Semantic understanding method and device, electronic equipment and storage medium

Abstract

The invention provides a semantic understanding method and device, electronic equipment and a storage medium. The method comprises the steps of obtaining a to-be-understood text and a tag description text; inputting the to-be-understood text and the label description text into a semantic understanding model to obtain a semantic understanding result of the to-be-understood text output by the semantic understanding model; the semantic understanding model is used for carrying out bidirectional semantic interaction on respective text representations of the to-be-understood text and the label description text to obtain respective interactive semantic representations of the to-be-understood text and the label description text, and carrying out semantic comparison and entity boundary regression based on the interactive semantic representations to obtain a semantic understanding result. According to the method and device, the electronic equipment and the storage medium, bidirectional semantic interaction and entity boundary regression are combined, and optimization and balance of semantic understanding in universality, precision and response speed are effectively guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a semantic understanding method, device, electronic device and storage medium. Background Art

[0002] With the vigorous development of artificial intelligence technology, the field of Natural Language Processing (NLP) has made great progress. Named Entity Recognition (NER), as a key task in NLP, is used to identify entities with specific meanings in texts. It plays a vital role in many application fields such as information retrieval, question-answering systems, and machine translation.

[0003] In related technologies, entity span sequences and entity categories can be generated by autoregression. Although the autoregressive generation method improves the accuracy of entity recognition to a certain extent, it affects the reasoning speed of the system and has a high computational cost. In addition, there are also solutions for entity recognition through large language models (LLM), but large language models have many parameters and consume a lot of computing resources, which makes it difficult to balance the versatility, accuracy and response speed of entity recognition in practical applications. Summary of the invention

[0004] The present invention provides a semantic understanding method, device, electronic device and storage medium to solve the defect that it is difficult to balance the universality, accuracy and response speed of entity recognition in the related art.

[0005] The present invention provides a semantic understanding method, comprising: Get the text to be understood and the label description text; Inputting the text to be understood and the label description text into a semantic understanding model, and obtaining a semantic understanding result of the text to be understood output by the semantic understanding model; The semantic understanding model is used to perform two-way semantic interaction on the text representations of the text to be understood and the label description text, respectively, to obtain the interactive semantic representations of the text to be understood and the label description text, and to perform semantic comparison and entity boundary regression based on the interactive semantic representations to obtain the semantic understanding result.

[0006] According to a semantic understanding method provided by the present invention, the inputting of the text to be understood and the label description text into a semantic understanding model to obtain a semantic understanding result of the text to be understood output by the semantic understanding model includes: Based on the bidirectional interaction module in the semantic understanding model, bidirectional semantic interaction is performed on the text representations of the text to be understood and the label description text to obtain the interactive semantic representations of the text to be understood and the label description text; Based on the semantic comparison module in the semantic understanding model, the interactive semantic representations of the text to be understood and the label description text are applied to perform semantic comparison on the text to be understood and the label description text to obtain a semantic comparison result; Based on the entity boundary regression module in the semantic understanding model, the interactive semantic representations of the text to be understood and the label description text are applied to perform entity boundary regression on the text to be understood, so as to obtain an entity boundary regression result; Based on the output module in the semantic understanding model, the semantic comparison result and the entity boundary regression result are applied to generate the semantic understanding result.

[0007] According to a semantic understanding method provided by the present invention, the text representations of the text to be understood and the label description text are subjected to bidirectional semantic interaction to obtain interactive semantic representations of the text to be understood and the label description text, including: Extracting a semantic representation of the text to be understood based on the text representation of the text to be understood; Based on the semantic representation of the text to be understood, semantic interaction is performed on the text representation of the label description text to obtain an interactive semantic representation of the label description text; Based on the interactive semantic representation of the tag description text, semantic interaction is performed on the text representation of the text to be understood to obtain the interactive semantic representation of the text to be understood.

[0008] According to a semantic understanding method provided by the present invention, the applying the interactive semantic representations of the text to be understood and the label description text to perform semantic comparison between the text to be understood and the label description text, comprises: Applying the interactive semantic representations of the text to be understood and the label description text, respectively, a semantic comparison is performed on the intentions described by the text to be understood and the label description text, and a semantic comparison is performed on each anchor point in the text to be understood and the entity type described by the label description text.

[0009] According to a semantic understanding method provided by the present invention, the method of applying the interactive semantic representations of the text to be understood and the label description text to perform entity boundary regression on the text to be understood includes: The interactive semantic representations of the text to be understood and the label description text are applied to perform entity anchor activation on each anchor point in the text to be understood, and entity boundary regression is performed on the activated anchor points.

[0010] According to a semantic understanding method provided by the present invention, the training step of the semantic understanding model includes: Get sample text; Extracting fragments from the sample text to obtain key fragments in the sample text, and describing the key fragments to obtain description text of the key fragments, wherein the fragment extraction and the fragment description are implemented based on a large language model; The sample text is used as the text to be understood during training, the description text is used as the label description text during training, and the key fragment is used as the label of the semantic understanding result during training to construct a training sample set; The semantic understanding model is trained based on the training sample set.

[0011] According to a semantic understanding method provided by the present invention, the training step further includes: Obtaining training samples for an intent recognition task and / or training samples for a named entity recognition task; The training samples are added to the training sample set.

[0012] The present invention also provides a semantic understanding device, comprising: An acquisition unit, used for acquiring the text to be understood and the label description text; A comprehension unit, used for inputting the text to be understood and the label description text into a semantic comprehension model, and obtaining a semantic comprehension result of the text to be understood output by the semantic comprehension model; The semantic understanding model is used to perform two-way semantic interaction on the text representations of the text to be understood and the label description text, respectively, to obtain the interactive semantic representations of the text to be understood and the label description text, and to perform semantic contrast learning and entity boundary regression based on the interactive semantic representations to obtain the semantic understanding result.

[0013] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, any of the semantic understanding methods described above is implemented.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the semantic understanding methods described above.

[0015] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the semantic understanding methods described above.

[0016] The semantic understanding method, device, electronic device and storage medium provided by the present invention, in the process of semantic understanding, obtains interactive semantic representations of the text to be understood and the label description text after semantic enhancement through bidirectional semantic interaction between the text to be understood and the label description text, thereby improving the accuracy of semantic understanding; and, based on the interactive semantic representations of the text to be understood and the label description text, entity boundary regression is performed, thereby reducing the computational cost and improving the response speed of semantic understanding. Combining bidirectional semantic interaction with entity boundary regression effectively ensures the optimization and balance of semantic understanding in terms of versatility, accuracy and response speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 This is one of the flow charts of the semantic understanding method provided by the present invention.

[0019] Figure 2 This is the second flowchart of the semantic understanding method provided by the present invention.

[0020] Figure 3 It is a structural schematic diagram of the semantic understanding model provided by the present invention.

[0021] Figure 4 It is a structural schematic diagram of the two-way interaction module provided by the present invention.

[0022] Figure 5 It is a schematic diagram of the model training process provided by the present invention.

[0023] Figure 6 It is a flowchart of the method for obtaining a training sample set provided by the present invention.

[0024] Figure 7 It is a structural schematic diagram of the semantic understanding device provided by the present invention.

[0025] Figure 8 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] With the vigorous development of artificial intelligence technology, the field of natural language processing has made great progress. Named Entity Recognition (NER), as a key task in natural language processing, is used to identify entities with specific meanings in texts, such as names of people, places, and institutions. It plays a vital role in many application fields such as information retrieval, question-answering systems, and machine translation.

[0028] In the related technology, there is a paradigm based on prompt learning, which guides the pre-trained model to handle specific tasks by designing templated prompts. For example, methods such as TemplateNER and LightNER generate entity span sequences and entity categories through generative models. In specific implementation, methods such as TemplateNER and LightNER convert NER tasks into template filling problems and use the BART (Bidirectional and Auto-RegressiveTransformers) model for few-shot learning. These methods use an autoregressive generation method, that is, when the model generates each entity, it needs to rely on the previously generated entity information. Although the autoregressive generation method improves the accuracy of entity recognition to a certain extent, it affects the reasoning speed of the system and has a high computational cost.

[0029] In addition, the QaNER method innovatively transforms NER questions into machine question-answering tasks, generates NER prompts for the question-answering model, and uses a small number of labeled NER examples to prompt and tune the question-answering model, achieving good zero-shot NER performance through prompting the question-answering model. QaNER has shown significant advantages in terms of insensitivity to prompt quality, robustness to hyperparameters, low-resource performance, and zero-shot capabilities. However, compared with traditional sequence labeling methods, QaNER still has problems with high computational cost and slow inference speed.

[0030] In addition, there are also solutions for named entity recognition through large language models. However, large language models have many parameters and consume a lot of computing resources, which makes it difficult to balance the versatility, accuracy and response speed of entity recognition in practical applications.

[0031] In view of the above problems, the present invention provides a semantic understanding method. Figure 1is one of the flow charts of the semantic understanding method provided by the present invention, such as Figure 1 As shown, the method includes: Step 110, obtaining the text to be understood and the label description text.

[0032] The semantic understanding here can be named entity recognition, or it can be intent recognition or key segment recognition. The text to be understood is the text that needs to be semantically understood, such as "Turn on the air conditioner, and play Xiaohong's song". The semantic understanding of the text to be understood can achieve at least one of named entity recognition, intent recognition and key segment recognition for the text to be understood. The text to be understood can be directly input by the user, or it can be obtained by voice transcription of the collected audio, or it can be obtained by collecting images through image acquisition devices such as scanners, mobile phones, cameras, etc., and performing OCR (Optical Character Recognition) on the images. The embodiments of the present invention do not specifically limit this.

[0033] The tag description text is a text used to describe the tags used in semantic understanding in natural language. For example, when performing named entity recognition, the tag description text may include descriptions of various named entity types; for another example, when performing intent recognition, the tag description text may include descriptions of various intent types. For example, the tag description text may include descriptions of multiple user-defined tag tags, such as descriptions of tags such as "equipment" and "singer".

[0034] It is understandable that applying the label description text to the semantic understanding task can help the model better understand the various labels that may be applied in the semantic understanding task, thereby effectively improving the performance of the semantic understanding task. The label description text can be pre-set, for example, it can be pre-written by humans, or pre-collected.

[0035] Step 120, inputting the text to be understood and the label description text into a semantic understanding model, and obtaining a semantic understanding result of the text to be understood output by the semantic understanding model; The semantic understanding model is used to perform two-way semantic interaction on the text representations of the text to be understood and the label description text, respectively, to obtain the interactive semantic representations of the text to be understood and the label description text, and to perform semantic comparison and entity boundary regression based on the interactive semantic representations to obtain the semantic understanding result.

[0036] Specifically, the semantic understanding model is a neural network model used to perform semantic understanding tasks. The semantic understanding model takes the text to be understood and the label description text as dual inputs and outputs the semantic understanding result. The semantic understanding model can be pre-trained.

[0037] After obtaining the text to be understood and the label description text, the text to be understood and the label description text can be input into the semantic understanding model.

[0038] Accordingly, after receiving the text to be understood and the label description text as input, the semantic understanding model can respectively perform vector representation on the text to be understood and the label description text to obtain the text representation of the text to be understood and the text representation of the label description text. Here, the text representation refers to a kind of information that represents the text with structured data (such as arrays, vectors). For example, a dual encoder structure can be set inside the semantic understanding model, where the dual encoder structure can include two encoders, one of which is used to encode the text to be understood to obtain the text representation of the text to be understood, and the other encoder is used to encode the label description text to obtain the text representation of the label description text.

[0039] Subsequently, the semantic understanding model can perform two-way semantic interaction on the text representations of the text to be understood and the label description text. The two-way semantic interaction here is used for the interaction and mutual influence of the text representation of the text to be understood and the text representation of the label description text at the semantic level, that is, in the process of two-way semantic interaction, the text representation of the text to be understood can have an effect on the text representation of the label description text, thereby, the text representation of the label description text can be updated with the help of the semantics of the text to be understood to obtain the interactive semantic representation of the label description text, in addition, the text representation of the label description text can have an effect on the text representation of the text to be understood, thereby, the text representation of the text to be understood can be updated with the help of the semantics of the label description text to obtain the interactive semantic representation of the text to be understood. In this process, the description of the label in the label description text can be more practical under the semantic conditions of the text to be understood, and the named entity or intention in the text to be understood can also be more prominent under the description of the label in the label description text, that is, through two-way semantic interaction, the semantic enhancement of the text representation of the text to be understood and the label description text can be achieved, thereby obtaining the interactive semantic representation of the text to be understood and the label description text.

[0040] Next, the semantic understanding model performs semantic comparison and entity boundary regression for the interactive semantic representations of the text to be understood and the label description text obtained by the two-way semantic interaction. That is, in the semantic understanding model, the semantic understanding task can be decoupled into two tasks: semantic comparison and entity boundary regression. The semantic comparison task can determine the labels of the same type as the text to be understood from the labels described in the label description text by comparing the similarity between the interactive semantic representations of the text to be understood and the label description text, which can be the named entity type, intent type, etc. contained in the text to be understood; the entity boundary regression task can combine the interactive semantic representations of the text to be understood and the label description text, and predict the boundary position of the entity in the text to be understood corresponding to the label type described in the label description text from the text to be understood, thereby realizing the entity location in the text to be understood. It can be understood that, by combining the results of the two tasks of semantic comparison and entity boundary regression, the semantic understanding model can obtain and output the semantic understanding results for the text to be understood. For example, the semantic understanding results for the text to be understood "Turn on the air conditioner and play Xiaohong's song" can include the intentions "operate the device" and "ask for music", as well as the device entity "air conditioner" and the singer entity "Xiaohong".

[0041] Among them, the application of entity boundary regression can, on the one hand, solve the problem of nested entity recognition; on the other hand, compared with traditional autoregressive tasks, it can reduce computational costs and accelerate the convergence of semantic understanding models.

[0042] In the method provided by the embodiment of the present invention, in the process of semantic understanding, through the two-way semantic interaction between the text to be understood and the label description text, the semantically enhanced interactive semantic representations of the text to be understood and the label description text are obtained, thereby improving the accuracy of semantic understanding; and based on the interactive semantic representations of the text to be understood and the label description text, entity boundary regression is performed, thereby reducing the computational cost and improving the response speed of semantic understanding. Combining the two-way semantic interaction with the entity boundary regression effectively ensures the optimization and balance of semantic understanding in terms of versatility, accuracy and response speed.

[0043] The method provided by the embodiment of the present invention can effectively improve the performance of multilingual human-computer interaction, multilingual information retrieval, and multilingual question-answering systems.

[0044] Based on the above embodiments, Figure 2 This is the second flow chart of the semantic understanding method provided by the present invention. Figure 2 As shown, step 120 includes: Step 121 , based on the bidirectional interaction module in the semantic understanding model, bidirectional semantic interaction is performed on the text representations of the text to be understood and the label description text to obtain interactive semantic representations of the text to be understood and the label description text.

[0045] Here, the semantic understanding model includes a two-way interaction module, which is used to achieve two-way semantic interaction between the text representations of the to-be-understood text and the label description text.

[0046] Specifically, the text representations of the text to be understood and the label description text can be input into the two-way interaction module to obtain the interactive semantic representations of the text to be understood and the label description text output by the two-way interaction module.

[0047] During this process, the bidirectional interaction module can perform bidirectional semantic interaction on the text representation of the input text to be understood and the text representation of the label description text, which is specifically reflected in updating the text representation of the label description text based on the text representation of the text to be understood to obtain an interactive semantic representation of the label description text, and updating the text representation of the text to be understood based on the text representation of the label description text to obtain an interactive semantic representation of the text to be understood.

[0048] Step 122, based on the semantic comparison module in the semantic understanding model, the interactive semantic representations of the text to be understood and the label description text are applied to perform semantic comparison on the text to be understood and the label description text to obtain a semantic comparison result.

[0049] Here, the semantic understanding model also includes a semantic comparison module, the input of which is connected to the output of the two-way interaction module. The semantic comparison module takes the interactive semantic representations of the text to be understood and the label description text as input and takes the semantic comparison result as output. The semantic comparison module is used to compare the semantic similarity between the interactive semantic representations of the text to be understood and the label description text.

[0050] In this process, the semantic comparison module can perform semantic comparison on the interactive semantic representations of the input text to be understood and the label description text. Through the semantic comparison, the similarity between the interactive semantic representations of the text to be understood and the label description text can be calculated, and then the labels described in the label description text are determined to be the labels contained in the text to be understood as the semantic comparison results. The semantic comparison results here can be whether the text to be understood contains each label, or the probability of each label being contained in the text to be understood, etc., which is not specifically limited in the embodiment of the present invention.

[0051] Step 123, based on the entity boundary regression module in the semantic understanding model, the interactive semantic representations of the text to be understood and the label description text are applied to perform entity boundary regression on the text to be understood to obtain an entity boundary regression result.

[0052] In addition, the semantic understanding model also includes an entity boundary regression module, the input of which is connected to the output of the two-way interaction module. The entity boundary regression module takes the interactive semantic representation of the text to be understood and the label description text as input, and takes the entity boundary regression result as output. The entity boundary regression module is used to predict the boundary position of the entity in the text to be understood corresponding to the label type described in the label description text, thereby realizing the entity location in the text to be understood.

[0053] In this process, the entity boundary regression module can combine the interactive semantic representations of the text to be understood and the label description text, and predict the boundary position of the entity in the text to be understood corresponding to the label type described in the label description text from the text to be understood, thereby obtaining and outputting the entity boundary regression result. The entity boundary regression result here includes the position of the entity in the text to be understood, for example, it can be the start and end position of the entity in the text to be understood.

[0054] It should be noted that step 122 and step 123 may be executed in parallel or sequentially, and the embodiment of the present invention does not specifically limit the execution order of step 122 and step 123.

[0055] Step 124, based on the output module in the semantic understanding model, apply the semantic comparison result and the entity boundary regression result to generate the semantic understanding result.

[0056] Specifically, the semantic understanding model may further include an output module, the input of which is the output of the semantic comparison module and the entity boundary regression module. The output module is used to integrate the semantic comparison and entity boundary regression results to generate a semantic understanding result of the text to be understood.

[0057] For example, Figure 3 Schematic diagram of the structure of the semantic understanding model provided by the present invention. Figure 3 As shown, the semantic structure model may include a label encoder, a text encoder, a bidirectional interaction module, a semantic comparison module, and an entity boundary regression module. In the process of semantic understanding based on the semantic understanding model: The text to be understood can be input into a text encoder for encoding, and the text encoder outputs a text representation of the text to be understood; in addition, the label description text can be input into a label encoder for encoding, and the label encoder outputs a text representation of the label description text. It should be noted that the operation of encoding the label description text based on the label encoder here can be implemented in the process of offline parameterization of the semantic understanding model, or in the process of online semantic understanding based on the semantic understanding model, and the embodiment of the present invention does not specifically limit this.

[0058] The text representation of the text to be understood output by the text encoder and the text representation of the label description text output by the label encoder are both input into the bidirectional interaction module, which performs bidirectional semantic interaction on the text representations of the two, thereby outputting the interactive semantic representations of the text to be understood and the label description text respectively.

[0059] Subsequently, the interactive semantic representations of the understood text and the label description text output by the bidirectional interaction module can be input into the semantic comparison module and the entity boundary regression module respectively. The semantic comparison module can perform semantic comparison on the interactive semantic representations of the text to be understood and the label description text and output the semantic comparison results. The entity boundary regression module can predict the boundary position of the entity in the text to be understood corresponding to the label type described in the label description text from the text to be understood based on the interactive semantic representations of the text to be understood and the label description text, thereby outputting the entity boundary regression result.

[0060] After obtaining the semantic comparison results and entity boundary regression results respectively, the semantic understanding results can be output by combining the two.

[0061] Based on any of the above embodiments, in step 121, performing bidirectional semantic interaction on the text representations of the text to be understood and the label description text to obtain interactive semantic representations of the text to be understood and the label description text, respectively, includes: Extracting a semantic representation of the text to be understood based on the text representation of the text to be understood; Based on the semantic representation of the text to be understood, semantic interaction is performed on the text representation of the label description text to obtain an interactive semantic representation of the label description text; Based on the interactive semantic representation of the tag description text, semantic interaction is performed on the text representation of the text to be understood to obtain the interactive semantic representation of the text to be understood.

[0062] Specifically, in the two-way interaction module, further semantic extraction can be performed based on the text representation of the input text to be understood, thereby obtaining the semantic representation of the text to be understood. For example, the semantic representation of the text to be understood can be a summary of the full text semantics of the text to be understood, such as the CLS vector of the text to be understood.

[0063] After obtaining the semantic representation of the text to be understood, the semantic representation of the text to be understood can be applied to influence the text representation of the label description text, that is, the text representation of the label description text is updated with the help of the semantic representation of the text to be understood, thereby obtaining the interactive semantic representation of the label semantic text. For example, for the text to be understood "I want to listen to song B by singer A", only when the semantic representation of the text to be understood is obtained, "song B" can be strongly semantically related to the label "song name", so in the embodiment of the present invention, the text representation of the label description text can be updated using the semantic representation of the entire text to be understood.

[0064] After obtaining the interactive semantic representation of the label semantic text, the interactive semantic representation of the label semantic text can also be used to influence the text representation of the text to be understood, that is, the text representation of the text to be understood is updated with the help of the interactive semantic representation of the label semantic text, thereby obtaining the interactive semantic representation of the text to be understood.

[0065] For example, Figure 4 is a schematic diagram of the structure of the two-way interaction module provided by the present invention, such as Figure 4 As shown, the CLS vector in the text representation of the text to be understood can be used as the semantic representation of the text to be understood, and the semantic representation of the text to be understood and the text representation of the label description text can be concatenated. After the concatenated vector representation passes through convolution conv and activation function sigmoid, it is input into MOE (Mixture of Experts) together with the semantic representation of the text to be understood and the text representation of the label description text for semantic interaction, and the output of MOE is concatenated with the text representation of the label description text as the interactive semantic representation of the label description text.

[0066] In addition, the text representation of the text to be understood can be weighted (the weight is denoted by W convq ) is used as the query vector of multi-head attention, and the interactive semantic representation of the label description text is weighted (the weight is recorded as W convk ) is used as the key vector of multi-head attention, and multi-head attention calculation is performed. The attention vector obtained by multi-head attention calculation is weighted with the text representation of the text to be understood to obtain weighted features. Subsequently, the weighted features are added to the text representation of the text to be understood, and the result of the addition is convolved conv to obtain the interactive semantic representation of the text to be understood.

[0067] Based on any of the above embodiments, in step 122, applying the interactive semantic representations of the text to be understood and the label description text to perform semantic comparison between the text to be understood and the label description text includes: Applying the interactive semantic representations of the text to be understood and the label description text, respectively, a semantic comparison is performed on the intentions described by the text to be understood and the label description text, and a semantic comparison is performed on each anchor point in the text to be understood and the entity type described by the label description text.

[0068] Specifically, in the tag description text, descriptions may be provided for various intents, and descriptions may also be provided for various entity types.

[0069] Accordingly, in the semantic comparison module, not only the semantic comparison of intent can be performed on the interactive semantic representations of the text to be understood and the label description text, but also the semantic comparison of entity types can be performed on the interactive semantic representations of the text to be understood and the label description text.

[0070] Moreover, when performing semantic comparison for intent, the interactive semantic representation of the entire text to be understood can be semantically compared with the interactive semantic representation of each intent in the interactive semantic representation of the label description text, thereby determining whether the intent described in the label description text exists in the text to be understood.

[0071] When performing semantic comparison for entity types, the interactive semantic representation of each anchor point in the interactive semantic representation of the text to be understood can be semantically compared with the interactive semantic representation of each entity type in the interactive semantic representation of the label description text, thereby determining whether each anchor point in the text to be understood belongs to the entity type described in the label description text. It is understandable that the anchor point in the text to be understood referred to here can specifically be a participle, phrase or sentence in the text to be understood.

[0072] It can be understood that the semantic comparison results thus obtained may include comparison results for intent and comparison results for entity types, thereby achieving intent recognition and entity recognition.

[0073] Based on any of the above embodiments, in step 123, applying the interactive semantic representations of the text to be understood and the label description text to perform entity boundary regression on the text to be understood includes: The interactive semantic representations of the text to be understood and the label description text are applied to perform entity anchor activation on each anchor point in the text to be understood, and entity boundary regression is performed on the activated anchor points.

[0074] Specifically, in the entity boundary regression module, entity anchor point activation can be first performed for each anchor point in the text to be understood, which is specifically reflected in activating the anchor points that may be named entities in the text to be understood; then, entity boundary regression is performed on the activated anchor points in the text to be understood, thereby determining the boundaries of the named entities in the text to be understood.

[0075] In an embodiment of the present invention, entity anchor point activation and entity boundary regression can be implemented with reference to the solutions in the field of image target detection. Specifically, when performing entity anchor point activation, the overlap between each anchor point in the text to be understood and the actual named entity can be calculated based on the interactive semantic representation of the text to be understood and the label description text, and the anchor points whose overlap exceeds the overlap threshold can be activated; when performing entity boundary regression, for the activated anchor point, the position of the anchor point in the text to be understood and the text length can be adjusted, so that the activated anchor point can better match the actual named entity, and the boundary of the adjusted anchor point is used as output.

[0076] In the method provided in the embodiment of the present invention, entity boundary regression is performed on the text to be understood, which can effectively combine the advantages of BIO annotation and Span annotation in the flexibility and accuracy of identifying named entity boundaries in named entity annotation, thereby alleviating the pressure of a single anchor point in semantic classification and achieving more efficient and flexible named entity recognition.

[0077] At present, traditional supervised learning-based named entity recognition methods rely on large-scale annotated datasets, which face two major challenges in a multilingual environment: first, data scarcity, i.e., annotated data in certain languages ​​or fields is difficult to obtain; second, resource consumption, i.e., building and maintaining large-scale annotated datasets requires a lot of manpower and material resources. These problems limit the application potential of named entity recognition technology in low-resource languages ​​and emerging fields, and also limit the development of multilingual semantic understanding technology.

[0078] For example, common named entity recognition methods are usually sequence labeling problems, that is, using pre-trained models such as BERT (Bidirectional Encoder Representations from Transformers) as encoders to generate sequence representations of text, and then using a classifier to map the sequence representations to predefined entity labels. This method works well in resource-rich environments, but faces the challenge of poor results when applied to multilingual environments and entity types with little or no labeled data.

[0079] In addition, there are also methods to achieve named entity recognition through cross-language transfer learning. Specifically, the model can be pre-trained in one language and fine-tuned in another language, thereby improving the performance of the model in the new language. Although this method has made progress in the task of named entity recognition with few samples, it still faces problems such as imbalanced language resources, differences in language characteristics, and the complexity of cross-language semantic alignment in a multilingual environment.

[0080] In view of the above problems, in an embodiment of the present invention, Figure 5 It is a schematic diagram of the model training process provided by the present invention, such as Figure 5 As shown, the training steps of the semantic understanding model include: Step 510, obtaining sample text.

[0081] Step 520, extracting fragments from the sample text to obtain key fragments in the sample text, and describing the key fragments to obtain description text of the key fragments. The fragment extraction and the fragment description are implemented based on a large language model.

[0082] Specifically, the sample text is a text collected in advance and can be used as a model training sample. In the embodiment of the present invention, the sample text can be an unlabeled text, and the sample text can be a text in various languages.

[0083] After obtaining the sample text, the sample text is input into the large language model, and the large language model extracts segments from the input sample text, thereby obtaining key segments in the sample text output by the large language model. The key segments here may be keywords in the sample text, or key phrases or key sentences in the sample text, which are not specifically limited in the embodiment of the present invention.

[0084] In addition, the sample text may be input into a large language model, or the key segments in the sample text may be input into the large language model, and the large language model may describe the key segments of the sample text, thereby obtaining the description text of the key segments output by the large language model. Here, the description text is used to describe the key segments.

[0085] Alternatively, the sample text may be input into a large language model, which extracts fragments from the input sample text and describes the extracted key fragments, thereby obtaining the key fragments and their description text output by the large language model. This is not specifically limited in the embodiments of the present invention.

[0086] Step 530, using the sample text as the text to be understood during training, the description text as the label description text during training, and the key fragment as the label of the semantic understanding result during training to construct a training sample set.

[0087] Specifically, the key segments of the sample text and the description text of the key segments can be obtained through step 520. In the semantic understanding scenario, the description text can be equivalent to the semantic description, and the key segments can be equivalent to the text segments. Therefore, the description text and the key segments can be abstracted into description segment pairs, that is, data pairs in the form of "semantic description-text segment".

[0088] It is understandable that in the semantic understanding scenario, key fragment extraction, intent recognition, and named entity recognition can all be abstracted into the form of "semantic description-text fragment". Among them, the description of various intents in intent recognition can be a semantic description, and the intent obtained from intent recognition can be a text fragment; the type description of various named entities in named entity recognition can be a semantic description, and the named entities obtained from named entity recognition can be a text fragment. Therefore, by constructing description fragment pairs in the form of "semantic description-text fragment", unified modeling of various semantic understanding tasks such as key fragment extraction, intent recognition, and named entity recognition can be achieved.

[0089] In this case, a training sample set can be constructed based on the sample text and the description segment pairs of the sample text. In the training sample set thus constructed, the sample text can be regarded as the text to be understood as the input for model training, the description text in the description segment pairs can be regarded as the label description text input for model training, and the key segment in the description segment pairs can be regarded as the label of the semantic understanding result output by the model training, so that the training sample set can be applied to the supervised training of the semantic understanding model.

[0090] Step 540: training the semantic understanding model based on the training sample set.

[0091] Specifically, the semantic understanding model training can be performed based on the training sample set constructed in step 530. In this process, the sample text in the training sample set and the description text in the description segment pair can be used as the training input of the semantic understanding model, and the key segment in the description segment pair can be used as the training label of the semantic understanding model, thereby realizing supervised training for the semantic understanding model.

[0092] Furthermore, in the process of training the semantic understanding model, parameter iteration can be performed based on the loss of semantic contrastive learning and the loss of entity boundary regression respectively to optimize the model effect.

[0093] In the method provided in the embodiment of the present invention, a large language model is combined to perform fragment extraction and fragment description on the sample text, thereby realizing the modeling of description fragment pairs in the form of "semantic description-text fragment", so that various semantic understanding tasks such as key fragment extraction, intent recognition, and named entity recognition can be uniformly modeled, thereby solving the problems of scarce annotated data and high acquisition cost in semantic understanding scenarios.

[0094] Based on any of the above embodiments, in step 530, the sample text is used as the text to be understood during training, the description text is used as the label description text during training, and the key segment is used as the label of the semantic understanding result during training to construct a training sample set, including: Translating the sample text and the description segment pair of the sample text to obtain a translated sample text and the description segment pair of the translated sample text, wherein the description segment pair includes a description text and a key segment; The training sample set is constructed based on the sample text, the description segment pair of the sample text, and the translation sample text, the description segment pair of the translation sample text.

[0095] Specifically, the sample text obtained in step 510 may be a sample text in a language with a better semantic understanding capability of the large language model, such as a sample text in English, Chinese, etc. Therefore, segment extraction and segment description based on the large language model can ensure the accuracy and reliability of the key segments and description texts obtained.

[0096] Considering that in a multilingual environment, the sample texts available in some minority languages ​​are limited, and large language models have poor understanding of such minority languages, after completing the fragment extraction and fragment description of sample texts in some languages ​​by large language models, the sample texts and their description fragment pairs that have been obtained in some languages ​​can be translated, thereby obtaining translated sample texts and their description fragment pairs in some minority languages.

[0097] On this basis, the sample text, the description fragment pair of the sample text, and the translated sample text, the description fragment pair of the translated sample text can be used together to construct the training sample set, thereby ensuring a balanced distribution of resources in different languages ​​in the training sample set.

[0098] For example, Figure 6 is a flow chart of the method for obtaining a training sample set provided by the present invention, such as Figure 6 As shown, unsupervised first language data may be obtained first. Here, the unsupervised first language data may be sample text in the first language, and the first language is specifically a language with better semantic understanding ability of the large language model.

[0099] Subsequently, segment extraction and segment description can be performed on the unsupervised first language data based on the large language model, thereby obtaining the key segments and description text of the sample text in the first language. Next, the language model in the first language can be used to calculate the perplexity of the key segments, thereby filtering out the key segments whose perplexity exceeds the perplexity threshold. After completing the perplexity-based data filtering, the remaining sample text in the first language and its key segments and description text can be translated into the second language. The second language here is a language with weak semantic understanding ability of the large language model, usually a minority language. The translation method here is based on the translation model, or it can be to analyze and process the named entities of the data in the first language, and then translate based on the analysis results and the translation template filled with information such as the scene and intention of the data in the first language. The embodiment of the present invention does not make specific limitations on this.

[0100] Next, for the translated sample text in the second language and its key segments and description text, the word frequencies of the segmented words in the key segments can be used to filter data, specifically filtering data with word frequencies below a word frequency threshold.

[0101] After completing the word frequency filtering, the remaining translated sample texts in the second language and their key fragments and description texts can be used as supervised second language data.

[0102] On this basis, a training sample set can be constructed based on supervised first language data and second language data.

[0103] Based on any of the above embodiments, the training step further includes: Obtaining training samples for an intent recognition task and / or training samples for a named entity recognition task; The training samples are added to the training sample set.

[0104] Specifically, in addition to extracting and describing fragments to obtain description fragment pairs of sample texts to construct training sample sets, open source data can also be collated and collected to obtain training samples for intent recognition tasks and / or named entity recognition tasks. The training samples here are annotated data, and large language models can be used to describe the labels in the training samples to obtain richer description texts.

[0105] After obtaining the training samples, the training samples can also be added to the training sample set, thereby further improving the quality of the training sample set.

[0106] Based on any of the above embodiments, the semantic understanding method may include the following steps: First, unified modeling is carried out for various semantic understanding tasks such as key fragment extraction, intent recognition, and named entity recognition. Specifically, a large language model can be used to extract and describe fragments of unsupervised sample texts, thereby constructing description fragment pairs in the form of "semantic description-text fragment", thereby realizing automatic labeling of unlabeled data. In addition, supervised data in various languages ​​can be obtained through translation, thereby solving the problem of language imbalance.

[0107] Subsequently, open source data can be collated to collect training samples for intent recognition tasks and / or named entity recognition tasks, and richer descriptive texts can be obtained through large language models to further expand the training sample set.

[0108] Next, a semantic understanding model can be constructed. The semantic understanding model here adopts a dual encoder structure to realize the encoding of the text to be understood and the label description text. On this basis, a two-way semantic interaction is performed for the text to be understood and the label description text, and the problem of nested entity extraction is solved through entity boundary regression to achieve a balance between semantic understanding accuracy and response speed. Among them, the encoder for the label description text can adopt a text similarity encoder.

[0109] Subsequently, the training sample set constructed in the above steps can be used to train and verify the semantic understanding model. In this process, the model parameters can be adjusted through semantic comparison learning loss and entity boundary regression loss.

[0110] Finally, the trained semantic recognition model can be applied to the text based on the label description, and model reasoning tasks such as intent recognition and named entity recognition can be performed on the text to be understood.

[0111] In the method provided by the embodiment of the present invention, based on two-way semantic interaction, the relationship between the text representation of the label description text and the text to be understood is deeply considered. Among them, the interactive semantic representation of the label description text needs to be updated with the help of the text representation of the text to be understood, so that the interactive semantic representation of the label description text has more practical significance under the condition of the overall semantics of the sentence. Moreover, in the method provided by the embodiment of the present invention, entity boundary regression can not only solve the problem of nested entity recognition, but also reduce the computational cost and accelerate the model convergence speed under the technical empowerment of cutting-edge computer graphics, effectively combine the advantages of BIO and SAPN entity extraction methods, alleviate the pressure on the semantic classification of a single anchor point, and play the role of a model self-voting integration, thereby making named entity recognition more flexible and efficient. Moreover, the combination of bidirectional semantic interaction and entity boundary regression can activate the anchor points in the text to be understood based on the interactive semantic representation of the label description text, further effectively improving the accuracy of entity boundary regression.

[0112] The semantic understanding device provided by the present invention is described below. The semantic understanding device described below and the semantic understanding method described above can be referenced to each other.

[0113] Figure 7 Schematic diagram of the structure of the semantic understanding device provided by the present invention. Figure 7 As shown, the device comprises: An acquisition unit 710 is used to acquire the text to be understood and the label description text; An understanding unit 720, configured to input the text to be understood and the label description text into a semantic understanding model, and obtain a semantic understanding result of the text to be understood output by the semantic understanding model; The semantic understanding model is used to perform two-way semantic interaction on the text representations of the text to be understood and the label description text, respectively, to obtain the interactive semantic representations of the text to be understood and the label description text, and to perform semantic contrast learning and entity boundary regression based on the interactive semantic representations to obtain the semantic understanding result.

[0114] In the device provided by the embodiment of the present invention, in the process of semantic understanding, the interactive semantic representations of the text to be understood and the label description text after semantic enhancement are obtained through the bidirectional semantic interaction between the text to be understood and the label description text, thereby improving the accuracy of semantic understanding; and entity boundary regression is performed based on the interactive semantic representations of the text to be understood and the label description text, thereby reducing the computational cost and improving the response speed of semantic understanding. Combining bidirectional semantic interaction with entity boundary regression effectively ensures the optimization and balance of semantic understanding in terms of versatility, accuracy and response speed.

[0115] Based on any of the above embodiments, the understanding unit includes: An interaction subunit, configured to perform a bidirectional semantic interaction on the text representations of the text to be understood and the label description text based on the bidirectional interaction module in the semantic understanding model, so as to obtain interactive semantic representations of the text to be understood and the label description text; A comparison subunit is used to apply the interactive semantic representations of the text to be understood and the label description text based on the semantic comparison module in the semantic understanding model, perform semantic comparison on the text to be understood and the label description text, and obtain a semantic comparison result; A boundary regression subunit is used to perform entity boundary regression on the text to be understood based on the entity boundary regression module in the semantic understanding model, apply the interactive semantic representations of the text to be understood and the label description text, and obtain an entity boundary regression result; An output subunit is used to generate the semantic understanding result based on the output module in the semantic understanding model by applying the semantic comparison result and the entity boundary regression result.

[0116] Based on any of the above embodiments, the interaction subunit is specifically used for: Extracting a semantic representation of the text to be understood based on the text representation of the text to be understood; Based on the semantic representation of the text to be understood, semantic interaction is performed on the text representation of the label description text to obtain an interactive semantic representation of the label description text; Based on the interactive semantic representation of the tag description text, semantic interaction is performed on the text representation of the text to be understood to obtain the interactive semantic representation of the text to be understood.

[0117] Based on any of the above embodiments, the comparison subunit is specifically used for: Applying the interactive semantic representations of the text to be understood and the label description text, respectively, a semantic comparison is performed on the intentions described by the text to be understood and the label description text, and a semantic comparison is performed on each anchor point in the text to be understood and the entity type described by the label description text.

[0118] Based on any of the above embodiments, the boundary regression subunit is specifically used for: The interactive semantic representations of the text to be understood and the label description text are applied to perform entity anchor activation on each anchor point in the text to be understood, and entity boundary regression is performed on the activated anchor points.

[0119] Based on any of the above embodiments, the device further includes a training unit, which is used to: Get sample text; Extracting fragments from the sample text to obtain key fragments in the sample text, and describing the key fragments to obtain description text of the key fragments, wherein the fragment extraction and the fragment description are implemented based on a large language model; The sample text is used as the text to be understood during training, the description text is used as the label description text during training, and the key fragment is used as the label of the semantic understanding result during training to construct a training sample set; The semantic understanding model is trained based on the training sample set.

[0120] Based on any of the above embodiments, the training unit is further used for: Obtaining training samples for an intent recognition task and / or training samples for a named entity recognition task; The training samples are added to the training sample set.

[0121] Figure 8An example of a physical structure diagram of an electronic device is shown in FIG. Figure 8 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication interface 820 and the memory 830 communicate with each other through the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute the semantic understanding method, which includes: Get the text to be understood and the label description text; Inputting the text to be understood and the label description text into a semantic understanding model, and obtaining a semantic understanding result of the text to be understood output by the semantic understanding model; The semantic understanding model is used to perform two-way semantic interaction on the text representations of the text to be understood and the label description text, respectively, to obtain the interactive semantic representations of the text to be understood and the label description text, and to perform semantic comparison and entity boundary regression based on the interactive semantic representations to obtain the semantic understanding result.

[0122] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention, or the part that contributes to the relevant 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 a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0123] On the other hand, the present invention further provides a computer program product, the computer program product comprising a computer program, the computer program can be stored on a non-transitory computer-readable storage medium, when the computer program is executed by a processor, the computer can execute the semantic understanding method provided by the above methods, the method comprising: Get the text to be understood and the label description text; Inputting the text to be understood and the label description text into a semantic understanding model, and obtaining a semantic understanding result of the text to be understood output by the semantic understanding model; The semantic understanding model is used to perform two-way semantic interaction on the text representations of the text to be understood and the label description text, respectively, to obtain the interactive semantic representations of the text to be understood and the label description text, and to perform semantic comparison and entity boundary regression based on the interactive semantic representations to obtain the semantic understanding result.

[0124] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to perform the semantic understanding method provided by the above methods, the method comprising: Get the text to be understood and the label description text; Inputting the text to be understood and the label description text into a semantic understanding model, and obtaining a semantic understanding result of the text to be understood output by the semantic understanding model; The semantic understanding model is used to perform two-way semantic interaction on the text representations of the text to be understood and the label description text, respectively, to obtain the interactive semantic representations of the text to be understood and the label description text, and to perform semantic comparison and entity boundary regression based on the interactive semantic representations to obtain the semantic understanding result.

[0125] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0126] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiment.

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A semantic understanding method, characterized in that: include: Get the text to be understood and the label description text; Inputting the text to be understood and the label description text into a semantic understanding model, and obtaining a semantic understanding result of the text to be understood output by the semantic understanding model; The semantic understanding model is used to perform two-way semantic interaction on the text representations of the text to be understood and the label description text, respectively, to obtain the interactive semantic representations of the text to be understood and the label description text, and to perform semantic comparison and entity boundary regression based on the interactive semantic representations to obtain the semantic understanding result.

2. The semantic understanding method according to claim 1, characterized in that: The step of inputting the text to be understood and the label description text into a semantic understanding model to obtain a semantic understanding result of the text to be understood outputted by the semantic understanding model includes: Based on the bidirectional interaction module in the semantic understanding model, bidirectional semantic interaction is performed on the text representations of the text to be understood and the label description text to obtain the interactive semantic representations of the text to be understood and the label description text; Based on the semantic comparison module in the semantic understanding model, the interactive semantic representations of the text to be understood and the label description text are applied to perform semantic comparison on the text to be understood and the label description text to obtain a semantic comparison result; Based on the entity boundary regression module in the semantic understanding model, the interactive semantic representations of the text to be understood and the label description text are applied to perform entity boundary regression on the text to be understood, so as to obtain an entity boundary regression result; Based on the output module in the semantic understanding model, the semantic comparison result and the entity boundary regression result are applied to generate the semantic understanding result.

3. The semantic understanding method according to claim 2, characterized in that: The performing bidirectional semantic interaction on the text representations of the text to be understood and the label description text to obtain interactive semantic representations of the text to be understood and the label description text, includes: Extracting a semantic representation of the text to be understood based on the text representation of the text to be understood; Based on the semantic representation of the text to be understood, semantic interaction is performed on the text representation of the label description text to obtain an interactive semantic representation of the label description text; Based on the interactive semantic representation of the tag description text, semantic interaction is performed on the text representation of the text to be understood to obtain the interactive semantic representation of the text to be understood.

4. The semantic understanding method according to claim 2, characterized in that: The applying the interactive semantic representations of the text to be understood and the label description text to perform semantic comparison between the text to be understood and the label description text, includes: Applying the interactive semantic representations of the text to be understood and the label description text, respectively, a semantic comparison is performed between the intentions described by the text to be understood and the label description text, and a semantic comparison is performed between each anchor point in the text to be understood and the entity type described by the label description text.

5. The semantic understanding method according to claim 2, characterized in that: The applying the interactive semantic representations of the text to be understood and the label description text to perform entity boundary regression on the text to be understood includes: The interactive semantic representations of the text to be understood and the label description text are applied to perform entity anchor activation on each anchor point in the text to be understood, and entity boundary regression is performed on the activated anchor points.

6. The semantic understanding method according to any one of claims 1 to 5, characterized in that: The training steps of the semantic understanding model include: Get sample text; Extracting fragments from the sample text to obtain key fragments in the sample text, and describing the key fragments to obtain description text of the key fragments, wherein the fragment extraction and the fragment description are implemented based on a large language model; The sample text is used as the text to be understood during training, the description text is used as the label description text during training, and the key fragment is used as the label of the semantic understanding result during training to construct a training sample set; The semantic understanding model is trained based on the training sample set.

7. The semantic understanding method according to claim 6, characterized in that: The training step also includes: Obtaining training samples for an intent recognition task and / or training samples for a named entity recognition task; The training samples are added to the training sample set.

8. A semantic understanding device, characterized in that: include: An acquisition unit, used for acquiring the text to be understood and the label description text; A comprehension unit, used for inputting the text to be understood and the label description text into a semantic comprehension model, and obtaining a semantic comprehension result of the text to be understood output by the semantic comprehension model; The semantic understanding model is used to perform two-way semantic interaction on the text representations of the text to be understood and the label description text, respectively, to obtain the interactive semantic representations of the text to be understood and the label description text, and to perform semantic contrast learning and entity boundary regression based on the interactive semantic representations to obtain the semantic understanding result.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the semantic understanding method according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the semantic understanding method according to any one of claims 1 to 7 is implemented.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the semantic understanding method according to any one of claims 1 to 7 is implemented.

Images