).
  • the second number is 2, then you can choose the two sets of encoding values of a1a2 and a2a3, of course
  • the two sets of coding values of a1a2 and a3a4 can also be selected, which is not limited in this embodiment.
  • the selection of the two sets of coding values of a1a2 and a2a3 is taken as an example for description. It should be noted that if the second number is 2, in the first selection, the two sets of coding values a1a2 and a2a3 are selected, then each subsequent model training, the two sets of coding values a1a2 and a2a3 still need to be selected .
  • the coding values of a1 and a2 are respectively mapped to the second negative corpus set.
  • the simplest mapping method here is that the coding values are included in the second negative corpus set. Take the remainder of the total number of negative corpus, and then concatenate the second negative corpus extracted according to the coding values of a1 and a2 to generate Bigram negative corpus, and use the generated Bigram negative corpus as the second type of negative corpus , And then add the generated Bigram negative corpus to the set of negative corpus required for training corpus. The same method can be followed to generate the second type of negative corpus corresponding to a2a3.
  • the third number of encoding values can be obtained from every three acquired encoding values (ie, a2a1a3 and a1a3a4).
  • the group code value is not limited in this embodiment.
  • the group code value a2a1a3 is selected as an example for description. It should be noted that if the third number is 1, the a2a1a3 group of coded values is selected in the first selection, then the group of coded values a2a1a3 still needs to be selected for each subsequent model training.
  • the coded values of a2, a1, and a3 can be respectively mapped to the second negative corpus set.
  • the simplest mapping method here is that the coded value is taken from the total number of negative corpus included in the second negative corpus set The remainder, then the second negative corpus extracted according to the coding values of a2, a1, and a3 are spliced to generate a Trigram negative corpus, and the generated Trigram negative corpus is used as the second type of negative corpus, and then the generated Trigram negative corpus is added to the set of negative corpus required for training corpus.
  • the purpose of reordering a1, a2, a3, and a4 is to make the generated Trigram negative corpus not include the generated Bigram negative corpus.
  • the extraction module 87 may include: a sampling value obtaining sub-module 872 and a corpus extraction sub-module 873;
  • the sampling value obtaining sub-module 872 is used to obtain the third sampling value of the negative corpus of the first type and the fourth sampling value of the negative corpus of the second type according to the coding value of each set of training corpus and the pre-learned mapping relationship ; Wherein, the above-mentioned pre-learned mapping relationship may include the remainder, which is not limited in this embodiment.
  • the corpus extraction sub-module 873 is configured to extract the negative corpus of the first type from the first negative corpus set according to the third sample value obtained by the sample value obtaining sub-module 872, and extract the negative corpus of the first type from the second negative corpus according to the fourth sample value. Extract the second type of negative corpus from the corpus.
  • the extraction module 87 may also include: a sample pair acquisition sub-module 877 and a mapping relationship learning sub-module 878;
  • the sample pair acquisition sub-module 877 is configured to acquire training sample pairs before the sample value acquisition sub-module 872 obtains the third sample value of the negative corpus of the first type.
  • the training sample pair includes the coding value of the training corpus and the corresponding The sample value of the negative corpus of; wherein the distance between the sample values of the negative corpus corresponding to the training corpus meets the preset constraint distance;
  • the mapping relationship learning sub-module 878 is configured to use the above-mentioned training sample pair to learn the mapping relationship, and the mapping relationship includes the mapping relationship between the coding value of the training corpus and the sampling value of the corresponding negative corpus.
  • the artificial intelligence-based semantic recognition device provided by the embodiment shown in FIG. 9 can be used to implement the technical solutions of the method embodiments shown in FIGS. 3 to 6 of this application. For its implementation principles and technical effects, please refer to the related descriptions in the method embodiments. .
  • each step of the above method or each of the above modules may be completed by an integrated logic circuit of hardware in the processor element or instructions in the form of software.
  • the above modules may be one or more integrated circuits configured to implement the above methods, such as: one or more specific integrated circuits (Application Specific Integrated Circuit; hereinafter referred to as ASIC), or, one or more micro-processing Digital Processor (Digital Singnal Processor; hereinafter referred to as DSP), or, one or more Field Programmable Gate Array (Field Programmable Gate Array; hereinafter referred to as FPGA), etc.
  • ASIC Application Specific Integrated Circuit
  • DSP Digital Singnal Processor
  • FPGA Field Programmable Gate Array
  • these modules can be integrated together and implemented in the form of System-On-a-Chip (hereinafter referred to as SOC).
  • FIG. 10 is a schematic structural diagram of an embodiment of an artificial intelligence-based semantic recognition device of this application.
  • the above-mentioned artificial intelligence-based semantic recognition device may include: a display screen; one or more processors; a memory; multiple applications; and one or Multiple computer programs.
  • the above-mentioned display screen may include the display screen of a vehicle-mounted computer (Mobile Data Center); the above-mentioned artificial intelligence-based semantic recognition device may be a cloud server, a mobile terminal (mobile phone), a smart screen, a drone, and an intelligent network connection.
  • Vehicle Intelligent Connected Vehicle; hereinafter referred to as ICV
  • smart/intelligent car smart/intelligent car
  • on-board equipment and other equipment may be used to be a vehicle-mounted computer (Mobile Data Center);
  • ICV Intelligent Connected Vehicle
  • smart/intelligent car smart/intelligent car or on-board equipment and other equipment.
  • the one or more computer programs are stored in the memory, and the one or more computer programs include instructions.
  • the instructions When the instructions are executed by the device, the device executes the following steps: Obtain user input Query statement;
  • the query sentence is recognized by a pre-trained semantic recognition model to obtain the intention of the query sentence; wherein, the pre-trained semantic recognition model is trained using training corpus and negative corpus, and the negative corpus is based on the above training The coded value of the corpus is mapped to the negative corpus extracted from the set;
  • the foregoing device when the foregoing instructions are executed by the foregoing device, the foregoing device is caused to specifically perform the following steps:
  • the device when the instruction is executed by the device, the device specifically executes the following steps:
  • Training is performed using the training corpus, the negative corpus of the first type, and the negative corpus of the second type to obtain the semantic recognition model.
  • the device when the instruction is executed by the device, the device specifically executes the following steps:
  • the device when the instruction is executed by the device, the device specifically executes the following steps: extracting the first negative corpus from the first negative corpus according to the first sampling value. After the first negative corpus of the type, calculate the first similarity between the first negative corpus and the training corpus;
  • first similarity is less than the first similarity threshold, it is determined that the first negative corpus is successfully sampled, and the first negative corpus is added to the sampled corpus.
  • the device when the instruction is executed by the device, the device specifically executes the following steps: after calculating the first similarity between the first negative corpus and the training corpus, If the first similarity is greater than or equal to the first similarity threshold, obtaining a second sampling value according to the first sampling value;
  • the second similarity is less than the first similarity threshold, it is determined that the second negative corpus is successfully sampled, and the second negative corpus is added to the sampled corpus.
  • the device when the instruction is executed by the device, the device specifically executes the following steps: after calculating the second similarity between the second negative corpus and the training corpus, If the second similarity is greater than or equal to the first similarity threshold, repeating the step of obtaining the second sampling value according to the first sampling value and subsequent steps;
  • the number of repetitions is greater than the preset repetition number threshold, if the similarity between the negative corpus obtained by the current sampling and the training corpus is less than the second similarity threshold, it is determined that the negative corpus obtained by the current sampling is successfully sampled, Add the negative corpus obtained by the current sampling to the sampled corpus; if the similarity between the negative corpus obtained by the current sampling and the training corpus is greater than or equal to the second similarity threshold, then the negative corpus that was successfully sampled last time Add the sampled corpus again.
  • the device when the instruction is executed by the device, the device is caused to specifically perform the following steps: sequentially obtaining every M coded values from the coded values;
  • the second type of negative corpus is extracted from the second negative corpus set according to the third number of coded values; wherein, M and N are positive integers, and M ⁇ N.
  • the device when the instruction is executed by the device, the device is caused to specifically perform the following steps: obtain the first training corpus according to the coding value of each set of training corpus and the pre-learned mapping relationship The third sampling value of the negative corpus of the first type and the fourth sampling value of the negative corpus of the second type;
  • the device when the instruction is executed by the device, the device is caused to specifically perform the following steps: obtaining the pre-learned mapping relationship according to the coding value of each set of training corpus Before the third sampling value of the negative corpus of the first type and the fourth sampling value of the negative corpus of the second type, a training sample pair is obtained, and the training sample pair includes the coding value of the training corpus and the corresponding negative corpus. Sampling values; wherein the distance between the sampling values of the negative corpus corresponding to the training corpus satisfies a preset constraint distance;
  • the training sample pair is used to learn the mapping relationship, and the mapping relationship includes the mapping relationship between the coding value of the training corpus and the sampling value of the corresponding negative corpus.
  • the artificial intelligence-based semantic recognition device shown in FIG. 10 may be an electronic device or a circuit device built in the above-mentioned electronic device.
  • the electronic device can be a cloud server, a mobile terminal (mobile phone), a smart screen, a drone, an ICV, a smart (automobile) vehicle or a vehicle-mounted device, etc.
  • the aforementioned artificial intelligence-based semantic recognition device can be used to execute the functions/steps in the methods provided in the embodiments shown in FIG. 3 to FIG. 7 of this application.
  • the artificial intelligence-based semantic recognition device 900 includes a processor 910 and a transceiver 920.
  • the artificial intelligence-based semantic recognition device 900 may further include a memory 930.
  • the processor 910, the transceiver 920, and the memory 930 can communicate with each other through an internal connection path to transfer control and/or data signals.
  • the memory 930 is used to store computer programs, and the processor 910 is used to download from the memory 930. Call and run the computer program.
  • the artificial intelligence-based semantic recognition device 900 may further include an antenna 940 for transmitting the wireless signal output by the transceiver 920.
  • the above-mentioned processor 910 and the memory 930 may be integrated into a processing device, and more commonly, are components independent of each other.
  • the processor 910 is configured to execute the program code stored in the memory 930 to implement the above-mentioned functions.
  • the memory 930 may also be integrated in the processor 910, or independent of the processor 910.
  • the artificial intelligence-based semantic recognition device 900 may also include an input unit 960, a display unit 970, an audio circuit 980, a camera 990, a sensor 901, etc.
  • the audio circuit may also include a speaker 982, a microphone 984, and the like.
  • the display unit 970 may include a display screen.
  • the aforementioned artificial intelligence-based semantic recognition device 900 may further include a power supply 950 for supplying power to various devices or circuits in the artificial intelligence-based semantic recognition device 900.
  • the artificial intelligence-based semantic recognition device 900 shown in FIG. 10 can implement various processes of the methods provided in the embodiments shown in FIG. 3 to FIG. 7.
  • the operations and/or functions of the various modules in the artificial intelligence-based semantic recognition device 900 are respectively intended to implement the corresponding processes in the foregoing method embodiments.
  • the processor 910 in the artificial intelligence-based semantic recognition device 900 shown in FIG. 10 may be a system-on-chip SOC, and the processor 910 may include a central processing unit (Central Processing Unit; hereinafter referred to as CPU), or It further includes other types of processors, such as: graphics processing unit (Graphics Processing Unit; hereinafter referred to as GPU) and so on.
  • CPU Central Processing Unit
  • GPU Graphics Processing Unit
  • each part of the processor or processing unit inside the processor 910 can cooperate to implement the previous method flow, and the corresponding software program of each part of the processor or processing unit can be stored in the memory 930.
  • the processors involved may include, for example, CPU, DSP, microcontroller or digital signal processor, and may also include GPU, embedded neural network processor (Neural-network Process Units; hereinafter referred to as NPU) and Image signal processing (Image Signal Processing; hereinafter referred to as ISP), which may also include necessary hardware accelerators or logic processing hardware circuits, such as ASIC, or one or more integrated circuits used to control the execution of the technical solutions of this application Circuit etc.
  • the processor may have a function of operating one or more software programs, and the software programs may be stored in a storage medium.
  • the embodiment of the present application also provides a computer-readable storage medium in which a computer program is stored, and when it is run on a computer, the computer can execute Figure 3, Figure 4, Figure 5, and Figure 6 of the present application. Or the method provided by the embodiment shown in FIG. 7.
  • the embodiments of the present application also provide a computer program product, the computer program product includes a computer program, when it runs on a computer, the computer executes the implementation shown in FIG. 3, FIG. 4, FIG. 5, FIG. 6 or FIG. 7 The method provided by the example.
  • At least one refers to one or more
  • multiple refers to two or more.
  • And/or describes the association relationship of the associated objects, indicating that there can be three types of relationships, for example, A and/or B, which can mean that A exists alone, A and B exist at the same time, and B exists alone. Among them, A and B can be singular or plural.
  • the character “/” generally indicates that the associated objects before and after are in an “or” relationship.
  • the following at least one item” and similar expressions refer to any combination of these items, including any combination of single items or plural items.
  • At least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single, or There can be more than one.
  • any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • the technical solution of the present application essentially or the part that contributes to the existing technology or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application.
  • the aforementioned storage media include: U disk, mobile hard disk, read-only memory (Read-Only Memory; hereinafter referred to as ROM), random access memory (Random Access Memory; hereinafter referred to as RAM), magnetic disks or optical disks, etc.
  • ROM read-only memory
  • RAM random access memory
  • magnetic disks or optical disks etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Artificial Intelligence (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • General Health & Medical Sciences (AREA)
  • Human Computer Interaction (AREA)
  • Acoustics & Sound (AREA)
  • Library & Information Science (AREA)
  • Machine Translation (AREA)

Abstract

本申请实施例提供一种基于人工智能的语义识别方法、装置和语义识别设备，上述基于人工智能的语义识别方法中，预先训练的语义识别模型是利用开发人员在模型训练平台，例如：Bot平台配置的训练语料及上述模型训练平台提供的负向语料进行训练的，所述负向语料是根据所述训练语料的编码值映射到负向语料集合中抽取的；从而实现了根据训练语料的编码值抽取负向语料，将负向语料生成的随机化方法变成稳定的方法，在训练语料没有添加、删除或修改的情况下，可以保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。

Description

基于人工智能的语义识别方法、装置和语义识别设备
本申请要求于2019年10月31日提交中国专利局、申请号为201911056617.4、发明名称为“基于人工智能的语义识别方法、装置和语义识别设备”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请实施例涉及人工智能中的语义识别技术领域，特别涉及基于人工智能的语义识别方法、装置和语义识别设备。
背景技术
人机对话系统是新一代的人机交互界面，Bot平台是一种模型训练平台，Bot平台为开发人员提供了一个快速能力构建的平台，支持诸如三方业务方快速构建对话的技能，用于其业务功能的交互触发。Bot平台为开发人员提供一键触发按钮可自动训练开发人员配置的技能模型。当开发人员重新训练模型时，重新训练获得的模型与之前训练获得的模型可能不一致，导致置信度波动较大，具体可以表现为上一次可以识别的语料在重新训练获得的模型下不能识别，或之前训练获得的模型无法识别的语料在重新训练获得的模型上可以识别。精心调整的语料，准确率和/或召回率等指标会有大幅波动不满足预期，这种重新训练模型导致的不稳定性最终会影响开发人员的体验。
发明内容
本申请实施例提供了一种基于人工智能的语义识别方法、装置和语义识别设备，本申请实施例还提供一种计算机可读存储介质，以实现在训练语料无添加、删除或修改的情况下，可以保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎 相同的置信度(差异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。
第一方面，本申请实施例提供了一种基于人工智能的语义识别方法，包括：获取用户输入的查询语句；
通过预先训练的语义识别模型对所述查询语句进行识别，获得所述查询语句的意图；其中，所述预先训练的语义识别模型是利用训练语料和负向语料训练的，所述负向语料是根据所述训练语料的编码值映射到负向语料集合中抽取的；其中，上述训练语料是开发人员在模型训练平台，例如：Bot平台上配置的，上述负向语料集合是上述模型训练平台提供的；
根据所述查询语句和所述查询语句的意图，获得所述查询语句对应的响应；
显示所述查询语句对应的响应。
上述基于人工智能的语义识别方法中，负向语料是根据训练语料的编码值通过映射关系抽取的，当训练语料无添加、删除或修改时，训练语料的编码值不变，并且映射关系也不会改变，因此根据训练语料的编码值通过映射关系抽取的负向语料也不会改变，由于训练语料和抽取的负向语料不变，因此利用训练语料和负向语料训练获得的模型的稳定性较高，可以实现在训练语料无添加、删除或修改的情况下，保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。
其中一种可能的实现方式中，所述获取用户输入的查询语句包括：获取用户通过文本输入的查询语句；或者，
获取用户通过语音输入的查询语句；或者，
获取用户输入的图片，对所述图片进行识别，获取所述图片中包括的查询语句。
也就是说，用户可以通过文本、语音或图片的方式输入上述查询语句。
其中一种可能的实现方式中，所述语义识别模型的训练过程包括：根据需要抽取的负向语料的数量，对训练语料进行分组；对每组训练语料进 行编码，获得每组训练语料的编码值；根据所述编码值抽取第一类型的负向语料和第二类型的负向语料，其中，上述第一类型的负向语料可以为闲聊负向语料，第二类型的负向语料可以为高频正向词汇负向语料；利用所述训练语料、所述第一类型的负向语料和所述第二类型的负向语料进行训练，获得所述语义识别模型。
具体地，在对每组训练语料进行分组之后，可以分别对每组训练语料进行编码，从而可以使每组训练语料具有唯一的编码值，编码方式可以包括：hash值或simHash值等；
进一步地，在根据需要抽取的负向语料的数量，对训练语料进行分组之前，还可以对上述训练语料进行排序。具体地，可以采用以下排序方式对上述训练语料进行排序：按照字符串、训练语料的哈希(Hash)值或者训练语料的simHash值等进行排序，当然还可以采用其他的排序方式对训练语料进行排序，本实施例不作限定。本实施例通过对训练语料进行排序，可以在训练语料完全相同的情况下，使分组后的编码值不因语料顺序的变化而有变化，保证训练语料的分组不发生变化。
本申请中，在对训练语料进行分组之后，对每组训练语料进行编码，根据上述编码值抽取第一类型的负向语料和第二类型的负向语料，然后利用上述训练语料、上述第一类型的负向语料和上述第二类型的负向语料进行训练，获得语义识别模型，从而实现了根据训练语料的编码值唯一地抽取负向语料，将负向语料生成的随机化方法变成稳定的生成方法，在训练语料没有添加、删除或修改的情况下，可以保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。
其中一种可能的实现方式中，所述根据所述编码值抽取第一类型的负向语料包括：获取第一负向语料集合中所包括的第一类型的负向语料的第一数量；其中，上述第一负向语料集合可以为闲聊负向语料集合，第一数量即为第一负向语料集合中所包括的第一类型的负向语料的总数量；
根据所述每组训练语料的编码值和所述第一数量，获得所述第一类型 的负向语料的第一采样值；具体地，根据上述每组训练语料的编码值和上述第一数量，获得上述第一类型的负向语料的第一采样值可以为：利用上述每组训练语料的编码值对上述第一数量取余数，将取余操作作为映射关系，上述余数作为上述第一采样值；
根据所述第一采样值从所述第一负向语料集合中抽取第一类型的第一负向语料。具体地，可以根据第一采样值在第一负向语料集合中进行查找，抽取标识(或索引)与上述第一采样值匹配的第一负向语料。
本实施例中，第一负向语料是根据训练语料的编码值通过映射关系抽取的，当训练语料无添加、删除或修改时，训练语料的编码值不变，并且映射关系也不会改变，因此根据训练语料的编码值通过映射关系抽取的第一负向语料也不会改变，由于训练语料和抽取的负向语料不变，因此利用训练语料和负向语料训练获得的模型的稳定性较高，可以实现在训练语料无添加、删除或修改的情况下，保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。
在一种可能的实现方式中，所述根据所述第一采样值从所述第一负向语料集合中抽取第一类型的第一负向语料之后，还包括：
计算所述第一负向语料与所述训练语料的第一相似度，这里的训练语料包括所有的正向语料，也即模型训练平台配置的所有正向训练语料；
如果所述第一相似度小于第一相似度阈值，则确定第一负向语料采样成功，将所述第一负向语料加入采样语料集合。
在一种可能的实现方式中，所述计算所述第一负向语料与所述训练语料的第一相似度之后，还包括：
如果所述第一相似度大于或等于第一相似度阈值，则根据所述第一采样值，获得第二采样值；在具体实现时，可以将上述第一采样值加上预先设定的数值，获得第二采样值；
根据所述第二采样值，从所述第一负向语料集合中抽取第一类型的第二负向语料；
计算所述第二负向语料与所述训练语料的第二相似度；
如果所述第二相似度小于第一相似度阈值，则确定所述第二负向语料采样成功，将所述第二负向语料加入所述采样语料集合。
在一种可能的实现方式中，所述计算所述第二负向语料与所述训练语料的第二相似度之后，还包括：
如果所述第二相似度大于或等于第一相似度阈值，则重复执行所述根据所述第一采样值，获得第二采样值的步骤及后续步骤；
当重复执行的次数大于预先设定的重复次数阈值时，如果当前采样获得的负向语料与所述训练语料的相似度小于第二相似度阈值，则确定当前采样获得的负向语料采样成功，将当前采样获得的负向语料加入所述采样语料集合；如果当前采样获得的负向语料与所述训练语料的相似度大于或等于第二相似度阈值，则将上一次采样成功的负向语料再次加入所述采样语料集合。
如果开发人员配置的训练语料与第一负向语料集合中的语料相近，那么将此语料当作负向语料，会影响训练语料的意图的识别，表现为将训练语料识别为负向意图或识别的正向意图的置信度较低，本实施例将与训练语料相似度较高的采样语料剔除，从而避免对正向意图带来的影响。本实施例可以实现将与训练语料相似度较低的负向语料添加到采样语料集合，不将与训练语料相似度高的负向语料添加到上述采样语料集合。
在一种可能的实现方式中，所述根据所述编码值抽取第二类型的负向语料包括：从所述编码值中按顺序获取每M个编码值；
从获取的每M个编码值中选择第二数量的编码值；
根据第二数量的编码值从第二负向语料集合中抽取第二类型的负向语料；
对所述编码值进行排序；
从所述排序后的编码值中按顺序获取每N个编码值；
从获取的每N个编码值中选择第三数量的编码值；
根据第三数量的编码值从第二负向语料集合中抽取第二类型的负向语料；其中，M，N为正整数，M≠N。
本实施例中，第二负向语料是根据训练语料的编码值通过映射关系抽 取的，当训练语料无添加、删除或修改时，训练语料的编码值不变，并且映射关系也不会改变，因此根据训练语料的编码值通过映射关系抽取的第二负向语料也不会改变，由于训练语料和抽取的负向语料不变，因此利用训练语料和负向语料训练获得的模型的稳定性较高，可以实现在训练语料无添加、删除或修改的情况下，保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。
在一种可能的实现方式中，所述根据所述编码值抽取第一类型的负向语料和第二类型的负向语料包括：根据所述每组训练语料的编码值和预先学习的映射关系，获得所述第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值；
根据所述第三采样值从所述第一负向语料集合中抽取第一类型的负向语料，并根据所述第四采样值从所述第二负向语料集合中抽取第二类型的负向语料。
在一种可能的实现方式中，所述根据所述每组训练语料的编码值和预先学习的映射关系，获得所述第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值之前，还包括：
获取训练样本对，所述训练样本对包括训练语料的编码值和对应的负向语料的采样值；其中，所述训练语料对应的负向语料的采样值之间的距离满足预先设定的约束距离；
利用所述训练样本对进行映射关系的学习，所述映射关系包括所述训练语料的编码值与对应的负向语料的采样值之间的映射关系。
本实现方式中，负向语料是根据训练语料的编码值通过映射关系抽取的，当训练语料无添加、删除或修改时，训练语料的编码值不变，并且映射关系也不会改变，因此根据训练语料的编码值通过映射关系抽取的负向语料也不会改变，由于训练语料和抽取的负向语料不变，因此利用训练语料和负向语料训练获得的模型的稳定性较高，可以实现在训练语料无添加、删除或修改的情况下，保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差 异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。
第二方面，本申请实施例提供一种基于人工智能的语义识别装置，包括：
获取模块，用于获取用户输入的查询语句；
识别模块，用于通过预先训练的语义识别模型对所述查询语句进行识别，获得所述查询语句的意图；其中，所述预先训练的语义识别模型是利用训练语料和负向语料训练的，所述负向语料是根据所述训练语料的编码值映射到负向语料集合中抽取的；
查询模块，用于根据所述获取模块获取的查询语句和所述识别模块识别的所述查询语句的意图，获得所述查询语句对应的响应；
显示模块，用于显示所述查询语句对应的响应。
其中一种可能的实现方式中，所述获取模块，具体用于获取用户通过文本输入的查询语句；或者，获取用户通过语音输入的查询语句；或者，获取用户输入的图片，对所述图片进行识别，获取所述图片中包括的查询语句。
其中一种可能的实现方式中，所述装置还包括：
分组模块，用于根据需要抽取的负向语料的数量，对训练语料进行分组；
编码模块，用于对每组训练语料进行编码，获得每组训练语料的编码值；
抽取模块，用于根据所述编码模块获得的编码值抽取第一类型的负向语料和第二类型的负向语料；
训练模块，用于利用所述训练语料、所述第一类型的负向语料和所述第二类型的负向语料进行训练，获得所述语义识别模型。
其中一种可能的实现方式中，所述抽取模块包括：
数量获取子模块，用于获取第一负向语料集合中所包括的第一类型的负向语料的第一数量；
采样值获得子模块，用于根据所述每组训练语料的编码值和所述第一数量，获得所述第一类型的负向语料的第一采样值；
语料抽取子模块，用于根据所述采样值获得子模块获得的第一采样值从所述第一负向语料集合中抽取第一类型的第一负向语料。
其中一种可能的实现方式中，所述抽取模块还包括：
相似度计算子模块，用于在所述语料抽取子模块抽取第一类型的第一负向语料之后，计算所述第一负向语料与所述训练语料的第一相似度；
所述语料抽取子模块，还用于当所述第一相似度小于第一相似度阈值时，确定第一负向语料采样成功，将所述第一负向语料加入采样语料集合。
其中一种可能的实现方式中，所述采样值获得子模块，还用于在所述相似度计算子模块计算第一相似度之后，如果所述第一相似度大于或等于第一相似度阈值，则根据所述第一采样值，获得第二采样值；
所述语料抽取子模块，还用于根据所述采样值获得子模块获得的第二采样值，从所述第一负向语料集合中抽取第一类型的第二负向语料；
所述相似度计算子模块，还用于计算所述第二负向语料与所述训练语料的第二相似度；
所述语料抽取子模块，还用于当所述第二相似度小于第一相似度阈值时，确定所述第二负向语料采样成功，将所述第二负向语料加入所述采样语料集合。
其中一种可能的实现方式中，所述采样值获得子模块，还用于在所述相似度计算子模块计算第二相似度之后，如果所述第二相似度大于或等于第一相似度阈值，则重复执行所述根据所述第一采样值，获得第二采样值的步骤及后续步骤；
所述语料抽取子模块，还用于当重复执行的次数大于预先设定的重复次数阈值时，如果当前采样获得的负向语料与所述训练语料的相似度小于第二相似度阈值，则确定当前采样获得的负向语料采样成功，将当前采样获得的负向语料加入所述采样语料集合；如果当前采样获得的负向语料与所述训练语料的相似度大于或等于第二相似度阈值，则将上一次采样成功的负向语料再次加入所述采样语料集合。
其中一种可能的实现方式中，所述抽取模块包括：
编码值获取子模块，用于从所述编码值中按顺序获取每M个编码值； 以及从获取的每M个编码值中选择第二数量的编码值；
语料抽取子模块，用于根据第二数量的编码值从第二负向语料集合中抽取第二类型的负向语料；
编码值排序子模块，用于对所述编码值进行排序；
所述编码值获取子模块，还用于从所述排序后的编码值中按顺序获取每N个编码值；以及从获取的每N个编码值中选择第三数量的编码值；
所述语料抽取子模块，还用于根据第三数量的编码值从第二负向语料集合中抽取第二类型的负向语料；其中，M，N为正整数，M≠N。
其中一种可能的实现方式中，所述抽取模块包括：
采样值获得子模块，用于根据所述每组训练语料的编码值和预先学习的映射关系，获得所述第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值；
语料抽取子模块，用于根据所述采样值获得子模块获得的第三采样值从所述第一负向语料集合中抽取第一类型的负向语料，并根据所述第四采样值从所述第二负向语料集合中抽取第二类型的负向语料。
其中一种可能的实现方式中，所述抽取模块还包括：
样本对获取子模块，用于在采样值获得子模块获得所述第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值之前，获取训练样本对，所述训练样本对包括训练语料的编码值和对应的负向语料的采样值；其中，所述训练语料对应的负向语料的采样值之间的距离满足预先设定的约束距离；
映射关系学习子模块，用于利用所述训练样本对进行映射关系的学习，所述映射关系包括所述训练语料的编码值与对应的负向语料的采样值之间的映射关系。
第三方面，本申请实施例提供一种基于人工智能的语义识别设备，包括：显示屏；一个或多个处理器；存储器；多个应用程序；以及一个或多个计算机程序，其中所述一个或多个计算机程序被存储在所述存储器中，所述一个或多个计算机程序包括指令，当所述指令被所述设备执行时，使得所述设备执行以下步骤：
获取用户输入的查询语句；
通过预先训练的语义识别模型对所述查询语句进行识别，获得所述查询语句的意图；其中，所述预先训练的语义识别模型是利用训练语料和负向语料训练的，所述负向语料是根据所述训练语料的编码值映射到负向语料集合中抽取的；
根据所述查询语句和所述查询语句的意图，获得所述查询语句对应的响应；
显示所述查询语句对应的响应。
其中一种可能的实现方式中，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：
获取用户通过文本输入的查询语句；或者，
获取用户通过语音输入的查询语句；或者，
获取用户输入的图片，对所述图片进行识别，获取所述图片中包括的查询语句。
其中一种可能的实现方式中，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：
根据需要抽取的负向语料的数量，对训练语料进行分组；
对每组训练语料进行编码，获得每组训练语料的编码值；
根据所述编码值抽取第一类型的负向语料和第二类型的负向语料；
利用所述训练语料、所述第一类型的负向语料和所述第二类型的负向语料进行训练，获得所述语义识别模型。
应当理解的是，本申请的第二至三方面与本申请的第一方面的技术方案一致，各方面及对应的可行实施方式所取得的有益效果相似，不再赘述。
第四方面，本申请实施例提供一种计算机可读存储介质，该计算机可读存储介质中存储有计算机程序，当其在计算机上运行时，使得计算机执行如第一方面所述的方法。
第五方面，本申请实施例提供一种计算机程序，当所述计算机程序被计算机执行时，用于执行第一方面所述的方法。
在一种可能的设计中，第五方面中的程序可以全部或者部分存储在与 处理器封装在一起的存储介质上，也可以部分或者全部存储在不与处理器封装在一起的存储器上。
附图说明
图1为现有相关技术中Bot平台生成的模型存在问题的示意图；
图2为采用本申请提供的方法训练获得的模型的置信度示意图；
图3为本申请基于人工智能的语义识别方法中语义识别模型的训练过程一个实施例的流程图；
图4为本申请基于人工智能的语义识别方法中语义识别模型的训练过程另一个实施例的流程图；
图5为本申请基于人工智能的语义识别方法中语义识别模型的训练过程再一个实施例的流程图；
图6为本申请基于人工智能的语义识别方法中语义识别模型的训练过程再一个实施例的流程图；
图7为本申请基于人工智能的语义识别方法一个实施例的流程图；
图8为本申请基于人工智能的语义识别装置一个实施例的结构示意图；
图9为本申请基于人工智能的语义识别装置另一个实施例的结构示意图；
图10为本申请基于人工智能的语义识别设备一个实施例的结构示意图。
具体实施方式
本申请的实施方式部分使用的术语仅用于对本申请的具体实施例进行解释，而非旨在限定本申请。
现有相关技术中，作为一种模型训练平台，Bot平台使用的机器学习意图识别算法在重新训练模型后，导致模型差异化的主要原因是存在两个随机因素。一个随机因素是随机对抗负向语料的生成，该种负向语料的生成方式是随机抽取式，例如：随机抽“你好啊”等闲聊语料；另外一个随 机因素是自动对抗负向语料的生成，该种负向语料的生成方式是生成式，例如：根据“卡”生成“银行卡”或“手机好卡”等负向语料。上述两种随机生成负向语料的方式会使得模型不稳定。
为此，本申请提出一种去随机化的稳定性采样方法，即将上文提到的两种生成负向语料的方式变成稳定的，同时具有随机化抽取负向语料的效果，从而使训练模型维持稳定，当开发人员的配置无修改时，利用Bot平台两次训练的模型保持一致，使同一语料在多次训练获得的模型下具有相同的预测效果。
需要说明的是，本申请中，Bot平台仅是模型训练平台的一种示例，上述模型训练平台还可以其他的平台，本申请对此不作限定。
图1为现有相关技术中Bot平台生成的模型存在问题的示意图，如图1所示，现有相关技术中，负向语料的随机生成方式存在不稳定性，导致Bot平台重新训练获得的模型与之前训练获得的模型有较大差异，表现为：
(1)准确率波动：重新训练后，分类算法置信度波动，导致预测准确率下降。精心调整的语料，不满足预期，严重影响准确率等指标。
(2)体验不一致：模型重新训练，模型的不稳定性，导致置信度波动，相同的语料在训练获得的不同模型上具有不同的结果，影响体验。
针对上述Bot平台存在的问题，本申请所要解决的技术问题是将负向语料的随机化生成方法变成稳定生成方法，在训练语料无添加、删除或修改的情况下，可以保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差异<0.01)，如图2所示，进而使准确率波动变小，从而提高开发人员的体验。图2为采用本申请提供的方法训练获得的模型的置信度示意图。
从图2中可以看出，现有相关技术中，对于查询语句“沈阳后天的天气怎么样”，采用两次训练获得的模型对上述查询语句进行识别，可以获得上述查询语句的意图为“天气查询”，但从图2中可以看出，现有相关技术中，前次训练获得的模型识别的意图的置信度为0.88，而重新训练获得的模型所识别的意图的置信度为0.78，两次训练获得的模型所识别的意图的置信度差异较大(为0.1)，导致两次训练获得的模型针对同一查询语 句给出的答案也不相同，从而影响开发人员体验。
而在采用本申请提供的方法生成负向语料之后，可以看出，两次训练获得的模型所识别的意图的置信度差异很小(为0.01)，这样，两次训练获得的模型针对同一查询语句给出的答案相同，从而可以提高开发人员体验。
下面对本申请提出的语义识别模型的训练过程进行介绍，本申请中的语义识别模型可以在模型训练平台，例如Bot平台上训练获得，上述模型训练平台可以部署在云服务器上，当然上述语义识别模型也可以在其他设备上训练获得，本实施例对训练上述语义识别模型的执行主体不作限定。
图3为本申请基于人工智能的语义识别方法中语义识别模型的训练过程一个实施例的流程图，如图3所示，可以包括：
步骤301，根据需要抽取的负向语料的数量，对训练语料进行分组。
其中，上述需要抽取的负向语料的数量可以在具体实现时，根据实现需求和/或系统性能等自行设定，本实施例对此不作限定，假设上述需要抽取的负向语料的数量为Num，那么就需要将训练语料划分为Num组，Num为正整数。
进一步地，在根据需要抽取的负向语料的数量，对训练语料进行分组之前，还可以对上述训练语料进行排序。具体地，可以采用以下排序方式对上述训练语料进行排序：按照字符串、训练语料的哈希(Hash)值或者训练语料的simHash值等进行排序，当然还可以采用其他的排序方式对训练语料进行排序，本实施例不作限定。本实施例通过对训练语料进行排序，可以在训练语料完全相同的情况下，使分组后的编码值不因语料顺序的变化而有变化，保证训练语料的分组不发生变化。
步骤302，对每组训练语料进行编码，获得每组训练语料的编码值。
具体地，本实施例中，在对每组训练语料进行分组之后，可以分别对每组训练语料进行编码，从而可以使每组训练语料具有唯一的编码值，编码方式可以包括：hash值或simHash值等，当然还可以采用其他的编码方式，本实施例不作限定。在具体实现时，本实施例可以对每组的训练语料按照N元语法(N-gram)，例如：unigram及bigranm方式，将计算的 simHash值作为每组训练语料的编码值。
步骤303，根据上述编码值抽取第一类型的负向语料和第二类型的负向语料。
其中，上述第一类型的负向语料可以为闲聊负向语料，第二类型的负向语料可以为高频正向词汇负向语料。
举例来说，闲聊负向语料可以包括“你好啊”等闲聊语料；高频正向词汇负向语料可以包括根据高频正向词汇“卡”生成的“银行卡”或“手机好卡”等负向语料，其中上述高频正向词汇包括在上述训练语料中出现频率较高的词汇。
步骤304，利用上述训练语料、上述第一类型的负向语料和上述第二类型的负向语料进行训练，获得上述语义识别模型。
本实施例中，在对训练语料进行分组之后，对每组训练语料进行编码，根据上述编码值抽取第一类型的负向语料和第二类型的负向语料，然后利用上述训练语料、上述第一类型的负向语料和上述第二类型的负向语料进行训练，获得语义识别模型，从而实现了根据训练语料的编码值唯一地抽取负向语料，将负向语料生成的随机化方法变成稳定的生成方法，在训练语料没有添加、删除或修改的情况下，可以保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。
图4为本申请基于人工智能的语义识别方法中语义识别模型的训练过程另一个实施例的流程图，如图4所示，本申请图3所示实施例中，步骤303中的根据上述编码值抽取第一类型的负向语料可以包括：
步骤401，获取第一负向语料集合中所包括的第一类型的负向语料的第一数量。
其中，上述第一负向语料集合可以为闲聊负向语料集合，第一数量即为第一负向语料集合中所包括的第一类型的负向语料的总数量。
步骤402，根据上述每组训练语料的编码值和上述第一数量，获得上 述第一类型的负向语料的第一采样值。
具体地，根据上述每组训练语料的编码值和上述第一数量，获得上述第一类型的负向语料的第一采样值可以为：利用上述每组训练语料的编码值对上述第一数量取余数，将取余操作作为映射关系，上述余数作为上述第一采样值。以上只是根据上述每组训练语料的编码值和上述第一数量，获得上述第一类型的负向语料的第一采样值的一种实现方式，还可以根据上述每组训练语料的编码值和上述第一数量，采用其他的实现方式获得上述第一类型的负向语料的第一采样值，本实施例对此不作限定。
步骤403，根据上述第一采样值从第一负向语料集合中抽取第一类型的第一负向语料。
具体地，可以根据第一采样值在第一负向语料集合中进行查找，抽取标识(或索引)与上述第一采样值匹配的第一负向语料。
进一步地，步骤403之后，还可以包括：
步骤404，计算第一负向语料与上述训练语料的第一相似度。然后执行步骤405或步骤406。
具体地，在根据第一采样值抽取第一负向语料之后，需要计算第一负向语料与上述训练语料的第一相似度，这里的训练语料包括所有的正向语料，也即模型训练平台配置的所有正向训练语料。
在具体实现时，可以采用lucene算法计算第一负向语料与上述训练语料的第一相似度。
步骤405，如果第一相似度小于第一相似度阈值，则确定第一负向语料采样成功，将第一负向语料加入采样语料集合。本次流程结束。
其中，上述第一相似度阈值可以在具体实现时，根据系统性能和/或实现需求等自行设定，本实施例对上述第一相似度阈值的大小不作限定。
步骤406，如果上述第一相似度大于或等于第一相似度阈值，则根据上述第一采样值，获得第二采样值。
在具体实现时，可以将上述第一采样值加上预先设定的数值，获得第二采样值。
其中，上述预先设定的数值的大小可以在具体实现时，根据系统性能 和/或实现需求等自行设定，本实施例对上述预先设定的数值的大小不作限定。
步骤407，根据上述第二采样值，从第一负向语料集合中抽取第一类型的第二负向语料。
同样，可以根据第二采样值在第一负向语料集合中进行查找，抽取标识(或索引)与上述第二采样值匹配的第二负向语料。
步骤408，计算第二负向语料与上述训练语料的第二相似度。然后执行步骤409或步骤410。
具体地，在根据第二采样值抽取第二负向语料之后，需要计算第二负向语料与上述训练语料的第一相似度，这里的训练语料包括所有的正向语料，也即模型训练平台配置的所有正向训练语料。
步骤409，如果上述第二相似度小于第一相似度阈值，则确定第二负向语料采样成功，将上述第二负向语料加入上述采样语料集合。本次流程结束。
步骤410，如果所述第二相似度大于或等于第一相似度阈值，则重复执行步骤406～步骤409；当重复执行的次数大于预先设定的重复次数阈值时，如果当前采样获得的负向语料与所述训练语料的相似度小于第二相似度阈值，则确定当前采样获得的负向语料采样成功，将当前采样获得的负向语料加入所述采样语料集合；如果当前采样获得的负向语料与所述训练语料的相似度大于或等于第二相似度阈值，则将上一次采样成功的负向语料再次加入上述采样语料集合。
其中，上述预先设定的重复次数阈值可以在具体实现时，根据系统性能和/或实现需求等自行设定，本实施例对上述预先设定的重复次数的大小不作限定，举例来说，上述预先设定的重复次数可以为5。
上述第二相似度阈值的大小可以在具体实现时，根据系统性能和/或实现需求等自行设定，本实施例对上述第二相似度阈值的大小不作限定，只要满足第二相似度阈值大于第一相似度阈值即可。
如果开发人员配置的训练语料与第一负向语料集合中的语料相近，那么将此语料当作负向语料，会影响训练语料的意图的识别，表现为将训练 语料识别为负向意图或识别的正向意图的置信度较低，本实施例将与训练语料相似度较高的采样语料剔除，从而避免对正向意图带来的影响。本实施例可以实现将与训练语料相似度较低的负向语料添加到采样语料集合，不将与训练语料相似度高的负向语料添加到上述采样语料集合。
图5为本申请基于人工智能的语义识别方法中语义识别模型的训练过程再一个实施例的流程图，如图5所示，本申请图3所示实施例中，步骤303中的根据上述编码值抽取第二类型的负向语料，可以包括：
步骤501，从所述编码值中按顺序获取每M个编码值。
步骤502，从获取的每M个编码值中选择第二数量的编码值。
其中，第二数量可以在具体实现时自行设定，本实施例对第二数量的大小不作限定。
步骤503，根据上述第二数量的编码值从第二负向语料集合中抽取第二类型的负向语料。
其中，第二负向语料集合可以为高频正向词汇负向语料集合，上述第二负向语料集合中包括的第二类型的负向语料可以为高频词。
步骤504，对上述编码值进行排序。
步骤505，从上述排序后的编码值中按顺序获取每N个编码值。
步骤506，从获取的每N个编码值中选择第三数量的编码值。
其中，第三数量可以在具体实现时自行设定，本实施例对第三数量的大小不作限定。
步骤507，根据第三数量的编码值从第二负向语料集合中抽取第二类型的负向语料。
其中，M，N为正整数，M≠N。具体地，M，N的大小可以在具体实现时，根据系统性能和/或实现需求等自行设定，本实施例对M，N的大小不作限定，举例来说，M可以为2，N可以为3。
假设，将训练语料分为了4组，这4组训练语料的编码值分别为a1、a2、a3和a4，那么可以先从上述编码值中按顺序获取每2个编码值，即a1a2、a2a3和a3a4，然后，可以从获取的每2个编码值(a1a2、a2a3和 a3a4)中选择第二数量的编码值，这里假设第二数量为2，那么可以选择a1a2和a2a3这两组编码值，当然也可以选择a1a2和a3a4这两组编码值，本实施例对此不作限定，这里以选择a1a2和a2a3这两组编码值为例进行说明。需要说明的是，如果第二数量为2，在第一次选择时，选择了a1a2和a2a3这两组编码值，那么后续每次进行模型训练时，仍需要选择a1a2和a2a3这两组编码值。
接下来，以a1a2这组编码值为例，首先将a1和a2的编码值分别映射到第二负向语料集合中，此处最简单的映射方法为编码值对第二负向语料集合所包括的负向语料的总数取余数，然后将分别根据a1和a2的编码值抽取的第二负向语料进行拼接，生成Bigram负向语料，以生成的Bigram负向语料作为第二类型的负向语料，然后将生成的Bigram负向语料加入到训练语料所需的负向语料集合中，同样可以按照相同的办法，生成a2a3对应的第二类型负向语料。
然后将a1、a2、a3和a4重新进行排序，假设重新排序后的编码值为a2、a1、a3和a4，从上述排序后的编码值中按顺序获取每3个编码值，即a2a1a3和a1a3a4，然后，可以从获取的每3个编码值(即a2a1a3和a1a3a4)中获取第三数量的编码值，这里假设第三数量为1，那么可以选择a2a1a3这组编码值，当然也可以选择a1a3a4这组编码值，本实施例对此不作限定，这里以选择a2a1a3这组编码值为例进行说明。需要说明的是，如果第三数量为1，在第一次选择时，选择了a2a1a3这组编码值，那么后续每次进行模型训练时，仍需要选择a2a1a3这组编码值。
接下来，可以将a2、a1和a3的编码值分别映射到第二负向语料集合中，此处最简单的映射方法为编码值对第二负向语料集合所包括的负向语料的总数取余数，然后将分别根据a2、a1和a3的编码值抽取的第二负向语料进行拼接，生成Trigram负向语料，以生成的Trigram负向语料作为第二类型的负向语料，然后将生成的Trigram负向语料加入到训练语料所需的负向语料集合中，这里，将a1、a2、a3和a4重新进行排序的目的是使生成的Trigram负向语料不包括生成的Bigram负向语料。
本实施例中，第二负向语料是根据训练语料的编码值通过映射关系抽 取的，当训练语料无添加、删除或修改时，训练语料的编码值不变，并且映射关系也不会改变，因此根据训练语料的编码值通过映射关系抽取的第二负向语料也不会改变，由于训练语料和抽取的负向语料不变，因此利用训练语料和负向语料训练获得的模型的稳定性较高，可以实现在训练语料无添加、删除或修改的情况下，保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。
本实施例对第一负向语料集合和第二负向语料集合中包括的负向语料的总数不作限定，此处仅提出根据编码值采取映射方法，映射到第一负向语料集合和第二负向语料集合中保证每次采样的一致性方法。
图6为本申请基于人工智能的语义识别方法中语义识别模型的训练过程再一个实施例的流程图，如图6所示，本申请图3所示实施例中，步骤303中的根据上述编码值抽取第一类型的负向语料和第二类型的负向语料，可以包括：
步骤601，根据每组训练语料的编码值和预先学习的映射关系，获得上述第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值。
其中，上述预先学习的映射关系可以包括取余，本实施例对此不作限定。
步骤602，根据上述第三采样值从上述第一负向语料集合中抽取第一类型的负向语料，并根据上述第四采样值从上述第二负向语料集合中抽取第二类型的负向语料。
进一步地，步骤601之前，还可以包括：
步骤603，获取训练样本对，上述训练样本对包括训练语料的编码值和对应的负向语料的采样值；其中，上述训练语料对应的负向语料的采样值之间的距离满足预先设定的约束距离。
具体地，可以通过以下方式获取上述训练样本对：
基于等距映射约束，如多维缩放(Multi-dimensional Scaling；以下简 称：MDS)等方法，把度量不同空间的概率分布问题，转化为不同空间内训练样本相对位置的等距约束问题，保证两个空间的概率分布在距离意义下一致。具体地，可以基于等距约束建模，求解配对训练样本；具体方法可以采用基于核(kernel)的学习方法，在不要求显式定义的映射下，直接通过对最近邻样本，定义相对映射关系，构造训练样本对。
步骤604，利用上述训练样本对进行映射关系的学习，上述映射关系包括上述训练语料的编码值与对应的负向语料的采样值之间的映射关系。
本实施例中，按照本实施例提供的基于编码值的采样方法，从第一负向语料集合中抽取第一类型的负向语料，可以获得采样语料集合。为使采样语料集合中的k个采样语料{k’i}的分布，与第一负向语料集合中随机抽取的k个采样语料{ki}的分布相同，可以等价为求{ki}的分布与{k’i}的分布之间的KL散度(Kullback-Leibler divergence)最小。具体地，KL散度可以通过式(1)所示的公式进行计算。
(1)
式(1)中，P(i)为{ki}的分布，Q(i)为{k’i}的分布，DKL(P||Q)为{ki}的分布与{k’i}的分布之间的KL散度。
假定hash方法已选定，以上问题的一般形式化问题可以归约为：求使{k’i＝hash(x’i)}的分布与{ki}的分布之间的KL散度最小的x’i的计算方法。
下面是求解分析：
hash方法选定为simhash，嵌入到第一负向语料集合内的最简单的方法为取余数(如本申请图4所示实施例所述)。首先，对训练语料进行分组，然后计算每组训练语料的simhash值作为编码值，最大程度上，可直观使得hash(x’i)的重叠程度(overlapping)最小，simhash保证了相似性约束(基于海明距离，一般64位的simhash，海明距离小于3可认为不相似)。
考虑原始训练语料集合是有标签(label)的，映射后的负向语料集合是无label的，这就可能出现以下两种情况：1)映射后的集合元素属于不同类；2)映射后的集合元素属于同一类。这里的类是指负向语料内在结 构类别，例如负向语料本身的类别标签，或没有类别标签，通过聚类等手段确定内在结构类别。
上述两种情况可以归约为，从更一般的分布保持上，直接取余数不一定是最好的映射关系，因为不能直接逼近原始分布，同时对内容也不敏感。
因此，本实施例提出基于投影到第一负向语料集合中的距离约束：反向求解映射前的训练语料的编码值x，这个x是通过待设计的hash函数映射的“虚拟样本”。Mapping(x)得到最终的负向语料的采样值。最简单的映射关系(Mapping)为取余数，最简单的hash为JDK String的hashcode。结合经典累积分布函数(Cumulative Distribution Function；以下简称：CDF)求逆采样操作：
1)通过距离约束hash：simhash提前控制overlapping，嵌入原始先验信息距离保持映射关系；
2)通过逆问题求解：嵌入目标性质距离约束，学习映射关系。
本实施例中，负向语料是根据训练语料的编码值通过映射关系抽取的，当训练语料无添加、删除或修改时，训练语料的编码值不变，并且映射关系也不会改变，因此根据训练语料的编码值通过映射关系抽取的负向语料也不会改变，由于训练语料和抽取的负向语料不变，因此利用训练语料和负向语料训练获得的模型的稳定性较高，可以实现在训练语料无添加、删除或修改的情况下，保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。
本申请图3～图5所示实施例通过对编码值取余数映射到负向语料集合的方法，以及本申请图6所示实施例通过学习映射关系将抽取负向语料的随机方法等价为稳定方法的抽取方法，不仅适用于本申请中抽取负向语料，同样也适用于其他随机因素存在每次需要保持一致的需求的情况。
通过本申请图3～图6所示实施例提供的方法，可以获得训练好的语义识别模型，然后可以利用上述训练好的语义识别模型对输入的查询语句进行语义识别。本申请中，上述语义识别模型可以搭载在模型训练平台， 例如：Bot平台上进行语义识别，如图2所示，图2的右侧方框中显示的是，按照本申请图3～图6所示实施例提供的方法，获得训练好的语义识别模型之后，在Bot平台上利用上述训练好的语义识别模型对用户输入的查询语句“沈阳后天的天气怎么样”进行查询，获得上述查询语句的意图为“天气查询”，进而获得上述查询语句的答案“沈阳后天晴，……”，并显示上述答案的一个示例。
也可以将上述语义识别模型搭载在其他电子设备，例如：服务器或终端上进行语义识别。上述电子设备可以包括：云服务器、移动终端(手机)、智慧屏、无人机、智能网联车(Intelligent Connected Vehicle；以下简称：ICV)、智能(汽)车(smart/intelligent car)或车载设备等设备。
图7为本申请基于人工智能的语义识别方法一个实施例的流程图，如图7所示，上述基于人工智能的语义识别方法可以包括：
步骤701，获取用户输入的查询语句。
具体地，获取用户输入的查询语句可以包括：
获取用户通过文本输入的查询语句；或者，
获取用户通过语音输入的查询语句；或者，
获取用户输入的图片，对上述图片进行识别，获取上述图片中包括的查询语句。
也就是说，用户可以通过文本、语音或图片的方式输入上述查询语句。
步骤702，通过预先训练的语义识别模型对上述查询语句进行识别，获得上述查询语句的意图；其中，上述预先训练的语义识别模型是利用训练语料和负向语料训练的，上述负向语料是根据上述训练语料的编码值映射到负向语料集合中抽取的。
其中，上述预先训练的语义识别模型是按照本申请图3～图6所示实施例提供的方法训练的，在此不再赘述。
步骤703，根据上述查询语句和上述查询语句的意图，获得上述查询语句对应的响应。
步骤704，显示上述查询语句对应的响应。
上述基于人工智能的语义识别方法中，负向语料是根据训练语料的编 码值通过映射关系抽取的，当训练语料无添加、删除或修改时，训练语料的编码值不变，并且映射关系也不会改变，因此根据训练语料的编码值通过映射关系抽取的负向语料也不会改变，由于训练语料和抽取的负向语料不变，因此利用训练语料和负向语料训练获得的模型的稳定性较高，可以实现在训练语料无添加、删除或修改的情况下，保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。
图8为本申请基于人工智能的语义识别装置一个实施例的结构示意图，如图8所示，上述基于人工智能的语义识别装置80可以包括：获取模块81、识别模块82、查询模块83和显示模块84。应当理解的是，基于人工智能的语义识别装置80可以对应于图10的设备900。其中，获取模块81、识别模块82和查询模块83的功能可以通过图10的设备900中的处理器910实现；显示模块84具体可以对应于图10的设备900中的显示单元970。
其中，获取模块81，用于获取用户输入的查询语句；本实施例中，获取模块81，具体用于获取用户通过文本输入的查询语句；或者，获取用户通过语音输入的查询语句；或者，获取用户输入的图片，对上述图片进行识别，获取上述图片中包括的查询语句。
也就是说，用户可以通过文本、语音或图片的方式输入上述查询语句。
识别模块82，用于通过预先训练的语义识别模型对上述查询语句进行识别，获得上述查询语句的意图；其中，上述预先训练的语义识别模型是利用训练语料和负向语料训练的，上述负向语料是根据所述训练语料的编码值映射到负向语料集合中抽取的；
查询模块83，用于根据获取模块81获取的查询语句和识别模块82识别的上述查询语句的意图，获得上述查询语句对应的响应；
显示模块84，用于显示上述查询语句对应的响应。
图8所示实施例提供的基于人工智能的语义识别装置可用于执行本 申请图7所示方法实施例的技术方案，其实现原理和技术效果可以进一步参考方法实施例中的相关描述。
图9为本申请基于人工智能的语义识别装置另一个实施例的结构示意图，与图8所示的基于人工智能的语义识别装置相比，不同之处在于，图9所示的基于人工智能的语义识别装置90还可以包括：分组模块85、编码模块86、抽取模块87和训练模块88；应当理解的是，基于人工智能的语义识别装置90可以对应于图10的设备900。其中，获取模块81、识别模块82和查询模块83的功能可以通过图10的设备900中的处理器910实现；显示模块84具体可以对应于图10的设备900中的显示单元970；分组模块85、编码模块86、抽取模块87和训练模块88的功能可以通过图10的设备900中的处理器910实现。
其中，分组模块85，用于根据需要抽取的负向语料的数量，对训练语料进行分组；其中，上述需要抽取的负向语料的数量可以在具体实现时，根据实现需求和/或系统性能等自行设定，本实施例对此不作限定，假设上述需要抽取的负向语料的数量为Num，那么就需要将训练语料划分为Num组，Num为正整数。
进一步地，基于人工智能的语义识别装置80还可以包括：排序模块89；
排序模块89，用于在分组模块85根据需要抽取的负向语料的数量，对训练语料进行分组之前，对上述训练语料进行排序。具体地，排序模块89可以采用以下排序方式对上述训练语料进行排序：按照字符串、训练语料的哈希(Hash)值或者训练语料的simHash值等进行排序，当然还可以采用其他的排序方式对训练语料进行排序，本实施例不作限定。本实施例通过排序模块89对训练语料进行排序，可以在训练语料完全相同的情况下，使分组后的编码值不因语料顺序的变化而有变化，保证训练语料的分组不发生变化。
编码模块86，用于对每组训练语料进行编码，获得每组训练语料的编码值；具体地，本实施例中，在分组模块85对每组训练语料进行分组 之后，编码模块86可以分别对每组训练语料进行编码，从而可以使每组训练语料具有唯一的编码值，编码方式可以包括：hash值或simHash值等，当然还可以采用其他的编码方式，本实施例不作限定。在具体实现时，编码模块86可以对每组的训练语料按照N元语法(N-gram)，例如：unigram及bigranm方式，将计算的simHash值作为每组训练语料的编码值。
抽取模块87，用于根据编码模块86获得的编码值抽取第一类型的负向语料和第二类型的负向语料；其中，上述第一类型的负向语料可以为闲聊负向语料，第二类型的负向语料可以为高频正向词汇负向语料。
训练模块88，用于利用上述训练语料、上述第一类型的负向语料和上述第二类型的负向语料进行训练，获得上述语义识别模型。
本实施例中，在分组模块85对训练语料进行分组之后，编码模块86对每组训练语料进行编码，抽取模块87根据上述编码值抽取第一类型的负向语料和第二类型的负向语料，然后训练模块88利用上述训练语料、上述第一类型的负向语料和上述第二类型的负向语料进行训练，获得语义识别模型，从而实现了根据训练语料的编码值唯一地抽取负向语料，将负向语料生成的随机化方法变成稳定的生成方法，在训练语料没有添加、删除或修改的情况下，可以保持两次及多次训练获得的模型无明显差异，使得开发人员的测试语料在多次训练获得的模型中具有几乎相同的置信度(差异<0.01)，进而使准确率波动变小，从而提高开发人员的体验。
本实施例中，抽取模块87可以包括：数量获取子模块871、采样值获得子模块872和语料抽取子模块873；
数量获取子模块871，用于获取第一负向语料集合中所包括的第一类型的负向语料的第一数量；其中，上述第一负向语料集合可以为闲聊负向语料集合，第一数量即为第一负向语料集合中所包括的第一类型的负向语料的总数量。
采样值获得子模块872，用于根据上述每组训练语料的编码值和上述第一数量，获得第一类型的负向语料的第一采样值；具体地，根据上述每组训练语料的编码值和上述第一数量，获得上述第一类型的负向语料的第 一采样值可以为：采样值获得子模块872利用上述每组训练语料的编码值对上述第一数量取余数，将取余操作作为映射关系，上述余数作为上述第一采样值。以上只是根据上述每组训练语料的编码值和上述第一数量，获得上述第一类型的负向语料的第一采样值的一种实现方式，采样值获得子模块872还可以根据上述每组训练语料的编码值和上述第一数量，采用其他的实现方式获得上述第一类型的负向语料的第一采样值，本实施例对此不作限定。
语料抽取子模块873，用于根据采样值获得子模块872获得的第一采样值从上述第一负向语料集合中抽取第一类型的第一负向语料。
具体地，语料抽取子模块873可以根据第一采样值在第一负向语料集合中进行查找，抽取标识(或索引)与上述第一采样值匹配的第一负向语料。
进一步地，上述抽取模块87还可以包括：相似度计算子模块874；
相似度计算子模块874，用于在语料抽取子模块873抽取第一类型的第一负向语料之后，计算第一负向语料与上述训练语料的第一相似度；具体地，在语料抽取子模块873根据第一采样值抽取第一负向语料之后，相似度计算子模块874需要计算第一负向语料与上述训练语料的第一相似度，这里的训练语料包括所有的正向语料，也即模型训练平台配置的所有正向训练语料。
语料抽取子模块873，还用于当第一相似度小于第一相似度阈值时，确定第一负向语料采样成功，将上述第一负向语料加入采样语料集合。其中，上述第一相似度阈值可以在具体实现时，根据系统性能和/或实现需求等自行设定，本实施例对上述第一相似度阈值的大小不作限定。
本实施例中，采样值获得子模块872，还用于在相似度计算子模块874计算第一相似度之后，如果上述第一相似度大于或等于第一相似度阈值，则根据上述第一采样值，获得第二采样值；在具体实现时，采样值获得子模块872可以将上述第一采样值加上预先设定的数值，获得第二采样值。
其中，上述预先设定的数值的大小可以在具体实现时，根据系统性能和/或实现需求等自行设定，本实施例对上述预先设定的数值的大小不作 限定。
语料抽取子模块873，还用于根据采样值获得子模块872获得的第二采样值，从第一负向语料集合中抽取第一类型的第二负向语料；同样，语料抽取子模块873可以根据第二采样值在第一负向语料集合中进行查找，抽取标识(或索引)与上述第二采样值匹配的第二负向语料。
相似度计算子模块874，还用于计算上述第二负向语料与上述训练语料的第二相似度；具体地，在语料抽取子模块873根据第二采样值抽取第二负向语料之后，相似度计算子模块874需要计算第二负向语料与上述训练语料的第一相似度，这里的训练语料包括所有的正向语料，也即模型训练平台配置的所有正向训练语料。
语料抽取子模块873，还用于当第二相似度小于第一相似度阈值时，确定第二负向语料采样成功，将第二负向语料加入上述采样语料集合。
采样值获得子模块872，还用于在相似度计算子模块874计算第二相似度之后，如果上述第二相似度大于或等于第一相似度阈值，则重复执行根据第一采样值，获得第二采样值的步骤及后续步骤；
语料抽取子模块873，还用于当重复执行的次数大于预先设定的重复次数阈值时，如果当前采样获得的负向语料与上述训练语料的相似度小于第二相似度阈值，则确定当前采样获得的负向语料采样成功，将当前采样获得的负向语料加入上述采样语料集合；如果当前采样获得的负向语料与上述训练语料的相似度大于或等于第二相似度阈值，则将上一次采样成功的负向语料再次加入上述采样语料集合。
其中，上述预先设定的重复次数阈值可以在具体实现时，根据系统性能和/或实现需求等自行设定，本实施例对上述预先设定的重复次数的大小不作限定，举例来说，上述预先设定的重复次数可以为5。
上述第二相似度阈值的大小可以在具体实现时，根据系统性能和/或实现需求等自行设定，本实施例对上述第二相似度阈值的大小不作限定，只要满足第二相似度阈值大于第一相似度阈值即可。
如果开发人员配置的训练语料与第一负向语料集合中的语料相近，那么将此语料当作负向语料，会影响训练语料的意图的识别，表现为将训练 语料识别为负向意图或识别的正向意图的置信度较低，本实施例将与训练语料相似度较高的采样语料剔除，从而避免对正向意图带来的影响。本实施例可以实现将与训练语料相似度较低的负向语料添加到采样语料集合，不将与训练语料相似度高的负向语料添加到上述采样语料集合。
本实施例中，抽取模块87可以包括：编码值获取子模块875、语料抽取子模块873和编码值排序子模块876；
编码值获取子模块875，用于从上述编码值中按顺序获取每M个编码值；以及从获取的每M个编码值中选择第二数量的编码值；
语料抽取子模块873，用于根据第二数量的编码值从第二负向语料集合中抽取第二类型的负向语料；其中，第二负向语料集合可以为高频正向词汇负向语料集合，上述第二负向语料集合中包括的第二类型的负向语料可以为高频词。
编码值排序子模块876，用于对上述编码值进行排序；
编码值获取子模块875，还用于从上述排序后的编码值中按顺序获取每N个编码值；以及从获取的每N个编码值中选择第三数量的编码值；
语料抽取子模块873，还用于根据第三数量的编码值从第二负向语料集合中抽取第二类型的负向语料；其中，M，N为正整数，M≠N。
具体地，M，N的大小可以在具体实现时，根据系统性能和/或实现需求等自行设定，本实施例对M，N的大小不作限定，举例来说，M可以为2，N可以为3。
假设，将训练语料分为了4组，这4组训练语料的编码值分别为a1、a2、a3和a4，那么可以先从上述编码值中按顺序获取每2个编码值，即a1a2、a2a3和a3a4，然后，可以从获取的每2个编码值(a1a2、a2a3和a3a4)中选择第二数量的编码值，这里假设第二数量为2，那么可以选择a1a2和a2a3这两组编码值，当然也可以选择a1a2和a3a4这两组编码值，本实施例对此不作限定，这里以选择a1a2和a2a3这两组编码值为例进行说明。需要说明的是，如果第二数量为2，在第一次选择时，选择了a1a2和a2a3这两组编码值，那么后续每次进行模型训练时，仍需要选择a1a2和a2a3这两组编码值。
接下来，以a1a2这组编码值为例，首先将a1和a2的编码值分别映射到第二负向语料集合中，此处最简单的映射方法为编码值对第二负向语料集合所包括的负向语料的总数取余数，然后将分别根据a1和a2的编码值抽取的第二负向语料进行拼接，生成Bigram负向语料，以生成的Bigram负向语料作为第二类型的负向语料，然后将生成的Bigram负向语料加入到训练语料所需的负向语料集合中，同样可以按照相同的办法，生成a2a3对应的第二类型负向语料。
然后将a1、a2、a3和a4重新进行排序，假设重新排序后的编码值为a2、a1、a3和a4，从上述排序后的编码值中按顺序获取每3个编码值，即a2a1a3和a1a3a4，然后，可以从获取的每3个编码值(即a2a1a3和a1a3a4)中获取第三数量的编码值，这里假设第三数量为1，那么可以选择a2a1a3这组编码值，当然也可以选择a1a3a4这组编码值，本实施例对此不作限定，这里以选择a2a1a3这组编码值为例进行说明。需要说明的是，如果第三数量为1，在第一次选择时，选择了a2a1a3这组编码值，那么后续每次进行模型训练时，仍需要选择a2a1a3这组编码值。
接下来，可以将a2、a1和a3的编码值分别映射到第二负向语料集合中，此处最简单的映射方法为编码值对第二负向语料集合所包括的负向语料的总数取余数，然后将分别根据a2、a1和a3的编码值抽取的第二负向语料进行拼接，生成Trigram负向语料，以生成的Trigram负向语料作为第二类型的负向语料，然后将生成的Trigram负向语料加入到训练语料所需的负向语料集合中，这里，将a1、a2、a3和a4重新进行排序的目的是使生成的Trigram负向语料不包括生成的Bigram负向语料。
本实施例中，抽取模块87可以包括：采样值获得子模块872和语料抽取子模块873；
采样值获得子模块872，用于根据每组训练语料的编码值和预先学习的映射关系，获得第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值；其中，上述预先学习的映射关系可以包括取余，本实施例对此不作限定。
语料抽取子模块873，用于根据采样值获得子模块872获得的第三采 样值从上述第一负向语料集合中抽取第一类型的负向语料，并根据上述第四采样值从第二负向语料集合中抽取第二类型的负向语料。
进一步地，抽取模块87还可以包括：样本对获取子模块877和映射关系学习子模块878；
其中，样本对获取子模块877，用于在采样值获得子模块872获得第一类型的负向语料的第三采样值之前，获取训练样本对，上述训练样本对包括训练语料的编码值和对应的负向语料的采样值；其中，上述训练语料对应的负向语料的采样值之间的距离满足预先设定的约束距离；
映射关系学习子模块878，用于利用上述训练样本对进行映射关系的学习，上述映射关系包括所述训练语料的编码值与对应的负向语料的采样值之间的映射关系。
图9所示实施例提供的基于人工智能的语义识别装置可用于执行本申请图3～图6所示方法实施例的技术方案，其实现原理和技术效果可以进一步参考方法实施例中的相关描述。
应理解以上图8～图9所示的基于人工智能的语义识别装置的各个模块的划分仅仅是一种逻辑功能的划分，实际实现时可以全部或部分集成到一个物理实体上，也可以物理上分开。且这些模块可以全部以软件通过处理元件调用的形式实现；也可以全部以硬件的形式实现；还可以部分模块以软件通过处理元件调用的形式实现，部分模块通过硬件的形式实现。例如，模块可以为单独设立的处理元件，也可以集成在电子设备的某一个芯片中实现。其它模块的实现与之类似。此外这些模块全部或部分可以集成在一起，也可以独立实现。在实现过程中，上述方法的各步骤或以上各个模块可以通过处理器元件中的硬件的集成逻辑电路或者软件形式的指令完成。
例如，以上这些模块可以是被配置成实施以上方法的一个或多个集成电路，例如：一个或多个特定集成电路(Application Specific Integrated Circuit；以下简称：ASIC)，或，一个或多个微处理器(Digital Singnal Processor；以下简称：DSP)，或，一个或者多个现场可编程门阵列(Field Programmable Gate Array；以下简称：FPGA)等。再如，这些模块可以 集成在一起，以片上系统(System-On-a-Chip；以下简称：SOC)的形式实现。
图10为本申请基于人工智能的语义识别设备一个实施例的结构示意图，上述基于人工智能的语义识别设备可以包括：显示屏；一个或多个处理器；存储器；多个应用程序；以及一个或多个计算机程序。
其中，上述显示屏可以包括车载计算机(移动数据中心Mobile Data Center)的显示屏；上述基于人工智能的语义识别设备可以为云服务器、移动终端(手机)、智慧屏、无人机、智能网联车(Intelligent Connected Vehicle；以下简称：ICV)、智能(汽)车(smart/intelligent car)或车载设备等设备。
其中所述一个或多个计算机程序被存储在所述存储器中，所述一个或多个计算机程序包括指令，当所述指令被所述设备执行时，使得所述设备执行以下步骤：获取用户输入的查询语句；
通过预先训练的语义识别模型对所述查询语句进行识别，获得上述查询语句的意图；其中，上述预先训练的语义识别模型是利用训练语料和负向语料训练的，上述负向语料是根据上述训练语料的编码值映射到负向语料集合中抽取的；
根据上述查询语句和上述查询语句的意图，获得上述查询语句对应的响应；
显示上述查询语句对应的响应。
在一种可能的实现方式中，当上述指令被上述设备执行时，使得上述设备具体执行以下步骤：
获取用户通过文本输入的查询语句；或者，
获取用户通过语音输入的查询语句；或者，
获取用户输入的图片，对所述图片进行识别，获取所述图片中包括的查询语句。
在一种可能的实现方式中，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：
根据需要抽取的负向语料的数量，对训练语料进行分组；
对每组训练语料进行编码，获得每组训练语料的编码值；
根据所述编码值抽取第一类型的负向语料和第二类型的负向语料；
利用所述训练语料、所述第一类型的负向语料和所述第二类型的负向语料进行训练，获得所述语义识别模型。
在一种可能的实现方式中，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：
获取第一负向语料集合中所包括的第一类型的负向语料的第一数量；
根据所述每组训练语料的编码值和所述第一数量，获得所述第一类型的负向语料的第一采样值；
根据所述第一采样值从所述第一负向语料集合中抽取第一类型的第一负向语料。
在一种可能的实现方式中，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：在根据所述第一采样值从所述第一负向语料集合中抽取第一类型的第一负向语料之后，计算所述第一负向语料与所述训练语料的第一相似度；
如果所述第一相似度小于第一相似度阈值，则确定第一负向语料采样成功，将所述第一负向语料加入采样语料集合。
在一种可能的实现方式中，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：在计算所述第一负向语料与所述训练语料的第一相似度之后，如果所述第一相似度大于或等于第一相似度阈值，则根据所述第一采样值，获得第二采样值；
根据所述第二采样值，从所述第一负向语料集合中抽取第一类型的第二负向语料；
计算所述第二负向语料与所述训练语料的第二相似度；
如果所述第二相似度小于第一相似度阈值，则确定所述第二负向语料采样成功，将所述第二负向语料加入所述采样语料集合。
在一种可能的实现方式中，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：在计算所述第二负向语料与所述训练语料的第二 相似度之后，如果所述第二相似度大于或等于第一相似度阈值，则重复执行所述根据所述第一采样值，获得第二采样值的步骤及后续步骤；
当重复执行的次数大于预先设定的重复次数阈值时，如果当前采样获得的负向语料与所述训练语料的相似度小于第二相似度阈值，则确定当前采样获得的负向语料采样成功，将当前采样获得的负向语料加入所述采样语料集合；如果当前采样获得的负向语料与所述训练语料的相似度大于或等于第二相似度阈值，则将上一次采样成功的负向语料再次加入所述采样语料集合。
在一种可能的实现方式中，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：从所述编码值中按顺序获取每M个编码值；
从获取的每M个编码值中选择第二数量的编码值；
根据所述第二数量的编码值从第二负向语料集合中抽取第二类型的负向语料；
对所述编码值进行排序；
从所述排序后的编码值中按顺序获取每N个编码值；
从获取的每N个编码值中选择第三数量的编码值；
根据所述第三数量的编码值从第二负向语料集合中抽取第二类型的负向语料；其中，M，N为正整数，M≠N。
在一种可能的实现方式中，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：根据所述每组训练语料的编码值和预先学习的映射关系，获得所述第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值；
根据所述第三采样值从所述第一负向语料集合中抽取第一类型的负向语料，并根据所述第四采样值从所述第二负向语料集合中抽取第二类型的负向语料。
在一种可能的实现方式中，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：在根据所述每组训练语料的编码值和预先学习的映射关系，获得所述第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值之前，获取训练样本对，所述训练样本对包括训练语料 的编码值和对应的负向语料的采样值；其中，所述训练语料对应的负向语料的采样值之间的距离满足预先设定的约束距离；
利用所述训练样本对进行映射关系的学习，所述映射关系包括所述训练语料的编码值与对应的负向语料的采样值之间的映射关系。
图10所示的基于人工智能的语义识别设备可以是电子设备也可以是内置于上述电子设备的电路设备。该电子设备可以为云服务器、移动终端(手机)、智慧屏、无人机、ICV、智能(汽)车或车载设备等设备。
上述基于人工智能的语义识别设备可以用于执行本申请图3～图7所示实施例提供的方法中的功能/步骤。
如图10所示，基于人工智能的语义识别设备900包括处理器910和收发器920。可选地，该基于人工智能的语义识别设备900还可以包括存储器930。其中，处理器910、收发器920和存储器930之间可以通过内部连接通路互相通信，传递控制和/或数据信号，该存储器930用于存储计算机程序，该处理器910用于从该存储器930中调用并运行该计算机程序。
可选地，基于人工智能的语义识别设备900还可以包括天线940，用于将收发器920输出的无线信号发送出去。
上述处理器910可以和存储器930可以合成一个处理装置，更常见的是彼此独立的部件，处理器910用于执行存储器930中存储的程序代码来实现上述功能。具体实现时，该存储器930也可以集成在处理器910中，或者，独立于处理器910。
除此之外，为了使得基于人工智能的语义识别设备900的功能更加完善，该基于人工智能的语义识别设备900还可以包括输入单元960、显示单元970、音频电路980、摄像头990和传感器901等中的一个或多个，所述音频电路还可以包括扬声器982、麦克风984等。其中，显示单元970可以包括显示屏。
可选地，上述基于人工智能的语义识别设备900还可以包括电源950，用于给基于人工智能的语义识别设备900中的各种器件或电路提供电源。
应理解，图10所示的基于人工智能的语义识别设备900能够实现图 3～图7所示实施例提供的方法的各个过程。基于人工智能的语义识别设备900中的各个模块的操作和/或功能，分别为了实现上述方法实施例中的相应流程。具体可参见图3～图7所示方法实施例中的描述，为避免重复，此处适当省略详细描述。
应理解，图10所示的基于人工智能的语义识别设备900中的处理器910可以是片上系统SOC，该处理器910中可以包括中央处理器(Central Processing Unit；以下简称：CPU)，还可以进一步包括其他类型的处理器，例如：图像处理器(Graphics Processing Unit；以下简称：GPU)等。
总之，处理器910内部的各部分处理器或处理单元可以共同配合实现之前的方法流程，且各部分处理器或处理单元相应的软件程序可存储在存储器930中。
以上各实施例中，涉及的处理器可以例如包括CPU、DSP、微控制器或数字信号处理器，还可包括GPU、嵌入式神经网络处理器(Neural-network Process Units；以下简称：NPU)和图像信号处理器(Image Signal Processing；以下简称：ISP)，该处理器还可包括必要的硬件加速器或逻辑处理硬件电路，如ASIC，或一个或多个用于控制本申请技术方案程序执行的集成电路等。此外，处理器可以具有操作一个或多个软件程序的功能，软件程序可以存储在存储介质中。
本申请实施例还提供一种计算机可读存储介质，该计算机可读存储介质中存储有计算机程序，当其在计算机上运行时，使得计算机执行本申请图3、图4、图5、图6或图7所示实施例提供的方法。
本申请实施例还提供一种计算机程序产品，该计算机程序产品包括计算机程序，当其在计算机上运行时，使得计算机执行本申请图3、图4、图5、图6或图7所示实施例提供的方法。
本申请实施例中，“至少一个”是指一个或者多个，“多个”是指两个或两个以上。“和/或”，描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示单独存在A、同时存在A和B、单独存在B的情况。其中A，B可以是单数或者复数。字符“/”一般表示前后 关联对象是一种“或”的关系。“以下至少一项”及其类似表达，是指的这些项中的任意组合，包括单项或复数项的任意组合。例如，a，b和c中的至少一项可以表示：a，b，c，a和b，a和c，b和c或a和b和c，其中a，b，c可以是单个，也可以是多个。
本领域普通技术人员可以意识到，本文中公开的实施例中描述的各单元及算法步骤，能够以电子硬件、计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
所属领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的系统、装置和单元的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在本申请所提供的几个实施例中，任一功能如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本申请各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(Read-Only Memory；以下简称：ROM)、随机存取存储器(Random Access Memory；以下简称：RAM)、磁碟或者光盘等各种可以存储程序代码的介质。
以上所述，仅为本申请的具体实施方式，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。本申请的保护范围应以所述权利要求的保护范围为准。

Claims (24)

  1. 一种基于人工智能的语义识别方法，其特征在于，包括：
    获取用户输入的查询语句；
    通过预先训练的语义识别模型对所述查询语句进行识别，获得所述查询语句的意图；其中，所述预先训练的语义识别模型是利用训练语料和负向语料训练的，所述负向语料是根据所述训练语料的编码值映射到负向语料集合中抽取的；
    根据所述查询语句和所述查询语句的意图，获得所述查询语句对应的响应；
    显示所述查询语句对应的响应。
  2. 根据权利要求1所述的方法，其特征在于，所述获取用户输入的查询语句包括：
    获取用户通过文本输入的查询语句；或者，
    获取用户通过语音输入的查询语句；或者，
    获取用户输入的图片，对所述图片进行识别，获取所述图片中包括的查询语句。
  3. 根据权利要求1或2所述的方法，其特征在于，所述语义识别模型的训练过程包括：
    根据需要抽取的负向语料的数量，对训练语料进行分组；
    对每组训练语料进行编码，获得每组训练语料的编码值；
    根据所述编码值抽取第一类型的负向语料和第二类型的负向语料；
    利用所述训练语料、所述第一类型的负向语料和所述第二类型的负向语料进行训练，获得所述语义识别模型。
  4. 根据权利要求3所述的方法，其特征在于，所述根据所述编码值抽取第一类型的负向语料包括：
    获取第一负向语料集合中所包括的第一类型的负向语料的第一数量；
    根据所述每组训练语料的编码值和所述第一数量，获得所述第一类型的负向语料的第一采样值；
    根据所述第一采样值从所述第一负向语料集合中抽取第一类型的第一负向语料。
  5. 根据权利要求4所述的方法，其特征在于，所述根据所述第一采样值从所述第一负向语料集合中抽取第一类型的第一负向语料之后，还包括：
    计算所述第一负向语料与所述训练语料的第一相似度；
    如果所述第一相似度小于第一相似度阈值，则确定第一负向语料采样成功，将所述第一负向语料加入采样语料集合。
  6. 根据权利要求5所述的方法，其特征在于，所述计算所述第一负向语料与所述训练语料的第一相似度之后，还包括：
    如果所述第一相似度大于或等于第一相似度阈值，则根据所述第一采样值，获得第二采样值；
    根据所述第二采样值，从所述第一负向语料集合中抽取第一类型的第二负向语料；
    计算所述第二负向语料与所述训练语料的第二相似度；
    如果所述第二相似度小于第一相似度阈值，则确定所述第二负向语料采样成功，将所述第二负向语料加入所述采样语料集合。
  7. 根据权利要求6所述的方法，其特征在于，所述计算所述第二负向语料与所述训练语料的第二相似度之后，还包括：
    如果所述第二相似度大于或等于第一相似度阈值，则重复执行所述根据所述第一采样值，获得第二采样值的步骤及后续步骤；
    当重复执行的次数大于预先设定的重复次数阈值时，如果当前采样获得的负向语料与所述训练语料的相似度小于第二相似度阈值，则确定当前采样获得的负向语料采样成功，将当前采样获得的负向语料加入所述采样语料集合；如果当前采样获得的负向语料与所述训练语料的相似度大于或等于第二相似度阈值，则将上一次采样成功的负向语料再次加入所述采样语料集合。
  8. 根据权利要求3所述的方法，其特征在于，所述根据所述编码值抽取第二类型的负向语料包括：
    从所述编码值中按顺序获取每M个编码值；
    从获取的每M个编码值中选择第二数量的编码值；
    根据所述第二数量的编码值从第二负向语料集合中抽取第二类型的负向语料；
    对所述编码值进行排序；
    从所述排序后的编码值中按顺序获取每N个编码值；
    从获取的每N个编码值中选择第三数量的编码值；
    根据所述第三数量的编码值从第二负向语料集合中抽取第二类型的负向语料；其中，M，N为正整数，M≠N。
  9. 根据权利要求3所述的方法，其特征在于，所述根据所述编码值抽取第一类型的负向语料和第二类型的负向语料包括：
    根据所述每组训练语料的编码值和预先学习的映射关系，获得所述第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值；
    根据所述第三采样值从所述第一负向语料集合中抽取第一类型的负向语料，并根据所述第四采样值从所述第二负向语料集合中抽取第二类型的负向语料。
  10. 根据权利要求9所述的方法，其特征在于，所述根据所述每组训练语料的编码值和预先学习的映射关系，获得所述第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值之前，还包括：
    获取训练样本对，所述训练样本对包括训练语料的编码值和对应的负向语料的采样值；其中，所述训练语料对应的负向语料的采样值之间的距离满足预先设定的约束距离；
    利用所述训练样本对进行映射关系的学习，所述映射关系包括所述训练语料的编码值与对应的负向语料的采样值之间的映射关系。
  11. 一种基于人工智能的语义识别装置，其特征在于，包括：
    获取模块，用于获取用户输入的查询语句；
    识别模块，用于通过预先训练的语义识别模型对所述查询语句进行识别，获得所述查询语句的意图；其中，所述预先训练的语义识别模型是利 用训练语料和负向语料训练的，所述负向语料是根据所述训练语料的编码值映射到负向语料集合中抽取的；
    查询模块，用于根据所述获取模块获取的查询语句和所述识别模块识别的所述查询语句的意图，获得所述查询语句对应的响应；
    显示模块，用于显示所述查询语句对应的响应。
  12. 根据权利要求11所述的装置，其特征在于，
    所述获取模块，具体用于获取用户通过文本输入的查询语句；或者，获取用户通过语音输入的查询语句；或者，获取用户输入的图片，对所述图片进行识别，获取所述图片中包括的查询语句。
  13. 根据权利要求11或12所述的装置，其特征在于，还包括：
    分组模块，用于根据需要抽取的负向语料的数量，对训练语料进行分组；
    编码模块，用于对每组训练语料进行编码，获得每组训练语料的编码值；
    抽取模块，用于根据所述编码模块获得的编码值抽取第一类型的负向语料和第二类型的负向语料；
    训练模块，用于利用所述训练语料、所述第一类型的负向语料和所述第二类型的负向语料进行训练，获得所述语义识别模型。
  14. 根据权利要求13所述的装置，其特征在于，所述抽取模块包括：
    数量获取子模块，用于获取第一负向语料集合中所包括的第一类型的负向语料的第一数量；
    采样值获得子模块，用于根据所述每组训练语料的编码值和所述第一数量，获得所述第一类型的负向语料的第一采样值；
    语料抽取子模块，用于根据所述采样值获得子模块获得的第一采样值从所述第一负向语料集合中抽取第一类型的第一负向语料。
  15. 根据权利要求14所述的装置，其特征在于，所述抽取模块还包括：
    相似度计算子模块，用于在所述语料抽取子模块抽取第一类型的第一负向语料之后，计算所述第一负向语料与所述训练语料的第一相似度；
    所述语料抽取子模块，还用于当所述第一相似度小于第一相似度阈值时，确定第一负向语料采样成功，将所述第一负向语料加入采样语料集合。
  16. 根据权利要求15所述的装置，其特征在于，
    所述采样值获得子模块，还用于在所述相似度计算子模块计算第一相似度之后，如果所述第一相似度大于或等于第一相似度阈值，则根据所述第一采样值，获得第二采样值；
    所述语料抽取子模块，还用于根据所述采样值获得子模块获得的第二采样值，从所述第一负向语料集合中抽取第一类型的第二负向语料；
    所述相似度计算子模块，还用于计算所述第二负向语料与所述训练语料的第二相似度；
    所述语料抽取子模块，还用于当所述第二相似度小于第一相似度阈值时，确定所述第二负向语料采样成功，将所述第二负向语料加入所述采样语料集合。
  17. 根据权利要求16所述的装置，其特征在于，
    所述采样值获得子模块，还用于在所述相似度计算子模块计算第二相似度之后，如果所述第二相似度大于或等于第一相似度阈值，则重复执行所述根据所述第一采样值，获得第二采样值的步骤及后续步骤；
    所述语料抽取子模块，还用于当重复执行的次数大于预先设定的重复次数阈值时，如果当前采样获得的负向语料与所述训练语料的相似度小于第二相似度阈值，则确定当前采样获得的负向语料采样成功，将当前采样获得的负向语料加入所述采样语料集合；如果当前采样获得的负向语料与所述训练语料的相似度大于或等于第二相似度阈值，则将上一次采样成功的负向语料再次加入所述采样语料集合。
  18. 根据权利要求13所述的装置，其特征在于，所述抽取模块包括：
    编码值获取子模块，用于从所述编码值中按顺序获取每M个编码值；以及从获取的每M个编码值中选择第二数量的编码值；
    语料抽取子模块，用于根据所述第二数量的编码值从第二负向语料集合中抽取第二类型的负向语料；
    编码值排序子模块，用于对所述编码值进行排序；
    所述编码值获取子模块，还用于从所述排序后的编码值中按顺序获取每N个编码值；以及从获取的每N个编码值中选择第三数量的编码值；
    所述语料抽取子模块，还用于根据所述第三数量的编码值从第二负向语料集合中抽取第二类型的负向语料；其中，M，N为正整数，M≠N。
  19. 根据权利要求13所述的装置，其特征在于，所述抽取模块包括：
    采样值获得子模块，用于根据所述每组训练语料的编码值和预先学习的映射关系，获得所述第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值；
    语料抽取子模块，用于根据所述采样值获得子模块获得的第三采样值从所述第一负向语料集合中抽取第一类型的负向语料，并根据所述第四采样值从所述第二负向语料集合中抽取第二类型的负向语料。
  20. 根据权利要求19所述的装置，其特征在于，所述抽取模块还包括：
    样本对获取子模块，用于在采样值获得子模块获得所述第一类型的负向语料的第三采样值和第二类型的负向语料的第四采样值之前，获取训练样本对，所述训练样本对包括训练语料的编码值和对应的负向语料的采样值；其中，所述训练语料对应的负向语料的采样值之间的距离满足预先设定的约束距离；
    映射关系学习子模块，用于利用所述训练样本对进行映射关系的学习，所述映射关系包括所述训练语料的编码值与对应的负向语料的采样值之间的映射关系。
  21. 一种基于人工智能的语义识别设备，其特征在于，包括：
    显示屏；一个或多个处理器；存储器；多个应用程序；以及一个或多个计算机程序，其中所述一个或多个计算机程序被存储在所述存储器中，所述一个或多个计算机程序包括指令，当所述指令被所述设备执行时，使得所述设备执行以下步骤：
    获取用户输入的查询语句；
    通过预先训练的语义识别模型对所述查询语句进行识别，获得所述查询语句的意图；其中，所述预先训练的语义识别模型是利用训练语料和负 向语料训练的，所述负向语料是根据所述训练语料的编码值映射到负向语料集合中抽取的；
    根据所述查询语句和所述查询语句的意图，获得所述查询语句对应的响应；
    显示所述查询语句对应的响应。
  22. 根据权利要求21所述的设备，其特征在于，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：
    获取用户通过文本输入的查询语句；或者，
    获取用户通过语音输入的查询语句；或者，
    获取用户输入的图片，对所述图片进行识别，获取所述图片中包括的查询语句。
  23. 根据权利要求21或22所述的设备，其特征在于，当所述指令被所述设备执行时，使得所述设备具体执行以下步骤：
    根据需要抽取的负向语料的数量，对训练语料进行分组；
    对每组训练语料进行编码，获得每组训练语料的编码值；
    根据所述编码值抽取第一类型的负向语料和第二类型的负向语料；
    利用所述训练语料、所述第一类型的负向语料和所述第二类型的负向语料进行训练，获得所述语义识别模型。
  24. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质中存储有计算机程序，当其在计算机上运行时，使得计算机执行如权利要求1-10任一项所述的方法。
PCT/CN2020/105908 2019-10-31 2020-07-30 基于人工智能的语义识别方法、装置和语义识别设备 WO2021082570A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP20881934.2A EP4030335A4 (en) 2019-10-31 2020-07-30 ARTIFICIAL INTELLIGENCE-BASED SEMANTIC IDENTIFICATION METHOD, SEMANTIC IDENTIFICATION DEVICE AND APPARATUS
US17/771,577 US20220414340A1 (en) 2019-10-31 2020-07-30 Artificial intelligence-based semantic recognition method, apparatus, and device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911056617.4 2019-10-31
CN201911056617.4A CN112749565B (zh) 2019-10-31 2019-10-31 基于人工智能的语义识别方法、装置和语义识别设备

Publications (1)

Publication Number Publication Date
WO2021082570A1 true WO2021082570A1 (zh) 2021-05-06

Family

ID=75645490

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/105908 WO2021082570A1 (zh) 2019-10-31 2020-07-30 基于人工智能的语义识别方法、装置和语义识别设备

Country Status (4)

Country Link
US (1) US20220414340A1 (zh)
EP (1) EP4030335A4 (zh)
CN (1) CN112749565B (zh)
WO (1) WO2021082570A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20230008868A1 (en) * 2021-07-08 2023-01-12 Nippon Telegraph And Telephone Corporation User authentication device, user authentication method, and user authentication computer program

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170004204A1 (en) * 2015-07-02 2017-01-05 International Business Machines Corporation Monitoring a Corpus for Changes to Previously Provided Answers to Questions
CN106777257A (zh) * 2016-12-28 2017-05-31 厦门快商通科技股份有限公司 基于话术的智能对话模型的构建系统及方法
CN108920622A (zh) * 2018-06-29 2018-11-30 北京奇艺世纪科技有限公司 一种意图识别的训练方法、训练装置和识别装置
CN109783820A (zh) * 2019-01-18 2019-05-21 广东小天才科技有限公司 一种语义解析方法及系统
CN110162611A (zh) * 2019-04-23 2019-08-23 苏宁易购集团股份有限公司 一种智能客服应答方法及系统

Family Cites Families (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102567304B (zh) * 2010-12-24 2014-02-26 北大方正集团有限公司 一种网络不良信息的过滤方法及装置
CN105138523A (zh) * 2014-05-30 2015-12-09 富士通株式会社 在文本中确定语义关键词的方法和装置
US10824812B2 (en) * 2016-06-07 2020-11-03 International Business Machines Corporation Method and apparatus for informative training repository building in sentiment analysis model learning and customization
CN108733722B (zh) * 2017-04-24 2020-07-31 北京京东尚科信息技术有限公司 一种对话机器人自动生成方法及装置
CN108460396B (zh) * 2017-09-20 2021-10-15 腾讯科技（深圳）有限公司 负采样方法和装置
CN108108754B (zh) * 2017-12-15 2022-07-22 北京迈格威科技有限公司 重识别网络的训练、重识别方法、装置和系统
US10733982B2 (en) * 2018-01-08 2020-08-04 Apple Inc. Multi-directional dialog
CN108520343B (zh) * 2018-03-26 2022-07-19 平安科技（深圳）有限公司 风险模型训练方法、风险识别方法、装置、设备及介质
US10573298B2 (en) * 2018-04-16 2020-02-25 Google Llc Automated assistants that accommodate multiple age groups and/or vocabulary levels
CN109344395B (zh) * 2018-08-30 2022-05-20 腾讯科技（深圳）有限公司 一种数据处理方法、装置、服务器及存储介质
CN110110860B (zh) * 2019-05-06 2023-07-25 南京大学 一种用于加速机器学习训练的自适应数据采样方法
CN110276075A (zh) * 2019-06-21 2019-09-24 腾讯科技（深圳）有限公司 模型训练方法、命名实体识别方法、装置、设备及介质

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170004204A1 (en) * 2015-07-02 2017-01-05 International Business Machines Corporation Monitoring a Corpus for Changes to Previously Provided Answers to Questions
CN106777257A (zh) * 2016-12-28 2017-05-31 厦门快商通科技股份有限公司 基于话术的智能对话模型的构建系统及方法
CN108920622A (zh) * 2018-06-29 2018-11-30 北京奇艺世纪科技有限公司 一种意图识别的训练方法、训练装置和识别装置
CN109783820A (zh) * 2019-01-18 2019-05-21 广东小天才科技有限公司 一种语义解析方法及系统
CN110162611A (zh) * 2019-04-23 2019-08-23 苏宁易购集团股份有限公司 一种智能客服应答方法及系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP4030335A4

Also Published As

Publication number Publication date
CN112749565B (zh) 2024-07-16
EP4030335A1 (en) 2022-07-20
EP4030335A4 (en) 2022-11-23
US20220414340A1 (en) 2022-12-29
CN112749565A (zh) 2021-05-04

Similar Documents

Publication Publication Date Title
CN108959246B (zh) 基于改进的注意力机制的答案选择方法、装置和电子设备
CN109101620B (zh) 相似度计算方法、聚类方法、装置、存储介质及电子设备
US11301637B2 (en) Methods, devices, and systems for constructing intelligent knowledge base
CN109815336B (zh) 一种文本聚合方法及系统
CN108763510A (zh) 意图识别方法、装置、设备及存储介质
CN112528637B (zh) 文本处理模型训练方法、装置、计算机设备和存储介质
WO2018005203A1 (en) Leveraging information available in a corpus for data parsing and predicting
WO2021056710A1 (zh) 多轮问答识别方法、装置、计算机设备及存储介质
WO2020114100A1 (zh) 一种信息处理方法、装置和计算机存储介质
CN112632226B (zh) 基于法律知识图谱的语义搜索方法、装置和电子设备
WO2020151690A1 (zh) 语句生成方法、装置、设备及存储介质
WO2024098533A1 (zh) 图文双向搜索方法、装置、设备及非易失性可读存储介质
WO2021063089A1 (zh) 规则匹配方法、规则匹配装置、存储介质及电子设备
CN108763202A (zh) 识别敏感文本的方法、装置、设备及可读存储介质
WO2023040742A1 (zh) 文本数据的处理方法、神经网络的训练方法以及相关设备
CN110852066B (zh) 一种基于对抗训练机制的多语言实体关系抽取方法及系统
US20230094730A1 (en) Model training method and method for human-machine interaction
CN111126084B (zh) 数据处理方法、装置、电子设备和存储介质
CN113436614A (zh) 语音识别方法、装置、设备、系统及存储介质
CN110874408B (zh) 模型训练方法、文本识别方法、装置及计算设备
WO2021082570A1 (zh) 基于人工智能的语义识别方法、装置和语义识别设备
WO2023137903A1 (zh) 基于粗糙语义的回复语句确定方法、装置及电子设备
CN115858776A (zh) 一种变体文本分类识别方法、系统、存储介质和电子设备
CN117235205A (zh) 命名实体识别方法、装置及计算机可读存储介质
CN114756655A (zh) 数据查询方法、装置、设备及存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20881934

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2020881934

Country of ref document: EP

Effective date: 20220413

NENP Non-entry into the national phase

Ref country code: DE