Semantic recognition method, device, robot and intelligent device

Abstract

The embodiment of the present invention provides a semantic recognition method, device, robot and intelligent device, the method comprising: obtaining a target sentence, determining the type of each word in the sentence, namely, a named entity, a position word and an action word indicating the intention of the sentence, and also determining the dependency relationship between the words in the target sentence. Then, a target phrase for describing the action object of the action word is determined according to the dependency relationship and the type of the word. Finally, the semantics of the target sentence is jointly described by the target phrase and the action word. In the above method, by determining the type of each word in the target sentence and the dependency relationship between different words, the modifying and modified relationship between different types of words can be obtained, thereby accurately extracting the action object of the action word from the target sentence, and ensuring the accuracy of semantic recognition.

Description

Technical Field

[0001] The present invention relates to the field of robots, and in particular to a semantic recognition method, device, robot and intelligent equipment. Background Art

[0002] With the development of artificial intelligence technology, various intelligent robots are increasingly entering people's lives, such as household or commercial cleaning robots, self-moving vending robots, commercial robots, etc. In order to facilitate users' use, intelligent robots usually support multiple semantic recognition methods, such as supporting human-computer dialogue in the form of text or voice.

[0003] Intelligent robots need to identify the intention of the dialogue sentences input by users and the named entities contained in the sentences to obtain the semantics of the dialogue sentences, and then realize human-computer dialogue based on the recognized semantics. However, in practice, the dialogue sentences input by users may be complex sentences containing multiple named entities, which increases the difficulty of semantic recognition. Summary of the invention

[0004] The embodiments of the present invention provide a semantic recognition method, apparatus, robot and intelligent device to ensure the accuracy of semantic recognition.

[0005] An embodiment of the present invention provides a semantic recognition method, including:

[0006] Determining named entities, positional terms, and action terms in a target sentence that indicate the intent of the target sentence;

[0007] Determining a dependency relationship between the named entity, the location term, and the action term;

[0008] Determining a target phrase containing different words in the target sentence according to the dependency relationship, wherein the target phrase is used to describe an object of action of the action word;

[0009] The semantics of the target sentence is determined according to the action word and the target phrase.

[0010] An embodiment of the present invention provides a semantic recognition device, including:

[0011] A first determination module is used to determine the named entities, position words and action words indicating the intention of the target sentence in the target sentence;

[0012] A second determination module, used to determine the dependency relationship between the named entity, the location word and the action word;

[0013] A third determination module is used to determine a target phrase containing different words in the target sentence according to the dependency relationship, wherein the target phrase is used to describe an object of action of the action word;

[0014] The fourth determination module is used to determine the semantics of the target sentence according to the action word and the target phrase.

[0015] An embodiment of the present invention provides a robot, comprising: a processor and a memory; wherein the memory is used to store one or more computer instructions, wherein when the one or more computer instructions are executed by the processor, the following are implemented:

[0016] Determining named entities, positional terms, and action terms in a target sentence that indicate the intent of the target sentence;

[0017] Determining a dependency relationship between the named entity, the location term, and the action term;

[0018] Determining a target phrase containing different words in the target sentence according to the dependency relationship, wherein the target phrase is used to describe an object of action of the action word;

[0019] The semantics of the target sentence is determined according to the action word and the target phrase.

[0020] An embodiment of the present invention provides an intelligent device. When a computer instruction stored in the intelligent device is executed by one or more processors, the one or more processors are caused to perform at least the following actions:

[0021] Determining named entities, positional terms, and action terms in a target sentence that indicate the intent of the target sentence;

[0022] Determining a dependency relationship between the named entity, the location term, and the action term;

[0023] Determining a target phrase containing different words in the target sentence according to the dependency relationship, wherein the target phrase is used to describe an object of action of the action word;

[0024] The semantics of the target sentence is determined according to the action word and the target phrase.

[0025] The semantic recognition method provided by the present invention obtains a target sentence, determines the type of each word in the sentence, namely, a named entity, a position word, and an action word indicating the intention of the sentence, and can also determine the dependency relationship between the words in the target sentence. Then, a target phrase for describing the object of action of the action word is determined according to the dependency relationship and the type of the word. Finally, the semantics of the target sentence are jointly described by the target phrase and the action word.

[0026] In the above method, by determining the type of each word in the target sentence and the dependency relationship between different words, the modifying and modified relationship between different types of words can be obtained, thereby accurately extracting the action object of the action word from the target sentence, ensuring the accuracy of semantic recognition. In particular, for complex sentences with multiple named entities and multiple positional words, the dependency relationship and word type can clearly reflect the modifying and modified relationship between multiple named entities and multiple positional words, thereby ensuring the accuracy of action object extraction and further ensuring the accuracy of semantic recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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.

[0028] Figure 1 A flowchart of a semantic recognition method provided by an embodiment of the present invention;

[0029] Figure 2 A flowchart of another semantic recognition method provided by an embodiment of the present invention;

[0030] Figure 3 is the dependency tree corresponding to a target sentence;

[0031] Figure 4 A flowchart of another semantic recognition method provided by an embodiment of the present invention;

[0032] Figure 5 A schematic diagram of applying the semantic recognition method provided by an embodiment of the present invention to a household cleaning scenario;

[0033] Figure 6 A schematic diagram of applying the semantic recognition method provided by an embodiment of the present invention to a shopping mall cleaning scenario;

[0034] Figure 7 A schematic diagram of the structure of a semantic recognition device provided by an embodiment of the present invention;

[0035] Figure 8 For Figure 7 A schematic diagram of the structure of a robot corresponding to the semantic recognition device provided in the illustrated embodiment;

[0036] Fig. 9 For Figure 7 A schematic diagram of the structure of an intelligent device corresponding to the semantic recognition apparatus provided in the illustrated embodiment. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments 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.

[0038] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms. Unless the context clearly indicates otherwise, "plurality" generally includes at least two.

[0039] As used herein, the words "if" and "if" may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0040] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a product or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such a product or system. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the product or system including the elements.

[0041] Optionally, the semantic recognition method provided by the embodiments of the present invention can be applied to intelligent robots such as public service robots, self-moving vending robots, cleaning robots, etc. The semantic recognition method can also be applied to human-computer dialogue plug-ins (or human-computer dialogue interfaces, human-computer dialogue function modules) integrated in online shopping systems and public service systems; the semantic recognition method can also be applied to smart terminals such as smart home appliances and smart wearable devices. Generally speaking, the semantic recognition method can be applied to any device or system that supports human-computer dialogue. Optionally, the following embodiments will illustrate the cleaning robot that provides cleaning functions as the executor of the semantic recognition method.

[0042] Based on the above description, some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the case where there is no conflict between the embodiments, the following embodiments and the features in the embodiments can be combined with each other. In addition, the step sequence in the following method embodiments is only an example and not a strict limitation.

[0043] Figure 1 A flowchart of a semantic recognition method provided by an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0044] 101, determine the named entities, position words, and action words that indicate the intention of the target sentence in the target sentence.

[0045] The user can generate a control instruction, i.e., a target sentence, to the cleaning robot to control the cleaning robot to clean a certain position in the area to be cleaned. Optionally, the cleaning robot can be provided with an operation screen or keyboard for the user to input a sentence, and in this case, the target sentence input by the user can be expressed in text form. More commonly, optionally, the cleaning robot can also be provided with a sound pickup device such as a microphone, and the cleaning robot can collect the target sentence in voice form through the sound pickup device.

[0046] After receiving the target sentence, the cleaning robot can also classify the words in the sentence, that is, determine the named entities, location words and action words in the target sentence. Among them, the action words are used to indicate the intention of the target sentence, and the location words are used to indicate the location of the object described by the named entity in the area to be cleaned. Optionally, the location words can specifically include directional words and / or spatial words.

[0047] Taking the cleaning scene as an example, the target sentence issued by the user may be: "Clean under the table in the bedroom". By identifying the action word "clean", the named entity "table", the location words "bedroom" and "under" in the target sentence, "bedroom" is specifically a spatial word, and "under" is specifically a directional word.

[0048] For determining the above different types of words, an optional method is that the cleaning robot can segment the target sentence and compare each word in the segmentation result with the words in the preset word set. The word set contains pre-collected named entities, action words and location words. The word set contains action words, named entities and location words commonly used in cleaning scenes.

[0049] Optionally, in another more common way, the cleaning robot can also determine the word type with the help of its own configured prediction model, that is, input the target sentence into the prediction model to determine the position words, action words and named entities in the target sentence according to the prediction results output by this model.

[0050] Optionally, the prediction model can be any one of a Convolutional Neural Network (CNN, CNN for short) model, a Recurrent Neural Network (RNN for short) network, a Bidirectional Encoder Representation from Transformers (BERT for short) model, and the like.

[0051] 102, determine the dependency relationships among named entities, location terms, and action terms.

[0052] At the same time, the cleaning robot can also analyze the dependencies between words in the target sentence.

[0053] In an optional manner, a generative syntactic analysis model, a discriminative syntactic analysis model, a decision-making syntactic analysis model, and a constraint satisfaction syntactic analysis model can be obtained based on a statistical method, and any of the above models can be used to perform syntactic dependency analysis to obtain the dependency relationship between the words in the target sentence. In another optional manner, the dependency analysis can also be implemented based on deep learning using different types of neural network models. In another optional manner, the dependency analysis can also be implemented based on a CYK algorithm containing preset rules. Optionally, the dependency relationship can be embodied in the form of a dependency tree.

[0054] It should be noted that, since the analysis of the word-to-word dependency relationship may not be based on the word type determined in step 101, the present invention does not limit the execution order of the above two steps. The execution order of the above steps 101 and 102 is only an example.

[0055] 103, determining a target phrase containing different words in the target sentence according to the dependency relationship.

[0056] 104, the semantics of the target sentence is determined based on the action words and the target phrase.

[0057] Next, by analyzing the dependency relationship and the types of words in the target sentence, the target phrase used to describe the action object of the action word in the target sentence is further determined. This target phrase can be composed of named entities and position words in the target sentence, or it can be composed of named entities alone. Finally, the semantics of the target sentence is determined based on the action words and the target phrase, that is, the action words and the target phrase jointly describe the semantics of the target sentence.

[0058] Among them, since the dependency relationship indicates the dependency relationship between any two words in the target sentence, it cannot reflect whether each word is a named entity, a location word or an action word. Therefore, in the process of determining the target phrase, it is also necessary to combine the types of each word in the target sentence identified in step 101. The process of determining the target phrase in combination with the word type can refer to the relevant description in the following embodiment.

[0059] Continuing with the above example of "clean under the table in the bedroom", after executing the above steps, it can be determined that "clean" is an action word, and the object of the action word is "under the table in the bedroom". The target word includes both named entities and location words.

[0060] In this embodiment, the recognition device obtains the target sentence, determines the type of each word in the sentence, namely, the named entity, the position word, and the action word that indicates the intention of the sentence, and can also determine the dependency relationship between the words in the target sentence. Then, based on the dependency relationship and the word type, the target phrase used to describe the object of action of the action word is determined. Finally, the semantics of the target sentence is jointly described by the target phrase and the action word.

[0061] In the above method, by determining the types of each word in the target sentence and the dependency relationship between different words, the modifying and modified relationship between different types of words can be obtained, so that the object of action of the action word can be accurately extracted from the target sentence to ensure the accuracy of semantic recognition.

[0062] Especially for complex sentences with multiple named entities and multiple positional terms, if the action terms and named entity extraction methods are used alone to determine the sentence semantics, it is impossible to accurately determine which named entity is the object of the action terms among the multiple named entities identified, nor is it possible to accurately determine the restrictive relationship between the positional terms and the named entities, thereby reducing the accuracy of semantic recognition. Figure 1 In the method of the illustrated embodiment, due to the existence of the dependency analysis process, it is possible to obtain the modifying and modified relationships between multiple positional words and multiple named entities contained in the target sentence, thereby ensuring the accuracy of the extraction of the action object, and further ensuring the accuracy of semantic recognition.

[0063] In the above embodiments, it has been mentioned that the model can be used to recognize named entities, location terms, and action terms, and determine dependency relationships. Figure 2 A flowchart of another semantic recognition method provided by an embodiment of the present invention, such as Figure 2 As shown, the method may include the following steps:

[0064] 201, input the target sentence into the prediction model, so that the prediction model outputs the action words in the target sentence and the labeling result of the target sentence.

[0065] 202, determining the position words and named entities in the target sentence according to the annotation results.

[0066] The target sentence can be input into the prediction model. Taking the BERT model as an example, the model can extract a first feature vector containing the information of the whole sentence, and a second feature vector corresponding to each word in the target sentence, and the second feature vector can actually be a fusion feature vector containing the context information and position information of the word in the target sentence. Based on this first feature vector, the first output layer of the BERT model can output the action words in the target sentence; based on this second feature vector, the second output layer of the BERT model can output the annotation result corresponding to the target sentence. The annotation result can reflect the position of the position words and named entities in the target sentence. Optionally, the BERT model can adopt the BIO annotation method (Begin, Inside, Outside, referred to as BIO) or the BIOE annotation method (B-begin, I-inside, O-outside, E-end, S-single, referred to as BIOE) to obtain the annotation result.

[0067] Taking the BIO annotation method as an example, assuming that the target sentence is "clean under the table in the bedroom", the BIO annotation sequence output by the prediction model is: "OOB-SI-SOB-EI-EB-DI-D", where O represents a non-named entity and a non-positional word, BS and IS represent the starting position and ending position of the spatial word in the positional word in the target sentence respectively; B-EI-E represents the starting position and ending position of the named entity in the target sentence respectively; B-DI-D represents the starting position and ending position of the locative word in the positional word in the target sentence respectively. According to the above annotation rules, the "bedroom" corresponding to "B-SI-S" in the annotation sequence is the spatial word in the positional word, the "below" corresponding to "B-DI-D" is the locative word in the positional word, and the "table" corresponding to "B-EI-E" is the named entity.

[0068] Optionally, the above model can simultaneously output the action words and named entities in the target sentence, wherein the prediction of the action words can be considered as a classification task, and the prediction of the named entities can be considered as a sequence labeling task. Optionally, the prediction model can also be trained in multi-task joint training, that is, the classification ability and sequence labeling ability of the prediction model are trained simultaneously, and the two tasks influence each other to improve the prediction effect of the model.

[0069] 203 , input the word segmentation result of the target sentence into the dependency syntactic analysis model, so that the dependency syntactic analysis model outputs the probability value of different dependency relations between any two words in the word segmentation result.

[0070] 204 , according to the dependency relationship corresponding to the maximum probability value between any two words, determine a dependency relationship tree reflecting the dependency relationship between the words in the word segmentation result.

[0071] At the same time, after obtaining the target sentence, it can also be segmented, and the segmentation results can be input into the dependency syntactic analysis model. The dependency syntactic analysis model can output the probability value of different dependency relationships between any two words in the target sentence. Based on the probability value output by the dependency syntactic analysis model, the dependency tree corresponding to the target sentence can be determined according to the dependency relationship corresponding to the maximum probability value between any two words.

[0072] The common dependency relationship between two words can be shown in Table 1 below:

[0073] Table 1

[0074] Relationship Tags illustrate root Root Node subj subject obj object pred predicate att attribute adv adverbial cmp complement coo Parallel structure adjct Additional Ingredients punc punctuation

[0075] Optionally, the dependency syntactic analysis model can be a neural network model. And in order to further improve the accuracy of dependency analysis, optionally, the part of speech of each word in the target sentence can also be input into the dependency syntactic analysis model at the same time as the word segmentation result is input. Optionally, the word segmentation processing and part of speech tagging processing of the target sentence can also be implemented by the corresponding neural network model.

[0076] Continuing with the above “clean under the table in the bedroom”, the dependency tree of the target sentence can be as follows Figure 3 The dependency tree shows that "below" is the object of "clean", "table" is the attributive of "below", "bedroom" is the attributive of "table", and "of" is the additional component of "bedroom".

[0077] It should be noted that the present invention does not limit the execution order of steps 201 to 202 and steps 203 to 204 .

[0078] 205 , determining a target phrase containing different words in the target sentence according to the dependency tree, wherein the target phrase is used to describe an object of action of the action word.

[0079] 206, determining the semantics of the target sentence based on the action words and the target phrase.

[0080] Finally, the target phrase can be determined according to the dependency tree, and the semantics of the target sentence can be described by the target phrase and action words.

[0081] For the specific implementation processes of the above steps 205 to 206, reference can be made to Figure 1 the relevant descriptions in the embodiments shown, which will not be elaborated here.

[0082] In this embodiment, the cleaning robot can respectively identify different types of words in the target statement by means of the prediction model configured by itself, and at the same time use the dependency syntax analysis model configured by itself to determine the dependency relationships between the words in the target statement, so as to determine the action words and the objects corresponding to the action words, and further realize the recognition of the semantics of the target statement. In addition, for the parts not described in detail in this embodiment and the achievable technical effects, reference can be made to Figure 1 the relevant descriptions of the embodiments shown, which will not be elaborated here.

[0083] Optionally, according to the descriptions in the above embodiments, the cleaning robot can determine the object of the action word in the target statement, that is, the target phrase, according to the dependency tree of the target statement. This target phrase is composed of multiple words in the target statement. Each node in the dependency tree corresponds to a word in the target statement, and the specific process of determining the target phrase can be as follows:

[0084] First, starting from the root node of the dependency tree, determine the target node corresponding to the action word. Among them, the node directly dependent on the root node is the target node. Then, traverse the different-level child nodes of this target node in turn. During the traversal process, determine the target word as the action object according to the types of the words corresponding to each level of child nodes. For the sake of clear description, according to the distance from the target node from near to far, the different-level child nodes of the target node can be respectively called the first-level child node, the second-level child node... the Nth-level child node, and the Nth-level child node is the leaf node of the entire dependency tree.

[0085] Then, during the process of traversing the different-level child nodes of the target node, in one case, if the word corresponding to the first-level child node of the target node is a location word, further traverse the first remaining-level child nodes of the target node, that is, traverse the second-level child node to the Nth-level child node mentioned above.

[0086] During the process of traversing the second-level child node to the Nth-level child node, if the Mth-level child node is traversed and the word corresponding to this Mth-level child node is a named entity, the core phrase is composed of the words corresponding to each level of child nodes from the first-level child node to the Mth-level child node. This Mth-level child node can be any level of child nodes from the second-level child node to the Nth-level child node, and in the dependency syntax tree, the Mth-level child node is the closest to the first-level child node. Among them, M < N, and M is an integer greater than or equal to 3.

[0087] Next, continue to traverse the target node's M+1th level child node to the Nth level child node (that is, the second remaining level child node of the Mth level child node), and determine the words corresponding to the M+1th level child node to the Nth level child node as the constraint position information of the core phrase. Finally, the core phrase and the constraint position information constitute the target phrase.

[0088] In the process of traversing different levels of child nodes of the target node, in another case, the word corresponding to the first-level child node of the target node is a named entity, then continue to traverse the remaining levels of child nodes of the target node, that is, the second-level child nodes to the Nth-level child nodes of the target node, and use the words corresponding to the remaining levels of child nodes as the position constraint information of the named entity, so that the target phrase is composed of the named entity and the position constraint information.

[0089] In practice, optionally, the words corresponding to each level of child nodes of the target node may be any one of named entities, directional words, and spatial words, and the core phrase, constrained position information and target phrase may also be composed of multiple words. The words corresponding to different levels of child nodes can be sorted in the order of spatial words, named entities, and directional words to obtain any one of the core phrases, constrained position information or target phrases.

[0090] In the process of traversing the different levels of child nodes of the target node, there is another situation where the words corresponding to the first-level child nodes of the target node are neither position words nor deployment named entities. At this time, it indicates that the target sentence may only include action words, that is, the user has not clearly given the object of the action word, then the preset phrase can be directly used as the target phrase.

[0091] Continuing with the cleaning scene as an example, the preset phrase can be the entire area to be cleaned, or a sub-area in the area to be cleaned, or an object in a certain area. Therefore, the preset phrase can usually be composed of named entities and / or spatial terms. For example, the target sentence input by the user is "start cleaning", the word corresponding to the target node is "cleaning", and the word corresponding to the first-level sub-node is "start", which is neither a positional word nor a named entity, indicating that the object of "cleaning" is not clearly given in the sentence, then the preset phrase "the whole house floor" can be determined as the target phrase, that is, the object of "cleaning".

[0092] The target phrase determined according to the above rules may include named entities as well as directional terms and / or spatial terms for accurately defining the positions of the named entities, so that the cleaning robot can accurately locate the position in the area to be cleaned through semantic recognition.

[0093] In practice, suppose that the named entity in the target phrase is used to describe an object in the area to be cleaned, that is, an obstacle in the area to be cleaned, and the obstacle exists in different sub-areas of the area to be cleaned, that is, there may be multiple obstacles described by the named entity in the entire area to be cleaned. In addition, the position words that define the position of the named entity may be omitted in the target sentence generated by the user. This default will cause the position of the obstacle described by the named entity in the target phrase to be inaccurate, which will further affect the accuracy of the cleaning robot in performing the cleaning task.

[0094] In order to improve the above problems, the obstacle map for the area to be cleaned established by the cleaning robot can also be used to complete the missing location words in the target phrase. Specifically, before the first cleaning task, the cleaning robot walks in the area to be cleaned to obtain sensor data collected by different sensors, and constructs an obstacle map of the area to be cleaned with this sensor data. Then, the missing location words in the target phrase are completed based on the positional relationship between the obstacles in the obstacle map. The obstacle map records the positional relationship between the sub-areas in the area to be cleaned, as well as the position of each obstacle in each sub-area.

[0095] For example, the area to be cleaned is the whole house, and there are tables in the bedroom and the living room. The target sentence issued by the user is "clean under the table next to the sofa", the action word in the sentence is "clean", the named entities are "table" and "sofa", and the directional words are "under" and "next to". The target phrase determined according to the dependency tree is "under the table next to the sofa", where the core phrase is "next to the sofa", and the position constraint information of the core phrase is "next to the sofa". It can be seen that the target phrase lacks a spatial word, and the absence of this spatial word will cause the target phrase to be unable to determine whether the object of "cleaning" is the table in the bedroom or the table in the living room.

[0096] At this point, the corresponding obstacle map of the whole house can be used to complete the spatial words: Since the target phrase contains the named entity "sofa", and the obstacle map describes that there is a positional association between "sofa" and "living room", the spatial word "living room" can be added to the target phrase, and the completed target phrase is "under the table next to the sofa in the living room". Through the completion of spatial words, the target phrase can uniquely determine that the object of "cleaning" is the table in the living room, making the location of the object of "cleaning" more precise. Among them, the principle of word supplementation is also in the order of spatial words, named entities, and directional words.

[0097] Continuing with the above cleaning scenario, the named entities and spatial terms contained in the target sentence can be a specific object or a specific space in the area to be cleaned. Continuing with the above example of "cleaning under the table in the bedroom", the named entity in this sentence is "table" and the spatial term is "bedroom". These terms are the default names of objects or spaces in the area to be cleaned. For these terms, users can also rename them according to their personal habits to get a custom name for the object or space. For example, rename "table" to "table 1" and rename "bedroom" to "Xiao Ming's bedroom".

[0098] In practice, since users have very flexible ways of renaming, it is unrealistic to collect all the custom names for the same object or space. Therefore, when the target sentence contains custom names of objects and / or regions, the prediction model often cannot accurately identify that the custom name belongs to a named entity, location term, or action term, which ultimately leads to inaccurate semantic recognition.

[0099] In order to avoid the above situation, Figure 4 A flowchart of another semantic recognition method provided by an embodiment of the present invention is as follows: Figure 4 As shown, before step 101, the method may further include the following steps:

[0100] 301, get the original statement.

[0101] 302, determine whether the original sentence includes a custom word, if it does, execute step 303, otherwise execute step 304.

[0102] 303, replacing the custom words with the default words to obtain the target sentence.

[0103] The user can generate control instructions in the form of text or voice for the cleaning robot, which is the original sentence. Then, the cleaning robot can perform word segmentation on the original sentence and match the word segmentation results with each word in the pre-maintained word set in turn to determine whether the original sentence contains the custom word in the sentence set.

[0104] Among them, for the pre-maintained word set, the set contains multiple word pairs with corresponding relationships.

[0105] Specifically, in the obstacle map corresponding to the area to be cleaned, each sub-area and each obstacle has a default name, and the user can manually modify the default name to obtain a custom name for each sub-area and each obstacle. For the same sub-area or obstacle, the default word used to describe the default name and the custom word used to describe the custom name are a pair of words with a corresponding relationship in the word set. If the original sentence contains a custom word in the word set, indicating that the original sentence has a custom name, the custom word can be replaced with a default word with a corresponding relationship.

[0106] For example, the original sentence is "Clean under Table 1 in Xiao Ming's bedroom". In this sentence, "Xiao Ming's bedroom" and "Table 1" are custom words. In the word set, "Xiao Ming's bedroom" corresponds to the default word "bedroom", and "Table 1" corresponds to the default word "table". After replacement, the target sentence obtained is "Clean under the table in the bedroom".

[0107] 304, taking the original sentence as the target sentence.

[0108] If the original sentence does not contain the custom word, indicating that the original sentence does not contain the user's custom name, the original sentence is the target sentence.

[0109] In this embodiment, when the original sentence contains a custom word, the prediction model often cannot accurately identify whether the custom word belongs to a named entity, a location word or an action word, thereby affecting the accuracy of semantic recognition. In order to improve the above problem, a pre-maintained word set can be used to replace the custom words in the sentence with default words that the prediction model can recognize, thereby ensuring the accuracy of action words, named entities and location words recognition, and ultimately ensuring the accuracy of semantic recognition.

[0110] In addition, if semantic recognition is performed using text matching, the original sentence and a sentence with known intent need to be input into the text matching model at the same time, and the semantics of the original sentence is determined by calculating the similarity between the feature vectors of the original sentence and the sentence with known intent. That is, the sentence intent is recognized by calculating the similarity between the feature vectors at the sentence level, and semantic recognition is further realized. Therefore, semantic recognition in the above manner requires the pre-writing and maintenance of multiple sentences with known intent. And in order to maximize the accuracy of semantic recognition, the number of sentences with known intent must be large.

[0111] According to the method in the above embodiment, since the default words are all words that can be recognized by the prediction model, the accuracy of semantic recognition can be guaranteed by replacing the custom words. And semantic recognition can be achieved by simply inputting a target sentence into the prediction model, and there is no need to maintain a large number of sentences with known intentions. And even if the user's renaming method is very flexible, due to the existence of the custom word replacement process, there is no need to continuously train the prediction model as the custom words are continuously updated, reducing the frequency and cost of model training.

[0112] In addition, according to Figure 2 In the description of the illustrated embodiment, a first feature vector at the sentence level is used to determine the action words in the target sentence. It is easy to understand that the first feature vector contains relevant information of both the named entities and the position words in the target sentence.

[0113] In order to reduce the influence of named entities and positional words in the sentence on the recognition of action words and improve the accuracy of action verb determination, after obtaining the target sentence, each word in the target sentence can also be matched with the default words in the above-mentioned word set. If the target sentence contains a default word, and the default word is a positional word or a named entity, the default word can be replaced with a preset character, which is used to indicate that the default word is a positional word or a named entity. Then the replaced sentence is input into the prediction model so that the prediction model can recognize the action words therein.

[0114] For example, if the target sentence is "Clean under the table in the bedroom", then "bedroom" and "table" are both default words. After replacement, the target sentence becomes "Clean under [ENTITY] of [SPACE]".

[0115] After the above replacement, for different sentences containing the same action word, their corresponding first feature vectors should be the same, that is, the interference of named entities and position words on the action word recognition process is reduced, thereby improving the accuracy of action word determination.

[0116] For ease of understanding, the specific implementation process of the semantic recognition method provided above is exemplified in combination with the following scenario.

[0117] (1) Home cleaning scenario: the area to be cleaned is the entire house where the user lives, and there are tables in the living room and bedroom. In this environment, the user can send the original sentence "clean under Xiao Ming's special table next to the sofa" to the cleaning robot.

[0118] After receiving the original sentence, the cleaning robot can first perform word segmentation on it and match the word segmentation result with the word set containing the custom words. Through matching, it can be determined that the target sentence contains the custom word "Xiao Ming's special table", and it can be replaced with the default word "table" that has a corresponding relationship with the custom word and is also included in the word set, thereby obtaining the target sentence "clean under the table next to the sofa".

[0119] After that, the target sentence is input into the prediction model to extract the action word "clean" from it; at the same time, the label sequence can be output: OOB-SI-SB-EI-EB-DI-DOB-DI-D. According to the label sequence, the named entities in the target sentence can be identified as "sofa" and "table"; the directional words are "next to" and "below". The target sentence is then input into the dependency parsing model, so that the model can output the dependency tree corresponding to the target sentence, such as Figure 5 As shown in (b).

[0120] Next, the action words, named entities, and location words identified by the model, as well as the dependency tree, are used to determine the objects of the action words:

[0121] Starting from the root node ROOT of the dependency tree, the node that has a direct dependency relationship with it is the target node, which corresponds to the action word "clean". Figure 5 As shown in (b), the "below" node, the "table" node, the "next to" node, and the "sofa" node are the first- to fourth-level child nodes of the target node, respectively. The "of" node is the child node of "next to", which can be considered as the fifth-level child node of the target node.

[0122] Further traverse the different levels of child nodes of the target node: If the word corresponding to the "below" node is a directional word, further traverse the "table" node. Since the word corresponding to the "table" node is a named entity and the "table" node is closest to the "below" node, "table" and "below" can together constitute the core phrase. Continue to traverse the "next to" node, the "sofa" node and the "of" child node. Since the "of" node is a leaf node, the traversal process ends. After the above traversal, the phrase "next to the sofa" can be combined in the order of spatial words, named entities, and directional words. This phrase can be used as the position constraint information of the above core phrase to limit the position of "under the table" in the entire area to be cleaned.

[0123] Finally, the core phrase "under the table" and the location constraint information "next to the sofa" can form a target phrase as the object of "cleaning", that is, to achieve semantic recognition. The cleaning robot can clean under the table.

[0124] The above target phrase "under the table next to the sofa" only contains directional words and named entities, but no spatial words. However, since there are tables in both the bedroom and the living room in the area to be cleaned, the above target phrase cannot determine whether the "table" to be cleaned is in the bedroom or the living room, that is, the semantic recognition result is not accurate enough, which may cause errors in the cleaning robot during cleaning.

[0125] In order to improve the above problem, after obtaining the target phrase, if there is a lack of directional words or position words in the target phrase, the missing spatial words in the target phrase can be supplemented according to the obstacle map corresponding to the area to be cleaned: According to the pre-established obstacle map, "sofa" and "living room" have an associated positional relationship, so the "table" in the target phrase should refer to the table in the living room, that is, the missing spatial word in the target phrase is "living room", and the spatial word "living room" can be supplemented to the target phrase "under the table next to the sofa". After completion, the object of action of the action word "clean" is "under the table next to the sofa in the living room", which points to the only object in the area to be cleaned, thereby improving the accuracy of semantic recognition, and the cleaning robot can clean under the table in the living room.

[0126] The above content can be combined with Figure 5 understand.

[0127] (2) In the shopping mall cleaning scenario, the user can send an original sentence to the service robot located in the shopping mall lobby: "Clean stores B and C next to store A on the first floor."

[0128] Through matching, it is found that the original sentence does not contain custom words. Then this original sentence can be used as the target sentence and input into the prediction model configured by the cleaning robot itself to extract the action word "clean" from it; at the same time, the annotation sequence can also be output: OO BS IS BE BE IE BD ID O BD BD ID O BD BD ID. According to the annotation sequence, the named entities in the target sentence can be identified as "Store A", "Store B" and "Store C"; the direction word is "next to"; and the space word is "first floor". Then the target sentence is input into the dependency syntactic analysis model, so that the model can output the dependency tree corresponding to the target sentence, such as Figure 6 As shown in (b).

[0129] Next, the action words, named entities, and location words identified by the model, as well as the dependency tree, are used to determine the objects of the action words:

[0130] Starting from the root node ROOT of the dependency tree, the node that has a direct dependency relationship with it is the target node, which corresponds to the action word "clean". Figure 6 As shown in (b), the "Shop B" node, the "Next to" node, the "Shop A" node, and the "First Floor" node are the first-level child nodes to the fourth-level child nodes of the target node, respectively; the "De" node is the child node of "Next to" and can be considered as the fifth-level child node of the target node; the "Shop C" node is a child node of the "Shop B" node.

[0131] Further traverse the different levels of child nodes of the target node: If the word corresponding to the "Store B" node is a named entity, then further traverse the child nodes of the "Store B" node: "Store C" node, "Next to" node, "Store A" node, "First floor" and "of" child node. Since the "of" node is a leaf node, the traversal ends. After the above traversal, the order of spatial words, named entities, and directional words can be combined into the phrase "Next to Shop A on the first floor". This phrase can be used as the location constraint information of "Store B" to limit the position of "Store B" in the entire shopping mall.

[0132] Finally, "Store B" and the location constraint information "next to Store A on the first floor" can together form target phrase 1, and "Store C" can constitute target phrase 2, so that these two target phrases can be used as the objects of "cleaning", that is, to achieve semantic recognition. The cleaning robot can clean Store B and Store C successively.

[0133] The above content can be combined with Figure 6 understand.

[0134] The semantic recognition device of one or more embodiments of the present invention will be described in detail below. Those skilled in the art will appreciate that these semantic recognition devices can be configured using commercially available hardware components through the steps taught in this solution.

[0135] Figure 7 A schematic diagram of the structure of a semantic recognition device provided by an embodiment of the present invention is shown in FIG. Figure 7 As shown, the device comprises:

[0136] The first determination module 11 is used to determine the named entities, position words and action words that indicate the intention of the target sentence in the target sentence.

[0137] The second determining module 12 is used to determine the dependency relationship between the named entity, the location word and the action word.

[0138] The third determination module 13 is used to determine a target phrase containing different words in the target sentence according to the dependency relationship, wherein the target phrase is used to describe an object of action of the action word.

[0139] The fourth determination module 14 is used to determine the semantics of the target sentence according to the action word and the target phrase.

[0140] Optionally, the first determination module 11 is specifically used to: input the target sentence into a prediction model so that the prediction model outputs the action words in the target sentence and the labeling result of the target sentence; and determine the position words and the named entities in the target sentence based on the labeling result.

[0141] Optionally, the device further includes: a statement acquisition module 21 and a replacement module 22.

[0142] The sentence acquisition module 21 is used to acquire the original sentence

[0143] The replacement module 22 is configured to replace the custom words with default words if the original sentence contains the custom words, so as to obtain the target sentence, and the custom words and the default words have a corresponding relationship.

[0144] Optionally, the second determination module 12 is specifically used to: input the word segmentation result of the target sentence into a dependency syntactic analysis model, so that the dependency syntactic analysis model outputs a probability value of different dependency relationships between any two words in the word segmentation result; and determine a dependency relationship tree reflecting the dependency relationship between the words in the word segmentation result based on the dependency relationship corresponding to the maximum probability value between the any two words.

[0145] Optionally, the nodes in the dependency tree correspond one-to-one to the words in the word segmentation result.

[0146] The third determination module 13 is specifically used to: determine the target node corresponding to the action word from the dependency tree;

[0147] If the word corresponding to the first-level child node of the target node is the position word, traverse the first remaining-level child nodes of the target node;

[0148] If the word corresponding to any one of the first remaining level sub-nodes is the named entity, then determining that the words corresponding to the sub-nodes at all levels from the first level sub-node to the any one of the first level sub-nodes are the core phrases, and the any one of the first level sub-nodes is closest to the first level sub-node;

[0149] Traversing the second remaining level child nodes of any one level child node, and taking the words corresponding to the second remaining level child nodes as the position constraint information of the core phrase,

[0150] The core phrase and the position constraint information are determined as the target phrase.

[0151] Optionally, the third determination module 13 is specifically configured to: if the word corresponding to the first-level child node of the target node is the named entity, traverse the remaining-level child nodes of the target node to use the words corresponding to the remaining-level child nodes as the position constraint information of the named entity;

[0152] The named entity and the location constraint information are determined as the target phrase.

[0153] Optionally, the third determination module 13 is specifically configured to: if the word corresponding to the first-level child node of the target node is neither the position word nor the named entity, determine a preset phrase as the target phrase.

[0154] Optionally, the device further comprises: a generating module 23, configured to generate at least one of the core phrase, the location constraint information and the target phrase according to the sequence of spatial words, named entities and directional words.

[0155] Optionally, the device further includes: a map acquisition module 24 and a completion module 25 .

[0156] The map acquisition module 24 is used to acquire an obstacle map corresponding to the area to be cleaned.

[0157] The completion module 25 is used to complete the missing position words in the target phrase according to the position relationship between different obstacles in the obstacle map.

[0158] Optionally, the device further comprises: an establishing module 26, configured to establish the obstacle map in response to the walking of the cleaning robot in the area to be cleaned.

[0159] Figure 7 The device shown can perform the aforementioned Figures 1 to 4 The semantic recognition method provided in the embodiment shown in the figure, the parts not described in detail in this embodiment and the technical effects that can be achieved can all be referred to in the Figures 1 to 4 The relevant description of the illustrated embodiment will not be repeated here.

[0160] The above describes the internal functions and structure of the semantic recognition device. Optionally, in a possible design, the structure of the semantic recognition device can be implemented as a robot, such as Figure 8 As shown, the robot may include: a processor 31 and a memory 32. The memory 32 is used to store the information that supports the robot to execute the aforementioned Figures 1 to 4 The program of the semantic recognition method provided in the illustrated embodiment, the processor 31 is configured to execute the program stored in the memory 32 .

[0161] The program includes one or more computer instructions, wherein the one or more computer instructions can implement the following steps when executed by the processor 31:

[0162] Determining named entities, positional terms, and action terms in a target sentence that indicate the intent of the target sentence;

[0163] Determining a dependency relationship between the named entity, the location term, and the action term;

[0164] Determining a target phrase containing different words in the target sentence according to the dependency relationship, wherein the target phrase is used to describe an object of action of the action word;

[0165] The semantics of the target sentence is determined according to the action word and the target phrase.

[0166] Optionally, the processor 31 is further configured to execute the aforementioned Figures 1 to 4 All or part of the steps in the illustrated embodiments.

[0167] The structure of the electronic device may further include a communication interface 33 for communicating with other devices or a communication network.

[0168] Optionally, in a possible design, the structure of the semantic recognition device can be implemented as an intelligent device, such as Figure 8 As shown, the smart device may include: a processor 41 and a memory 42. The memory 42 is used to store the information that supports the robot to execute the aforementioned Figures 1 to 4 The program of the semantic recognition method provided in the illustrated embodiment, the processor 41 is configured to execute the program stored in the memory 42 .

[0169] The program includes one or more computer instructions, wherein the one or more computer instructions can implement the following steps when executed by the processor 41:

[0170] Determining named entities, positional terms, and action terms in a target sentence that indicate the intent of the target sentence;

[0171] Determining a dependency relationship between the named entity, the location term, and the action term;

[0172] Determining a target phrase containing different words in the target sentence according to the dependency relationship, wherein the target phrase is used to describe an object of action of the action word;

[0173] The semantics of the target sentence is determined according to the action word and the target phrase.

[0174] Optionally, the processor 41 is further configured to execute the aforementioned Figures 1 to 4All or part of the steps in the illustrated embodiments.

[0175] The structure of the electronic device may further include a communication interface 43 for communicating with other devices or a communication network.

[0176] In addition, an embodiment of the present invention provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed by one or more processors, the one or more processors are caused to perform at least the following actions:

[0177] Determining named entities, positional terms, and action terms in a target sentence that indicate the intent of the target sentence;

[0178] Determining a dependency relationship between the named entity, the location term, and the action term;

[0179] Determining a target phrase containing different words in the target sentence according to the dependency relationship, wherein the target phrase is used to describe an object of action of the action word;

[0180] The semantics of the target sentence is determined according to the action word and the target phrase.

[0181] The device embodiments described above are merely illustrative, wherein the modules described as separate components may or may not be physically separate. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment.

[0182] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by adding a necessary general hardware platform, and of course can also be implemented by combining hardware and software. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a computer product.

[0183] 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 recognition method, characterized in that: include: Determining named entities, positional terms, and action terms in a target sentence that indicate the intent of the target sentence; Determining a dependency tree among the named entity, the location word, and the action word, wherein nodes in the dependency tree correspond one-to-one to words in the target sentence; Traverse the different levels of child nodes of the target node; If the word corresponding to the first-level child node of the target node is the location word or the named entity, a target phrase is determined according to the types of words corresponding to the remaining level child nodes of the target node, the target node corresponds to the action word, and the target phrase is used to describe the object of action of the action word; If the word corresponding to the first-level child node of the target node is neither the position word nor the named entity, determining the preset phrase as the target phrase; The semantics of the target sentence is determined according to the action word and the target phrase.

2. The method according to claim 1, characterized in that The step of determining the named entities, position words, and action words indicating the intention of the target sentence in the target sentence includes: Inputting the target sentence into a prediction model, so that the prediction model outputs the action words in the target sentence and a labeling result of the target sentence; The position words and the named entities in the target sentence are determined according to the labeling result.

3. The method according to claim 2, characterized in that The method of inputting the target sentence into the prediction model also includes: Get the original sentence; If the original sentence contains a custom word, the custom word is replaced with a default word to obtain the target sentence, and the custom word and the default word have a corresponding relationship.

4. The method according to claim 1, characterized in that: The determining of the dependency tree between the named entity, the location word and the action word comprises: Inputting the word segmentation result of the target sentence into a dependency syntactic analysis model, so that the dependency syntactic analysis model outputs a probability value that different dependency relationships exist between any two words in the word segmentation result; According to the dependency relationship corresponding to the maximum probability value between the arbitrary two words, a dependency relationship tree reflecting the dependency relationship between the words in the word segmentation result is determined.

5. The method according to claim 4, characterized in that The step of determining the target phrase according to the types of words corresponding to the remaining sub-nodes of the target node includes: If the word corresponding to the first-level child node of the target node is the position word, traverse the first remaining-level child nodes of the target node; If the word corresponding to any one of the first remaining level sub-nodes is the named entity, then determining that the words corresponding to the sub-nodes at all levels from the first level sub-node to the any one of the first level sub-nodes are core phrases, and the any one of the first level sub-nodes is closest to the first level sub-node; Traversing the second remaining level child nodes of any one level child node, and taking the words corresponding to the second remaining level child nodes as the position constraint information of the core phrase, The core phrase and the position constraint information are determined as the target phrase.

6. The method according to claim 5, characterized in that The step of determining the target phrase according to the types of words corresponding to the remaining sub-nodes of the target node includes: If the word corresponding to the first-level child node of the target node is the named entity, traverse the remaining-level child nodes of the target node to use the words corresponding to the remaining-level child nodes as the position constraint information of the named entity; The named entity and the location constraint information are determined as the target phrase.

7. The method according to claim 5 or 6, characterized in that: The position words include directional words and spatial words, and the method further includes: At least one of the core phrase, the location constraint information and the target phrase is generated in the order of the spatial words, the named entities and the directional words.

8. The method according to claim 1, characterized in that The target sentence includes a control instruction issued by a user to the cleaning robot; the named entity is used to describe obstacles in the area to be cleaned; The method further comprises: Obtaining an obstacle map corresponding to the area to be cleaned; According to the positional relationship between different obstacles in the obstacle map, the missing position words in the target phrase are completed.

9. The method according to claim 8, characterized in that: The method further comprises: In response to the cleaning robot walking in the area to be cleaned, the obstacle map is established.

10. A semantic recognition device, characterized in that: include: A first determination module is used to determine the named entities, position words and action words indicating the intention of the target sentence in the target sentence; A second determination module is used to determine a dependency tree between the named entity, the location word and the action word, wherein the nodes in the dependency tree correspond one-to-one to the words in the target sentence; A third determination module is used to traverse different levels of child nodes of the target node; If the word corresponding to the first-level child node of the target node is the location word or the named entity, a target phrase is determined according to the type of words corresponding to the remaining level child nodes of the target node, the target node corresponds to the action word, and the target phrase is used to describe the object of action of the action word; if the word corresponding to the first-level child node of the target node is neither the location word nor the named entity, a preset phrase is determined as the target phrase; The fourth determination module is used to determine the semantics of the target sentence according to the action word and the target phrase.

11. A robot, characterized in that: include: A processor and a memory; wherein the memory is used to store one or more computer instructions, wherein the one or more computer instructions, when executed by the processor, implement: Determining named entities, positional terms, and action terms in a target sentence that indicate the intent of the target sentence; Determining a dependency tree among the named entity, the location word, and the action word, wherein nodes in the dependency tree correspond one-to-one to words in the target sentence; Traverse the different levels of child nodes of the target node; If the word corresponding to the first-level child node of the target node is the location word or the named entity, a target phrase is determined according to the types of words corresponding to the remaining level child nodes of the target node, the target node corresponds to the action word, and the target phrase is used to describe the object of action of the action word; If the word corresponding to the first-level child node of the target node is neither the position word nor the named entity, determining the preset phrase as the target phrase; The semantics of the target sentence is determined according to the action word and the target phrase.

12. A smart device, characterized in that: When the computer instructions stored in the smart device are executed by one or more processors, the one or more processors are caused to perform at least the following actions: Determining named entities, positional terms, and action terms in a target sentence that indicate the intent of the target sentence; Determining a dependency tree among the named entity, the location word, and the action word, wherein nodes in the dependency tree correspond one-to-one to words in the target sentence; Traverse the different levels of child nodes of the target node; If the word corresponding to the first-level child node of the target node is the location word or the named entity, a target phrase is determined according to the types of words corresponding to the remaining level child nodes of the target node, the target node corresponds to the action word, and the target phrase is used to describe the object of action of the action word; If the word corresponding to the first-level child node of the target node is neither the position word nor the named entity, determining the preset phrase as the target phrase; The semantics of the target sentence is determined according to the action word and the target phrase.

Images