CN109800430B - Semantic understanding method and system - Google Patents

Abstract

The present invention provides a method and system for semantic understanding. The method includes: obtaining corpus samples, establishing a corpus according to the corpus samples; obtaining characteristic information according to the corpus samples, and generating corresponding regular expressions according to the characteristic information; respectively extracting the regular expressions in the regular expressions Feature information generates multiple different regular expression combinations; analyze the existence accuracy of each feature information in each regular expression combination, and filter the regular expression combination; analyze the regular expression probability corresponding to the regular expression according to the corpus, and the regular expression is a regular expression formula and the filtered regular formula; generate a probabilistic semantic model according to the regular formula and the probability of the regular formula; obtain the user corpus; generate the corresponding user regular formula according to the user corpus; compare the user regular formula and the probability semantic model to obtain the user corpus user semantics. The present invention analyzes the acquired user corpus based on the probabilistic semantic model, so as to obtain the most likely user semantics.

Description

A semantic understanding method and system

technical field

The invention relates to the technical field of language processing, in particular to a semantic understanding method and system.

Background technique

With the rapid development of the network, intelligent processing of information is becoming more and more common. Computers, smart devices, etc. may need to process thousands of information every day. Smart devices generally analyze the corpus to obtain the corresponding regular expressions, thereby parsing the corpus.

However, in the process of corpus processing using word segmentation technology, there will be multiple parts of speech in the formation of regular expressions after word segmentation, and the situation that the regular expression corresponding to the obtained user corpus matches multiple regular expressions. It is still impossible to accurately judge the semantics and entities. Therefore, it is necessary to analyze the user corpus through a semantic understanding method and system to obtain the most likely corresponding user semantics.

Contents of the invention

The purpose of the present invention is to provide a semantic understanding method and system, which can analyze the acquired user corpus based on the probabilistic semantic model, so as to obtain the most likely user semantics.

The technical scheme provided by the invention is as follows:

The present invention provides a semantic understanding method, comprising:

Obtaining a corpus sample, and establishing a corpus according to the corpus sample;

Obtain feature information according to the corpus sample, and generate a corresponding regular expression according to the feature information;

respectively extracting the feature information in the regular expression to generate a plurality of different regular expression combinations;

Analyzing the existence accuracy of each feature information in each regular expression combination, and screening the regular expression combination;

Analyzing the regular expression probability corresponding to the regular expression according to the corpus, the regular expression is a combination of the regular expression and the filtered regular expression;

generating a probabilistic semantic model according to the regular expression and the probability of the regular expression;

Get user corpus;

Generate a corresponding user regular expression according to the user corpus;

Comparing the user regular expression with the probabilistic semantic model to obtain the user semantics of the user corpus.

Further, said acquiring feature information according to said corpus samples, and generating corresponding regular expressions according to said feature information specifically includes:

Segmenting the corpus sample by word segmentation technology to obtain the sample word segmentation and the part of speech corresponding to the sample word segmentation;

determining the sample successor words in the sample word segmentation according to the sample word segmentation and the part of speech of the word segmentation;

Analyzing the sentence structure of the corpus sample to obtain the association relationship between the sample word segments;

A corresponding regular expression is generated according to the feature information, where the feature information includes the sample word segmentation, the part of speech of the word segmentation, the sample successor word, and the association relationship.

Further, the analysis of the existence accuracy of each feature information in each regular expression combination, and screening the regular expression combination specifically includes:

Analyzing the existence accuracy of each feature information in each regular expression combination, and screening the regular expression combination;

If the existence of each characteristic information in a certain regular expression combination is accurate, then there is an accurate regular expression combination for storing the characteristic information;

If the existence of at least one characteristic information in a certain regular expression combination is inaccurate, then delete the regular expression combination in which the characteristic information is inaccurate.

Further, the analysis of the regular expression probability corresponding to the regular expression according to the corpus, the regular expression being the combination of the regular expression and the filtered regular expression specifically includes:

Determine the sample keywords in the corresponding regular formula according to the feature information in each regular formula;

Analyzing the sample keyword probability that the sample keyword appears in the corpus;

Determine the corresponding regular expression probability of the regular expression according to the sample keyword probability.

Further, comparing the user regular expression with the probabilistic semantic model to obtain the user semantics of the user corpus specifically includes:

comparing the user regular expression with the probabilistic semantic model;

If the user regular expression matches multiple regular expressions in the probabilistic semantic model, then determine the target regular expression according to the regular expression probabilities corresponding to the multiple matching regular expressions;

Parsing the user corpus according to the target regular expression to obtain the user semantics.

The present invention also provides a semantic understanding system, comprising:

The corpus building module obtains a corpus sample and builds a corpus according to the corpus sample;

The expression generation module obtains feature information according to the corpus sample obtained by the corpus building module, and generates a corresponding regular expression according to the feature information;

An expression combination module, respectively extracting the feature information in the regular expressions generated by the expression generation module to generate a plurality of different regular expression combinations;

A screening module, analyzing the existence accuracy of each feature information in each regular expression combination generated by the expression combination module, and screening the regular expression combination;

The probability analysis module analyzes the regular expression probability corresponding to the regular expression according to the corpus established by the corpus building module, and the regular expression is a combination of the regular expression and the filtered regular expression;

A model generation module that generates a probabilistic semantic model according to the regular expression obtained by the expression generation module and the screening module and the probability of the regular expression obtained by the probability analysis module;

The corpus acquisition module acquires user corpus;

A processing module that generates a corresponding user regular expression according to the user corpus acquired by the corpus acquisition module;

The comparison module compares the user regular expression obtained by the processing module with the probabilistic semantic model generated by the model generation module to obtain the user semantics of the user corpus.

Further, the expression generation module specifically includes:

The word segmentation unit performs word segmentation on the corpus sample obtained by the corpus building module through word segmentation technology, to obtain the sample word segmentation and the part of speech corresponding to the sample word segmentation;

The successor word determining unit determines the sample successor word in the sample word segmentation according to the sample word segmentation and the part of speech obtained by the word segmentation unit;

A relationship analysis unit, analyzing the sentence structure of the corpus sample obtained by the corpus building module to obtain the association relationship between the sample word segmentation obtained by the word segmentation unit;

An expression generation unit generates a corresponding regular expression according to the feature information, the feature information includes the sample word segmentation obtained by the word segmentation unit, the part of speech of the word segmentation, and the sample succession obtained by the succession word determination unit words and the association relationship determined by the relationship analysis unit.

Further, the screening module specifically includes:

The accuracy analysis unit analyzes the existence accuracy of each feature information in each regular expression combination generated by the expression combination module, and screens the regular expression combination;

The screening unit, if the accuracy analysis unit obtains that the existence of each feature information in a certain regular expression combination is accurate, then save the accurate regular expression combination in the feature information;

The screening unit is configured to delete the regular expression combination in which the feature information is inaccurate if the accuracy analysis unit obtains that at least one feature information in a certain regular expression combination is inaccurate.

Further, the probability analysis module specifically includes:

The keyword analysis unit determines the sample keywords in the corresponding regular expressions according to the feature information in each regular expression;

a probability analysis unit, analyzing the probability of the sample keywords determined by the keyword analysis unit appearing in the corpus established by the corpus building module;

A probability determining unit is configured to determine the corresponding regular expression probability of the regular expression according to the sample keyword probability obtained by the probability analyzing unit.

Further, the comparison module specifically includes:

a comparing unit, comparing the user regular expression with the probabilistic semantic model;

A processing unit, if the comparison unit obtains that the user regular expression matches multiple regular expressions in the probabilistic semantic model, then determine the target regular expression according to the regular expression probabilities corresponding to the multiple matching regular expressions;

A parsing unit, for parsing the user corpus according to the target regular expression obtained by the processing unit to obtain the user semantics.

The semantic understanding method and system provided by the present invention can bring at least one of the following beneficial effects:

1. In the present invention, the corresponding regular expression is generated according to the corpus sample, and then the feature information in the regular expression is extracted to obtain the regular expression combination, and the regular expression and the regular expression probability of the regular expression combination are analyzed in combination with the corpus, thereby analyzing the acquired user The largest possible user semantics of the corpus.

2. In the present invention, the corpus sample is segmented according to the word segmentation technology, and the sentence structure of the corpus sample is analyzed, thereby generating a corresponding regular expression, which is convenient for subsequently summarizing the corresponding regular expression combination and analyzing the regular expression probability.

3. In the present invention, the characteristic information in the regular expression combination generated by extracting the characteristic information in the regular expression is analyzed, and it is judged whether the regular expression combination is logical and has actual semantics, thereby screening the regular expression combination.

Description of drawings

In the following, a preferred implementation mode will be described in a clear and understandable manner in conjunction with the accompanying drawings, and the above-mentioned characteristics, technical features, advantages and implementation methods of a semantic understanding method and system will be further described.

Fig. 1 is the flowchart of an embodiment of a kind of semantic understanding method of the present invention;

Fig. 2 is the flowchart of another embodiment of a kind of semantic understanding method of the present invention;

Fig. 3 is the flowchart of another embodiment of a kind of semantic understanding method of the present invention;

Fig. 4 is the flowchart of another embodiment of a kind of semantic understanding method of the present invention;

Fig. 5 is the flowchart of another embodiment of a kind of semantic understanding method of the present invention;

Fig. 6 is a schematic structural diagram of an embodiment of a semantic understanding system of the present invention;

Fig. 7 is a schematic structural diagram of another embodiment of a semantic understanding system of the present invention.

Explanation of reference numbers:

100 semantic understanding system

110 Corpus Building Module

120 Expression generation module 121 Word segmentation unit 122 Successor word determination unit 123 Relationship analysis unit 124 Expression generation unit

130 expression combination module

140 Screening Module 141 Accuracy Analysis Unit 142 Screening Unit

150 Probability Analysis Module 151 Keyword Analysis Unit 152 Probability Analysis Unit 153 Probability Determination Unit

160 model generation module

170 Corpus acquisition module

180 processing modules

190 Comparison module 191 Comparison unit 192 Processing unit 193 Analysis unit

Detailed ways

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the specific implementation manners of the present invention will be described below with reference to the accompanying drawings. Apparently, the accompanying drawings in the following description are only some embodiments of the present invention, and those skilled in the art can obtain other accompanying drawings based on these drawings and obtain other implementations.

In order to make the drawings concise, each drawing only schematically shows the parts related to the present invention, and they do not represent the actual structure of the product. In addition, to make the drawings concise and easy to understand, in some drawings, only one of the components having the same structure or function is schematically shown, or only one of them is marked. Herein, "a" not only means "only one", but also means "more than one".

One embodiment of the present invention, as shown in Figure 1, a kind of semantic understanding method, comprises:

S100 Acquire a corpus sample, and build a corpus according to the corpus sample.

Specifically, a large number of corpus samples are obtained, and a corpus is established according to the corpus samples. The corpus samples can be standardized written language, or user voice, audio, etc., because user voice input and text input are the mainstream interaction methods in the process of human-computer interaction.

In addition, since the entire analysis process is aimed at written text, if the user's voice, audio and other voice files are collected, the voice file needs to be converted into recognized text first, and then the recognized text should be processed accordingly.

S200 Acquire feature information according to the corpus sample, and generate a corresponding regular expression according to the feature information.

Specifically, the part of speech and the sentence structure included in the sentence of the corpus sample are analyzed to obtain the feature information contained in the corpus sample, and then a regular expression corresponding to the corpus sample is generated according to the feature information.

S300 Extract the feature information in the regular expression to generate multiple different regular expression combinations.

Specifically, extracting feature information in the regular expressions to generate multiple different regular expression combinations is equivalent to arranging and combining the feature information contained in the regular expressions to obtain several regular expression combinations. For example, first arbitrarily select two pieces of feature information to combine into a regular expression combination, then arbitrarily select three feature information to combine into a regular expression combination, until all the regular expression combinations are obtained.

S400 Analyze the existence and accuracy of each feature information in each regular expression combination, and screen the regular expression combinations.

Specifically, there may be some unreasonable combinations of sentence components in the several regular formula combinations obtained from the above permutations and combinations, so the existence accuracy of each feature information in each regular formula combination is analyzed, so as to screen the regular formula combinations .

S500 Analyze the regular expression probability corresponding to the regular expression according to the corpus, where the regular expression is a combination of the regular expression and a filtered regular expression.

Specifically, the regular expressions obtained directly from the user corpus, and the combined regular expressions obtained after screening are regular expressions that conform to the logic of the sentence structure. According to all corpus samples in the corpus, each regular expression corresponding to the corpus is analyzed. The regular expression probability that appears in .

S600 Generate a probabilistic semantic model according to the regular expression and the probability of the regular expression.

Specifically, a probability semantic model is generated according to the regular expression and the corresponding regular expression probability, and a mapping relationship between the regular expression and the regular expression probability is established in the probability semantic model.

The S700 acquires user corpus.

Specifically, to obtain user corpus, in the process of obtaining user corpus by smart devices, users' voice input and text input are the mainstream interaction methods, but no matter what form the acquired user corpus is in, the final system will process text form, so if a speech form is obtained, it needs to be converted into text form first.

S800 Generate a corresponding user regular expression according to the user corpus.

Specifically, a corresponding user regular expression is generated according to part of speech and sentence structure included in the sentence of the user corpus obtained above.

S900 Compare the user regular expression with the probabilistic semantic model to obtain the user semantics of the user corpus.

Specifically, compare the user regular expressions obtained above with the regular expressions in the probabilistic semantic model one by one, select the regular expression with the highest regular expression probability among the matching regular expressions, analyze the user corpus, and obtain the corresponding user semantics.

In this embodiment, the corresponding regular expression is generated according to the corpus sample, and then the feature information in the regular expression is extracted to obtain the regular expression combination, and the regular expression probability of the regular expression and the regular expression combination is analyzed in conjunction with the corpus, thereby analyzing the obtained user corpus Maximum possible user semantics.

Another embodiment of the present invention is an optimized embodiment of the above-mentioned embodiment, as shown in Figure 2, including:

S100 Acquire a corpus sample, and build a corpus according to the corpus sample.

S200 Acquire feature information according to the corpus sample, and generate a corresponding regular expression according to the feature information.

The step S200 is to obtain feature information according to the corpus sample, and to generate a corresponding regular expression according to the feature information specifically includes:

S210 Segment the corpus sample by using a word segmentation technology to obtain the sample word segmentation and the part of speech corresponding to the sample word segmentation.

Specifically, the corpus sample is segmented according to the word segmentation technology, the part of speech in each sentence in the corpus sample is identified, and then the entire sentence in each sentence in the corpus sample is divided into words, words, and phrases according to the part of speech of the word. participle composition. Therefore, the sample part of speech contained in the corpus sample and the corresponding part of speech of the part of speech are obtained.

For example, if a corpus sample is "whale can spout water", the sample word segmentation obtained by word segmentation is "whale", "hui", "water spray", the part of speech corresponding to "whale" is a noun, and the part of speech corresponding to "hui" is The part of speech is a pronoun, and the part of speech corresponding to "spray water" is a noun.

S220 Determine a sample successor word in the sample word segmentation according to the sample word segmentation and the part of speech of the word segmentation.

Specifically, the sample successor words in the sample word segmentation are determined according to the sample word segmentation and part of speech, such as "and", "or", "not only", "and" and other sample successor words, which help to determine the relationship between the sample word segmentation.

S230 Analyze the sentence structure of the corpus sample to obtain the association relationship between the sample word segments.

Specifically, the above-mentioned word segmentation technology obtains the sample word segmentation and part of speech included in the corpus sample, and then analyzes the relationship between the sample word segmentation included in the corpus sample according to the sentence structure of the corpus sample.

For example, if a corpus sample is "whale can spout water", the sample word segmentation obtained by word segmentation is "whale", "hui", "spout water", the part of speech corresponding to "whale" is a noun, and the part of speech corresponding to "hui" is The part of speech of the participle is a pronoun, and the part of part of part of speech corresponding to "spray water" is a noun. By analyzing the sentence structure of the corpus sample, it is concluded that the noun "whale" and the verb "spray water" are subject-predicate relations.

S240 Generate a corresponding regular expression according to the feature information, where the feature information includes the sample word segmentation, the part of speech of the word segmentation, the sample successor word, and the association relationship.

Specifically, the corresponding regular expressions are generated according to the sample word segmentation, part of speech of the word segmentation, sample successor words, and association relations. For example, if a certain corpus sample is "whale can spout water", the content word segmentation obtained by word segmentation is "whale", " The part of speech corresponding to "hui", "spray water", and "whale" is a noun, the part of speech corresponding to "hui" is a pronoun, and the part of speech corresponding to "spray water" is a noun. The sentence structure of the entity content is analyzed to obtain the noun "whale " and the verb "water spray" are subject-predicate relations, and the regular expression obtained is: noun (whale)#pronoun (will)#verb (water spray).

S300 Extract the feature information in the regular expression to generate multiple different regular expression combinations.

S400 Analyze the existence and accuracy of each feature information in each regular expression combination, and screen the regular expression combinations.

S500 Analyze the regular expression probability corresponding to the regular expression according to the corpus, where the regular expression is a combination of the regular expression and a filtered regular expression.

S600 Generate a probabilistic semantic model according to the regular expression and the probability of the regular expression.

The S700 acquires user corpus.

S800 Generate a corresponding user regular expression according to the user corpus.

S900 Compare the user regular expression with the probabilistic semantic model to obtain the user semantics of the user corpus.

In this embodiment, the corpus sample is segmented according to the word segmentation technology, and the sentence structure of the corpus sample is analyzed to generate a corresponding regular expression, which facilitates the subsequent summary of the corresponding regular expression combination and analysis of the regular expression probability.

Another embodiment of the present invention is an optimized embodiment of the above-mentioned embodiment, as shown in Figure 3, including:

S100 Acquire a corpus sample, and build a corpus according to the corpus sample.

S200 Acquire feature information according to the corpus sample, and generate a corresponding regular expression according to the feature information.

S300 Extract the feature information in the regular expression to generate multiple different regular expression combinations.

S400 Analyze the existence and accuracy of each feature information in each regular expression combination, and screen the regular expression combinations.

The S400 analyzes the existence accuracy of each feature information in each regular expression combination, and screening the regular expression combination specifically includes:

S410 Analyze the existence accuracy of each feature information in each regular expression combination, and screen the regular expression combinations.

Specifically, the feature information in the regular expression is extracted to generate several different regular expression combinations, but there may be some combinations of sentence components with unreasonable structures. The corresponding regular expression is: noun (whale)#pronoun (hui)#verb (spray), and the part of the regular expression generated by extracting the feature information is: noun (whale)#pronoun (hui), pronoun (hui) #verb (spray water), noun (whale)#verb (spray water), of which the regular expression combination noun (whale)# pronoun (hui) has an incomplete structure, and the combination of noun (whale) and pronoun (hui) is not accurate , does not have complete semantics.

Therefore, it is necessary to analyze the existence and accuracy of each feature information in each regular expression combination, and to screen the regular expression combinations.

S420 If the presence of each feature information in a certain regular expression combination is accurate, save the exact regular expression combination in which the feature information exists.

Specifically, if the existence of each feature information in a certain regular expression combination is accurate, it means that the regular expression combination has real semantics, so the regular expression combination is saved.

S430 If the presence of at least one feature information in a certain regular expression combination is inaccurate, delete the regular expression combination in which the feature information is inaccurate.

Specifically, if the existence of at least one feature information in a certain regular expression combination is inaccurate, such as the regular expression combination generated by extracting nouns, pronouns, and subject-predicate relationship feature information, where the subject-predicate relationship is inaccurate, delete the Regular combination.

S500 Analyze the regular expression probability corresponding to the regular expression according to the corpus, where the regular expression is a combination of the regular expression and a filtered regular expression.

S600 Generate a probabilistic semantic model according to the regular expression and the probability of the regular expression.

The S700 acquires user corpus.

S800 Generate a corresponding user regular expression according to the user corpus.

S900 Compare the user regular expression with the probabilistic semantic model to obtain the user semantics of the user corpus.

In this embodiment, the characteristic information in the regular expression combination generated by extracting the characteristic information in the regular expression is analyzed, and it is judged whether the regular expression combination is logical and has actual semantics, so as to filter the regular expression combination.

Another embodiment of the present invention is an optimized embodiment of the above-mentioned embodiment, as shown in Figure 4, including:

S100 Acquire a corpus sample, and build a corpus according to the corpus sample.

S200 Acquire feature information according to the corpus sample, and generate a corresponding regular expression according to the feature information.

S300 Extract the feature information in the regular expression to generate multiple different regular expression combinations.

S400 Analyze the existence and accuracy of each feature information in each regular expression combination, and screen the regular expression combinations.

S500 Analyze the regular expression probability corresponding to the regular expression according to the corpus, where the regular expression is a combination of the regular expression and a filtered regular expression.

The S500 analyzes the probability of the regular expression corresponding to the regular expression according to the corpus, and the regular expression is a combination of the regular expression and the filtered regular expression and specifically includes:

S510 Determine the sample keywords in the corresponding regular expression according to the feature information in each regular expression.

Specifically, the sample keywords are determined according to the feature information in each regular expression, for example, the sample keywords are determined according to the association relationship, the sample participle corresponding to the subject in the subject-predicate relationship is determined as the sample keyword, or the sample key is determined according to the part of speech word, determine the sample participle corresponding to the verb or noun in the regular expression as the sample keyword. The determination rule of the sample keyword is set by the user according to actual needs, and the number of the sample keyword can be one or more.

For example, if a certain corpus sample is "whale can spout water", the regular expression obtained is: noun (whale)#pronoun (will)#verb (spray water). If the sample participle corresponding to the subject in the subject-predicate relationship is determined as the sample keyword, the sample keyword is "whale", and if the sample participle corresponding to the verb and the noun is determined as the sample keyword, then the sample keyword is "whale" and "water spray", so the number of sample keywords is not limited to one.

S520 Analyze the probability of the sample keyword appearing in the corpus of the sample keyword.

Specifically, after determining the sample keyword in each regular expression, the probability of the sample keyword appearing in all corpus samples in the corpus is counted. For example, if a certain corpus sample is "whale can spout", the regular expression obtained is: noun (whale)#pronoun (will)#verb (spray), and the determined sample keywords are "whale" and "spout". ", then count the sample keyword probabilities of "whale" and "spray" in all corpus samples respectively. For example, the corpus contains 100 corpus samples, of which 20 corpus samples contain "whale", and 10 corpus samples contains "water spray", the sample keyword probability corresponding to the sample keyword "whale" is 0.2, and the sample keyword probability corresponding to the sample keyword "water spray" is 0.1.

S530 Determine the corresponding regular expression probability of the regular expression according to the sample keyword probability.

Specifically, the regular expression probability of the corresponding regular expression is determined according to the sample keyword probability. If the regular expression has only one sample keyword, the sample keyword probability corresponding to the sample keyword is the regular expression probability corresponding to the regular expression. If the regular expression has multiple sample keywords, the probability of the sample keyword corresponding to the sample keyword is the regular expression probability corresponding to the regular expression.

S600 Generate a probabilistic semantic model according to the regular expression and the probability of the regular expression.

The S700 acquires user corpus.

S800 Generate a corresponding user regular expression according to the user corpus.

S900 Compare the user regular expression with the probabilistic semantic model to obtain the user semantics of the user corpus.

In this embodiment, the sample keywords in each regular expression are determined through preset rules, and then the sample keyword probability of the sample keywords is analyzed in combination with the corpus, so as to obtain the regular expression probability of the corresponding regular expression.

Another embodiment of the present invention is an optimized embodiment of the above-mentioned embodiment, as shown in Figure 5, including:

S100 Acquire a corpus sample, and build a corpus according to the corpus sample.

S200 Acquire feature information according to the corpus sample, and generate a corresponding regular expression according to the feature information.

S300 Extract the feature information in the regular expression to generate multiple different regular expression combinations.

S400 Analyze the existence and accuracy of each feature information in each regular expression combination, and screen the regular expression combinations.

S500 Analyze the regular expression probability corresponding to the regular expression according to the corpus, where the regular expression is a combination of the regular expression and a filtered regular expression.

S600 Generate a probabilistic semantic model according to the regular expression and the probability of the regular expression.

The S700 acquires user corpus.

S800 Generate a corresponding user regular expression according to the user corpus.

S900 Compare the user regular expression with the probabilistic semantic model to obtain the user semantics of the user corpus.

The S900 compares the user regular expression with the probabilistic semantic model, and obtains the user semantics of the user corpus, which specifically includes:

S910 compares the user regular expression with the probabilistic semantic model.

Specifically, the user regular expressions obtained from the user corpus and the regular expressions in the probabilistic semantic model are compared and matched one by one.

S920 If the user regular expression matches multiple regular expressions in the probabilistic semantic model, determine a target regular expression according to regular expression probabilities corresponding to multiple matching regular expressions.

S930: Parse the user corpus according to the target regular expression to obtain the user semantics.

Specifically, since the user regular expression contains multiple feature information, part of the feature information may match the regular expression 1 in the probabilistic semantic model, and the other part of the feature information matches the regular expression 2 in the probabilistic semantic model. If multiple regular expressions match in the semantic model, then compare and compare the regular expression probabilities corresponding to the matching regular expressions to determine the target regular expression, such as comparing the regular expression probabilities corresponding to the above regular expression 1 and regular expression 2 to determine the target regular expression formula, the obtained target regular formula is the most likely regular formula for parsing the user corpus, and then the user corpus is parsed according to the target regular formula to obtain the corresponding user semantics.

In this embodiment, the corresponding user regular expression is obtained according to the user corpus, and then compared with the semantic probability model to obtain the corresponding target regular expression, and the most likely user semantics is obtained by analyzing the user corpus through the target regular expression.

One embodiment of the present invention, as shown in Figure 6, a semantic understanding system 100, comprising:

The corpus building module 110 acquires a corpus sample, and builds a corpus according to the corpus sample.

Specifically, the corpus building module 110 acquires a large number of corpus samples, and builds a corpus based on the corpus samples. The corpus samples can be standardized written language, or user voice, audio, etc., because user voice input and text input are the mainstream interaction methods in the process of human-computer interaction.

In addition, since the entire analysis process is aimed at written text, if the user's voice, audio and other voice files are collected, the voice file needs to be converted into recognized text first, and then the recognized text should be processed accordingly.

The expression generating module 120 acquires feature information according to the corpus sample acquired by the corpus building module 110, and generates a corresponding regular expression according to the feature information.

Specifically, the expression generation module 120 analyzes the part of speech and sentence structure contained in the sentence of the corpus sample, so as to obtain the feature information contained in the corpus sample, and then generates a regular expression corresponding to the corpus sample according to the feature information.

The expression combination module 130 extracts the feature information in the regular expressions generated by the expression generation module 120 to generate multiple different regular expression combinations.

Specifically, the expression combination module 130 respectively extracts the feature information in the regular expressions to generate multiple different regular expression combinations, which is equivalent to arranging and combining the feature information contained in the regular expressions to obtain several regular expression combinations. For example, first arbitrarily select two pieces of feature information to combine into a regular expression combination, then arbitrarily select three feature information to combine into a regular expression combination, until all the regular expression combinations are obtained.

The screening module 140 analyzes the existence accuracy of each feature information in each regular expression combination generated by the expression combination module 130, and screens the regular expression combination.

Specifically, there may be some unreasonable combinations of sentence components in the several regular expression combinations obtained from the above permutations and combinations. Therefore, the screening module 140 analyzes the existence accuracy of each feature information in each regular expression combination, so that the regular expression combination to filter.

The probability analysis module 150 analyzes the regular expression probability corresponding to the regular expression according to the corpus built by the corpus building module 110, and the regular expression is a combination of the regular expression and the filtered regular expression.

Specifically, the regular expressions obtained directly from the user corpus, and the combined regular expressions obtained after screening are regular expressions that conform to the logic of the sentence structure, and the probability analysis module 150 analyzes each regular expression according to all corpus samples in the corpus. The corresponding regular expression probability of occurrence in the corpus.

The model generation module 160 generates a probabilistic semantic model according to the regular expression obtained by the expression generation module 120 and the screening module 140 and the probability of the regular expression obtained by the probability analysis module 150 .

Specifically, the model generation module 160 generates a probability semantic model according to the regular expression and the corresponding regular expression probability, and establishes a mapping relationship between the regular expression and the regular expression probability in the probability semantic model.

The corpus acquisition module 170 acquires user corpus.

Specifically, the corpus acquisition module 170 acquires the user corpus. During the process of acquiring the user corpus by the smart device, the user's voice input and text input are the mainstream interaction methods, but no matter what form the acquired user corpus is, the final system will process it. All are in text form, so if you get the speech form, you need to convert it into text form first.

The processing module 180 generates a corresponding user regular expression according to the user corpus acquired by the corpus acquisition module 170 .

Specifically, the processing module 180 generates a corresponding user regular expression according to part of speech and sentence structure included in the sentence of the user corpus obtained above.

The comparison module 190 compares the user regular expression obtained by the processing module 180 with the probabilistic semantic model generated by the model generation module 160 to obtain the user semantics of the user corpus.

Specifically, the comparison module 190 compares the user regular expressions obtained above with the regular expressions in the probabilistic semantic model one by one, selects the regular expression with the highest regular expression probability among the matching regular expressions, analyzes the user corpus, and obtains the corresponding user semantics.

In this embodiment, the corresponding regular expression is generated according to the corpus sample, and then the feature information in the regular expression is extracted to obtain the regular expression combination, and the regular expression probability of the regular expression and the regular expression combination is analyzed in combination with the corpus, thereby analyzing the obtained user corpus Maximum possible user semantics.

Another embodiment of the present invention is an optimized embodiment of the above-mentioned embodiment, as shown in FIG. 7 , including:

The corpus building module 110 acquires a corpus sample, and builds a corpus according to the corpus sample.

The expression generating module 120 acquires feature information according to the corpus sample acquired by the corpus building module 110, and generates a corresponding regular expression according to the feature information.

The expression generating module 120 specifically includes:

The word segmentation unit 121 performs word segmentation on the corpus sample acquired by the corpus building module 110 by word segmentation technology, and obtains the sample word segmentation and the part of speech corresponding to the sample word segmentation.

Specifically, the word segmentation unit 121 performs word segmentation on the corpus sample according to the word segmentation technology, identifies the part of speech in each sentence in the corpus sample, and then divides the entire sentence in each sentence in the corpus sample according to the part of speech of the word into words, words, and words. and phrase equivalence. Therefore, the sample part of speech contained in the corpus sample and the corresponding part of speech of the part of speech are obtained.

For example, if a corpus sample is "whale can spout water", the sample word segmentation obtained by word segmentation is "whale", "hui", "water spray", the part of speech corresponding to "whale" is a noun, and the part of speech corresponding to "hui" is The part of speech is a pronoun, and the part of speech corresponding to "spray water" is a noun.

The successor word determining unit 122 determines the sample successor word in the sample word segmentation according to the sample word segmentation and the part of speech obtained by the word segmentation unit 121 .

Specifically, the successor word determination unit 122 determines the sample successor words in the sample word segmentation according to the sample word segmentation and the part of speech of the word segmentation, such as "and", "or", "not only", "and" and other sample successor words, which help to determine the sample word segmentation The relationship between.

The relationship analysis unit 123 analyzes the sentence structure of the corpus sample acquired by the corpus building module 110 to obtain the association relationship between the sample word segments obtained by the word segmentation unit 121 .

Specifically, the above-mentioned word segmentation technology obtains the sample word segmentation and part of speech included in the corpus sample, and then the relationship analysis unit 123 analyzes the relationship between the sample word segmentation included in the corpus sample according to the sentence structure of the corpus sample.

For example, if a corpus sample is "whale can spout water", the sample word segmentation obtained by word segmentation is "whale", "hui", "spout water", the part of speech corresponding to "whale" is a noun, and the part of speech corresponding to "hui" is The part of speech of the participle is a pronoun, and the part of part of part of speech corresponding to "spray water" is a noun. By analyzing the sentence structure of the corpus sample, it is concluded that the noun "whale" and the verb "spray water" are subject-predicate relations.

The expression generating unit 124 generates a corresponding regular expression according to the feature information, the feature information including the sample word segmentation obtained by the word segmentation unit 121, the part of speech of the word segmentation, and the obtained word segmentation unit 122. The sample bearer word and the association relationship determined by the relationship analysis unit 123.

Specifically, the expression generating unit 124 generates a corresponding regular expression according to the sample word segmentation, part of speech of the word segmentation, sample successor words, and association relations. For example, a certain corpus sample is "whale can spout water", and the content word segmentation obtained by word segmentation is, "Whale", "hui", "spray", the part of speech corresponding to "whale" is a noun, the part of speech corresponding to "hui" is a pronoun, and the part of speech corresponding to "spray" is a noun, analyze the sentence structure of the entity content It is obtained that the noun "whale" and the verb "spray" are subject-predicate relations, and the regular expression obtained is: noun (whale)#pronoun (will)#verb (spray).

The expression combination module 130 extracts the feature information in the regular expressions generated by the expression generation module 120 to generate multiple different regular expression combinations.

The screening module 140 analyzes the existence accuracy of each feature information in each regular expression combination generated by the expression combination module 130, and screens the regular expression combination.

The screening module 140 specifically includes:

The accuracy analysis unit 141 analyzes the existence accuracy of each feature information in each regular expression combination generated by the expression combination module 130, and screens the regular expression combinations.

Specifically, the feature information in the regular expression is extracted to generate several different regular expression combinations, but there may be some combinations of sentence components with unreasonable structures. The corresponding regular expression is: noun (whale)#pronoun (hui)#verb (spray), and the part of the regular expression generated by extracting the feature information is: noun (whale)#pronoun (hui), pronoun (hui) #verb (spray water), noun (whale)#verb (spray water), of which the regular expression combination noun (whale)# pronoun (hui) has an incomplete structure, and the combination of noun (whale) and pronoun (hui) is not accurate , does not have complete semantics.

Therefore, the accuracy analysis unit 141 is required to analyze the existence accuracy of each feature information in each regular expression combination, and screen the regular expression combinations.

The screening unit 142, if the accuracy analysis unit 141 obtains that the existence of each feature information in a certain regular expression combination is accurate, save the exact regular expression combination in which the feature information exists.

Specifically, if the screening unit 142 judges that the existence of each feature information in a certain regular expression combination is accurate, it means that the regular expression combination has real semantics, so the regular expression combination is saved.

The screening unit 142, if the accuracy analysis unit 141 obtains that the existence of at least one feature information in a certain regular expression combination is inaccurate, deletes the regular expression combination in which the feature information is inaccurate.

Specifically, if the screening unit 142 judges that the existence of at least one feature information in a certain regular expression combination is inaccurate, for example, the regular expression combination generated by extracting nouns, pronouns, and subject-predicate relationship feature information, wherein the subject-predicate relationship is inaccurate , the regular expression combination is deleted.

The probability analysis module 150 analyzes the regular expression probability corresponding to the regular expression according to the corpus built by the corpus building module 110, and the regular expression is a combination of the regular expression and the filtered regular expression.

The probability analysis module 150 specifically includes:

The keyword analysis unit 151 determines the sample keywords in the corresponding regular expression according to the feature information in each regular expression.

Specifically, the keyword analysis unit 151 determines the sample keywords according to the feature information in each regular expression, for example, determines the sample keywords according to the association relationship, determines the sample participle corresponding to the subject in the subject-predicate relationship as the sample keyword, or determines the sample keywords according to The part of speech of the part of speech determines the sample keyword, and the sample participle corresponding to the verb or noun in the regular expression is determined as the sample keyword. The determination rule of the sample keyword is set by the user according to actual needs, and the number of the sample keyword can be one or more.

For example, if a certain corpus sample is "whale can spout water", the regular expression obtained is: noun (whale)#pronoun (will)#verb (spray water). If the sample participle corresponding to the subject in the subject-predicate relationship is determined as the sample keyword, the sample keyword is "whale", and if the sample participle corresponding to the verb and the noun is determined as the sample keyword, then the sample keyword is "whale" and "water spray", so the number of sample keywords is not limited to one.

The probability analysis unit 152 is configured to analyze the probability of the sample keyword determined by the keyword analysis unit 151 appearing in the corpus created by the corpus building module 110 .

Specifically, after determining the sample keywords in each regular expression, the probability analysis unit 152 counts the probability of the sample keywords appearing in all corpus samples in the corpus. For example, if a certain corpus sample is "whale can spout", the regular expression obtained is: noun (whale)#pronoun (will)#verb (spray), and the determined sample keywords are "whale" and "spout". ", then count the sample keyword probabilities of "whale" and "spray" in all corpus samples respectively. For example, the corpus contains 100 corpus samples, of which 20 corpus samples contain "whale", and 10 corpus samples contains "water spray", the sample keyword probability corresponding to the sample keyword "whale" is 0.2, and the sample keyword probability corresponding to the sample keyword "water spray" is 0.1.

The probability determining unit 153 determines the corresponding regular expression probability of the regular expression according to the sample keyword probability obtained by the probability analyzing unit 152 .

Specifically, the probability determination unit 153 determines the regular expression probability of the corresponding regular expression according to the sample keyword probability. If the regular expression has only one sample keyword, the sample keyword probability corresponding to the sample keyword is the regular expression corresponding to the regular expression. formula probability. If the regular expression has multiple sample keywords, the probability of the sample keyword corresponding to the sample keyword is the regular expression probability corresponding to the regular expression.

The model generation module 160 generates a probabilistic semantic model according to the regular expression obtained by the expression generation module 120 and the screening module 140 and the probability of the regular expression obtained by the probability analysis module 150 .

The corpus acquisition module 170 acquires user corpus.

The processing module 180 generates a corresponding user regular expression according to the user corpus acquired by the corpus acquisition module 170 .

The comparison module 190 compares the user regular expression obtained by the processing module 180 with the probabilistic semantic model generated by the model generation module 160 to obtain the user semantics of the user corpus.

The comparison module 190 specifically includes:

The comparing unit 191 compares the user regular expression with the probabilistic semantic model.

Specifically, the comparison unit 191 compares and matches the user's regular expressions obtained according to the user corpus and the regular expressions in the probabilistic semantic model one by one.

The processing unit 192, if the comparison unit 191 obtains that the user regular expression matches multiple regular expressions in the probabilistic semantic model, then determine the target regular expression according to the regular expression probabilities corresponding to the multiple matching regular expressions.

The parsing unit 193 parses the user corpus according to the target regular expression obtained by the processing unit 192 to obtain the user semantics.

Specifically, since the user regular expression contains multiple feature information, part of the feature information may match the regular expression 1 in the probabilistic semantic model, and the other part of the feature information matches the regular expression 2 in the probabilistic semantic model. In the case that multiple regular expressions match in the semantic model, the processing unit 192 compares and compares the regular expression probabilities corresponding to the matching regular expressions to determine the target regular expression, such as comparing the regular expression probabilities corresponding to the above regular expression 1 and regular expression 2 The target regular expression is determined, and the obtained target regular expression is the most likely regular expression for parsing the user corpus, so the parsing unit 193 parses the user corpus according to the target regular expression to obtain the corresponding user semantics.

In this embodiment, the corpus sample is segmented according to the word segmentation technology, and the sentence structure of the corpus sample is analyzed to generate a corresponding regular expression, and the feature information in the regular expression combination generated by extracting the feature information in the regular expression is performed. Analyze and judge whether the regular expression combination is logical and has actual semantics, so as to screen the regular expression combination.

The sample keyword in each regular expression is determined through preset rules, and the sample keyword probability of the sample keyword is analyzed in combination with the corpus, so as to obtain the regular expression probability of the corresponding regular expression. According to the user corpus, the corresponding user regular expression is obtained, and then compared with the semantic probability model to obtain the corresponding target regular expression, and the user corpus is analyzed through the target regular expression to obtain the most likely user semantics.

It should be noted that the above embodiments can be freely combined as required. The above is only a preferred embodiment of the present invention, it should be pointed out that, for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made, and these improvements and modifications can also be made. It should be regarded as the protection scope of the present invention.

Claims (8)

1. A semantic understanding method, comprising:

obtaining a corpus sample, and establishing a corpus according to the corpus sample;

acquiring feature information according to the corpus sample, and generating a corresponding regular expression according to the feature information;

respectively extracting the characteristic information in the regular expression to generate a plurality of different regular expression combinations;

analyzing the existence accuracy of each characteristic information in each regular pattern combination, and screening the regular pattern combinations;

analyzing the regular expression probability corresponding to the regular expression according to the corpus, wherein the regular expression is a combination of the regular expression and the screened regular expression;

generating a probability semantic model according to the regular expression and the regular expression probability;

acquiring a user corpus;

generating a corresponding user regular pattern according to the user corpus;

comparing the user regular expression with the probability semantic model to obtain user semantics of the user corpus;

analyzing the existence accuracy of each piece of characteristic information in each regular pattern combination, and screening the regular pattern combination specifically comprises the following steps:

analyzing the existence accuracy of each characteristic information in each regular pattern combination, and screening the regular pattern combinations;

If the existence of each piece of characteristic information in a certain regular pattern combination is accurate, the accurate regular pattern combination of the characteristic information is stored;

if the existence of at least one piece of characteristic information in a certain regular pattern combination is inaccurate, deleting the characteristic information to exist the inaccurate regular pattern combination.

2. The semantic understanding method according to claim 1, wherein the obtaining feature information according to the corpus sample, and generating a corresponding regular expression according to the feature information specifically includes:

performing word segmentation on the corpus sample through a word segmentation technology to obtain sample word segmentation and word segmentation part of speech corresponding to the sample word segmentation;

determining a sample carrier word in the sample word segmentation according to the sample word segmentation and the word part of the word segmentation;

analyzing sentence pattern structures of the corpus sample to obtain association relations among the sample word segmentation;

and generating a corresponding regular expression according to the characteristic information, wherein the characteristic information comprises the sample word segmentation, the word segmentation part of speech, the sample carrier word and the association relation.

3. The semantic understanding method according to claim 1 or 2, wherein the analyzing, according to the corpus, a regular expression probability corresponding to a regular expression, where the regular expression is a combination of the regular expression and the screened regular expression specifically includes:

Determining sample keywords in the corresponding regular expressions according to the characteristic information in each regular expression;

analyzing the probability of the sample keywords appearing in the corpus;

and determining the regular expression probability of the corresponding regular expression according to the sample keyword probability.

4. The semantic understanding method according to claim 1, wherein comparing the user regular expression and the probabilistic semantic model to obtain the user semantics of the user corpus specifically comprises:

comparing the user regular expression with the probability semantic model;

if the user regular pattern matches with a plurality of regular patterns in the probability semantic model, determining a target regular pattern according to the regular pattern probabilities corresponding to the matched regular patterns;

and analyzing the user corpus according to the target regular expression to obtain the user semantics.

5. A semantic understanding system, comprising:

the corpus establishing module is used for acquiring a corpus sample and establishing a corpus according to the corpus sample;

the expression generation module is used for acquiring characteristic information according to the corpus sample acquired by the corpus establishment module and generating a corresponding regular expression according to the characteristic information;

The expression combination module is used for respectively extracting the characteristic information in the regular expression generated by the expression generation module to generate a plurality of different regular expression combinations;

the screening module is used for analyzing the existence accuracy of each piece of characteristic information in each regular pattern combination generated by the expression combination module and screening the regular pattern combination;

the probability analysis module is used for analyzing the regular expression probability corresponding to the regular expression according to the corpus established by the corpus establishment module, wherein the regular expression is a combination of the regular expression and the screened regular expression;

the model generation module is used for generating a probability semantic model according to the regular expression obtained by the expression generation module and the screening module and the regular expression probability obtained by the probability analysis module;

the corpus acquisition module is used for acquiring user corpus;

the processing module generates a corresponding user regular pattern according to the user corpus acquired by the corpus acquisition module;

the comparison module is used for comparing the user regular expression obtained by the processing module with the probability semantic model generated by the model generation module to obtain user semantics of the user corpus;

The screening module specifically comprises:

the accuracy analysis unit is used for analyzing the existence accuracy of each piece of characteristic information in each regular pattern combination generated by the expression combination module and screening the regular pattern combination;

the screening unit is used for storing the exact regular pattern combination of the characteristic information if the accuracy analysis unit obtains that the existence of each characteristic information in a certain regular pattern combination is exact;

and the screening unit is used for deleting the regular pattern combination with inaccurate characteristic information if the accuracy analysis unit obtains that at least one characteristic information exists in a certain regular pattern combination.

6. The semantic understanding system of claim 5, wherein the expression generation module specifically comprises:

the word segmentation unit is used for segmenting the corpus sample acquired by the corpus establishment module through a word segmentation technology to obtain sample word segmentation and word segmentation part of speech corresponding to the sample word segmentation;

the receiving word determining unit is used for determining sample receiving words in the sample word segmentation according to the sample word segmentation and the word segmentation part of speech obtained by the word segmentation unit;

The relation analysis unit is used for analyzing the sentence pattern structure of the corpus sample acquired by the corpus establishment module to obtain the association relation between the sample word segmentation obtained by the word segmentation unit;

and the expression generating unit is used for generating a corresponding regular expression according to the characteristic information, wherein the characteristic information comprises the sample word segmentation obtained by the word segmentation unit, the word part of speech of the word segmentation, the sample adapting word obtained by the adapting word determining unit and the association relation determined by the relation analyzing unit.

7. The semantic understanding system according to claim 5 or 6, wherein the probability analysis module specifically comprises:

the keyword analysis unit is used for determining sample keywords in the corresponding regular expressions according to the characteristic information in each regular expression;

the probability analysis unit is used for analyzing the probability of the sample keywords, which are determined by the keyword analysis unit, in the corpus established by the corpus establishment module;

and the probability determining unit is used for determining the regular pattern probability of the corresponding regular pattern according to the sample keyword probability obtained by the probability analyzing unit.

8. The semantic understanding system according to claim 5, wherein the comparison module specifically comprises:

The comparison unit is used for comparing the user regular expression with the probability semantic model;

the processing unit is used for determining a target regular expression according to regular expression probabilities corresponding to the matched and matched regular expressions if the comparison unit obtains that the user regular expression is matched and matched with the regular expressions in the probability semantic model;

and the analysis unit is used for analyzing the user corpus according to the target regular expression obtained by the processing unit to obtain the user semantics.

Images