CN113762301A

CN113762301A - Training of information matching model, information matching method and device

Info

Publication number: CN113762301A
Application number: CN202010619956.5A
Authority: CN
Inventors: 王泽勋; 冯明超
Original assignee: Beijing Jingdong Century Trading Co Ltd; Beijing Wodong Tianjun Information Technology Co Ltd
Current assignee: Beijing Jingdong Century Trading Co Ltd; Beijing Wodong Tianjun Information Technology Co Ltd
Priority date: 2020-07-01
Filing date: 2020-07-01
Publication date: 2021-12-07

Abstract

The invention discloses a training method and a training device of an information matching model and an information matching method and device, and relates to the technical field of computers. One embodiment of the method comprises: determining a plurality of training information sets, wherein each training information set comprises two sections of training information and matching labels of the training information sets; generating corresponding feature score vectors for the training information groups according to the two sections of training information; and training a preset model to be trained by using the feature score vector and the matching label to obtain an information matching model. The implementation method can effectively reduce the operation cost of information matching.

Description

Training of information matching model, information matching method and device

Technical Field

The invention relates to the technical field of computers, in particular to a training and information matching method and device of an information matching model.

Background

The information matching is to calculate the similarity of two pieces of information by using an algorithm, and is a very core application in various current dialog systems. For example, many systems can respond accordingly based on the matching result by matching the problem with a large number of samples.

In the process of implementing the invention, the inventor finds that at least the following problems exist in the prior art:

the existing information matching method needs human resources to maintain and expand a great number of similar questions, and operation cost is high.

Disclosure of Invention

In view of this, embodiments of the present invention provide a method and an apparatus for training an information matching model, which complete information matching through the information matching model, and do not need a large amount of human resources to maintain and expand similar statements, so that the operation cost can be effectively reduced.

To achieve the above object, according to an aspect of an embodiment of the present invention, there is provided an information matching model training method, including:

determining a plurality of training information sets, wherein each training information set comprises two pieces of training information and matching labels of the training information sets;

generating corresponding feature score vectors for the training information groups according to the two pieces of training information;

training a preset model to be trained by using the feature score vector and the matching label;

and generating an information matching model according to the training result.

Preferably, generating a corresponding feature score vector for the training information set comprises:

calculating semantic feature scores for the two segments of training information in the training information set;

calculating entity characteristic scores and attribute characteristic scores corresponding to the training information sets;

and generating a feature score vector corresponding to the training information group by using the entity feature score, the attribute feature score and the semantic feature score.

Preferably, the first and second electrodes are formed of a metal,

before calculating the entity feature scores and the attribute feature scores corresponding to the training information sets, the method further includes: generating corresponding entity feature vectors and attribute feature vectors for each section of the training information in the training information;

calculating the entity characteristic score and the attribute characteristic score corresponding to the training information group, wherein the step comprises the following steps: and calculating entity feature scores and attribute feature scores corresponding to the training information sets according to the entity feature vectors and the attribute feature vectors corresponding to each section of the training information.

Preferably, the first and second electrodes are formed of a metal,

the training information is marked with an entity label and an attribute label;

generating corresponding entity feature vectors and attribute feature vectors for each section of the training information in the training information, including:

determining entity characteristics and attribute characteristics included in the training information according to the entity labels and the attribute labels;

converting the entity features into corresponding entity feature vectors according to a preset first entity dictionary;

and converting the attribute features into corresponding attribute feature vectors according to a preset first attribute dictionary.

Preferably, calculating semantic feature scores for the two pieces of training information in the training information group includes:

calculating semantic feature vectors corresponding to each section of the training information in the training information group;

and determining a distance score between the semantic feature vectors corresponding to the two sections of training information, and taking the distance score as a first semantic feature score.

and calculating a second semantic feature score according to the length and the preset length of each section of training information in the training information group.

determining the maximum matching length between the two pieces of training information in the training information group;

and calculating a third semantic feature score according to the maximum matching length and the length of each section of training information in the training information group.

Preferably, the training method of the information matching model further includes: carrying out normalization processing on original training information;

the determining a plurality of training information sets comprises: and determining a plurality of training information groups according to the normalized result.

Preferably, the determining a plurality of training information sets includes:

when first training information is received, primarily screening second training information of a preset number of sections for the first training information from a preset information base;

for each piece of the second training information, performing:

and forming a training information group by using the second training information, the first training information and the matching label determined by the first training information and the second training information.

In a second aspect, an embodiment of the present invention provides an information matching method implemented based on a trained information matching model in any one of the above embodiments, including:

when inquiry information is received, generating corresponding feature score vectors for the inquiry information and a plurality of pieces of information to be matched in a preset matching library respectively;

inputting the feature score vector corresponding to the inquiry information and the feature score vector corresponding to the information to be matched into the information matching model;

and determining matching information corresponding to the inquiry information from the plurality of information to be matched according to the matching result determined by the information matching model.

Preferably, the information matching method further includes:

constructing an inverted index for the preset matching library;

screening a preset amount of information to be matched for the inquiry information from the preset matching library by using the inverted index;

generating a corresponding feature score vector for information to be matched in a preset matching library, wherein the step comprises the following steps;

and generating corresponding feature score vectors for the screened preset number of information to be matched.

Preferably, the step of generating corresponding feature score vectors for the query information and the information to be matched in a preset matching library respectively includes:

calculating corresponding semantic feature scores, entity feature scores and attribute feature scores for the inquiry information and the information to be matched respectively;

and generating a corresponding feature score vector by using the entity feature score, the attribute feature score and the semantic feature score.

Preferably, the information matching method further includes: generating corresponding entity feature vectors and attribute feature vectors for the query information and the information to be matched respectively according to the entity features and attribute features included by the query information and the entity features and attribute features included by the information to be matched;

calculating corresponding entity feature scores and attribute feature scores for the query information, comprising: and calculating an entity feature score and an attribute feature score corresponding to the inquiry information according to the entity feature vector, the attribute feature vector corresponding to the inquiry information and the entity feature vector and the attribute feature vector corresponding to the information to be matched.

Preferably, the first and second electrodes are formed of a metal,

the information to be matched is marked with an entity label and an attribute label;

generating corresponding entity feature vectors and attribute feature vectors for the query information and the information to be matched respectively, including:

determining entity characteristics and attribute characteristics included in the information to be matched according to the entity label marked by the information to be matched and the attribute label;

respectively converting the entity features in the query information and the information to be matched into corresponding entity feature vectors according to a preset second entity dictionary;

and respectively converting the attribute features in the inquiry information and the information to be matched into corresponding attribute feature vectors according to a preset second attribute dictionary.

Preferably, the first and second electrodes are formed of a metal,

the information to be matched is configured with answer information;

the information matching method further comprises the following steps:

and outputting the answer information of the matching information corresponding to the inquiry information to the user who initiates the inquiry information.

In a third aspect, an embodiment of the present invention provides a training apparatus for an information matching model, including: a determination unit, a processing unit and a training unit, wherein,

the determining unit is configured to determine a plurality of training information sets, where each training information set includes two pieces of training information and a matching label of the training information set;

the processing unit is used for generating a corresponding feature score vector for the training information group according to the two pieces of training information in the training information group determined by the determining unit;

the training unit is used for training a preset model to be trained by using the feature score vector generated by the processing unit and the matching label; and generating an information matching model according to the training result.

Preferably, the first and second electrodes are formed of a metal,

the processing unit is used for calculating semantic feature scores for the two sections of training information in the training information group; calculating entity characteristic scores and attribute characteristic scores corresponding to the training information sets; and generating a feature score vector corresponding to the training information group by using the entity feature score, the attribute feature score and the semantic feature score.

In a fourth aspect, an embodiment of the present invention provides an information matching apparatus, including: a vector generation unit and a matching unit, wherein,

the vector generation unit is used for respectively generating corresponding feature score vectors for the inquiry information and the information to be matched in a preset matching library when the inquiry information is received;

the matching unit is used for inputting the feature score vector corresponding to the inquiry information generated by the vector generation unit and the feature score vector corresponding to the information to be matched into the information matching model; and determining matching information corresponding to the query information from the information to be matched according to the matching result determined by the information matching model, wherein the information matching model is obtained based on the training method of the information matching model provided by any embodiment.

One embodiment of the above invention has the following advantages or benefits: the method comprises the steps of generating a corresponding feature score vector and a matching label of a training information group for a training information group, and training a model to be trained, wherein the feature score vector is obtained based on the training information and can well reflect the training information, so that the matching relationship between information can be reflected through the feature score vector and the matching label of the training information group. The information matching can be realized based on the obtained information matching model, a large amount of human resources are not needed for maintaining and expanding similar sentences, and the operation cost can be effectively reduced.

Further effects of the above-mentioned non-conventional alternatives will be described below in connection with the embodiments.

Drawings

The drawings are included to provide a better understanding of the invention and are not to be construed as unduly limiting the invention. Wherein:

FIG. 1 is a schematic diagram of a main flow of a training method of an information matching model according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a main process flow for generating corresponding feature score vectors for a training information set, according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a main process for generating corresponding entity feature vectors and attribute feature vectors for training information according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a main flow of an information matching method according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of the main elements of a training apparatus for information matching models according to an embodiment of the present invention;

fig. 6 is a schematic diagram of main units of an information matching apparatus according to an embodiment of the present invention;

FIG. 7 is an exemplary system architecture diagram in which embodiments of the present invention may be employed;

FIG. 8 is a schematic block diagram of a computer system suitable for use with a server implementing an embodiment of the invention.

Detailed Description

Exemplary embodiments of the present invention are described below with reference to the accompanying drawings, in which various details of embodiments of the invention are included to assist understanding, and which are to be considered as merely exemplary. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the invention. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

The training information refers to text information or voice information used to train the information matching model, which is generally a complete sentence. For example, there are some famous tourist attractions in city A and some food in area B.

The structure of one training information set may be (training information 1, training information 2).

The matching label of the training information group is mainly used for representing the matching condition between two pieces of training information in one training information group, if the two pieces of training information in one training information group are matched (namely the meanings of the two pieces of training information are identical or similar), the matching label is represented as 1, and if the two pieces of training information in one training information group are not matched (namely the meanings of the two pieces of training information are completely different), the matching label is represented as 0.

Fig. 1 is a training method of an information matching model according to an embodiment of the present invention, and as shown in fig. 1, the training method of the information matching model may include the following steps:

s101: determining a plurality of training information sets, wherein each training information set comprises two sections of training information and matching labels of the training information sets;

s102: generating corresponding feature score vectors for the training information groups according to the two sections of training information;

s103: and training a preset model to be trained by using the feature score vector and the matching label to obtain an information matching model.

The model to be trained can be an SVR support vector regression model.

The feature score vector is a vector composed of feature scores calculated from the features of the training information.

In the embodiment shown in fig. 1, the model to be trained is trained by generating the corresponding feature score vector and the matching label of the training information group for the training information group, wherein the feature score vector is obtained based on the training information and can better reflect the training information, and then the matching relationship between the information can be reflected by the feature score vector and the matching label of the training information group. The information matching can be realized based on the obtained information matching model, a large amount of human resources are not needed for maintaining and expanding similar sentences, and the operation cost can be effectively reduced.

In one embodiment of the present invention, as shown in fig. 2, generating a corresponding feature score vector for a training information set may include the following steps:

s201: calculating semantic feature scores for two sections of training information in the training information group;

s202: calculating entity characteristic scores and attribute characteristic scores corresponding to the training information groups;

s203: and generating a feature score vector corresponding to the training information group by using the entity feature score, the attribute feature score and the semantic feature score.

Generally, the features of the training information (sentence) mainly include semantic features, entity features, and attribute features. The training information (sentences) can be embodied more accurately through the semantic features, the entity features and the attribute features, wherein the semantic features are mainly used for expressing the semantics expressed by the training information. For entities and attributes, for example, for training information: the entity of the famous tourist attractions in the city A is the city A, and the attribute of the famous tourist attractions is the tourist attraction. In order to facilitate training of the model to be trained, in the embodiment of the invention, the semantic feature score, the entity feature score and the attribute feature score corresponding to the semantic feature, the entity feature and the attribute feature are calculated respectively, so that training information can be expressed more accurately through the semantic feature score, the entity feature score and the attribute feature score.

The feature score vector corresponding to the training information group is generated by mainly splicing the semantic feature score, the entity feature score and the attribute feature score obtained by the two segments of training information. Accordingly, one expression of the feature score vector obtained by splicing may be (semantic feature score, entity feature score, attribute feature score). It should be noted that the semantic feature scores, the entity feature scores, and the attribute feature scores may be arbitrarily spliced to form feature score vectors (for example, another expression manner of the feature score vectors obtained by splicing may be (entity feature scores, semantic feature scores, and attribute feature scores)), and it is only necessary to ensure that the expression manners of the feature score vectors corresponding to a plurality of training information sets used for training the same model to be trained are consistent.

Through the process, the training information is converted into data, the complexity of the training information is simplified, the entity characteristic score, the attribute characteristic score and the semantic characteristic score can reflect the corresponding training information really, correspondingly, the characteristic score vector can reflect the training information group really, the training can be carried out smoothly, and the accuracy of the training result can be guaranteed.

In an embodiment of the present invention, before calculating the entity feature scores and the attribute feature scores corresponding to the training information sets, the method may further include: generating corresponding entity characteristic vectors and attribute characteristic vectors for each section of training information in the training information; correspondingly, the step of calculating the entity feature score and the attribute feature score corresponding to the training information set comprises the following steps: and calculating entity characteristic scores and attribute characteristic scores corresponding to the training information sets according to the entity characteristic vectors and the attribute characteristic vectors corresponding to each section of training information.

The entity characteristic score corresponding to the training information group is calculated by utilizing entity characteristic vectors corresponding to two sections of training information in the training information group;

and calculating the attribute feature score corresponding to the training information group by utilizing the attribute feature vector corresponding to the two sections of training information in the training information group.

The entity feature score or the attribute feature score may be calculated using the following calculation formula (1), respectively.

Calculating formula (1):

wherein, when F_iWhen representing entity characteristic scores corresponding to the training information group i, A_i1Characterizing an entity feature vector, A, corresponding to a piece of training information in the training information set i_i2Representing an entity feature vector corresponding to another piece of training information in the training information group i; when F is present_iWhen the attribute feature score corresponding to the training information group i is represented, A_i1Characterizing an attribute feature vector, A, corresponding to a piece of training information in a training information set i_i2And characterizing the attribute feature vector corresponding to another piece of training information in the training information group i.

The entity characteristic score and the attribute characteristic score of the training information set are calculated through the process.

The entity characteristic score and the attribute characteristic score can reflect the matching condition between two pieces of training information in the training information group to a certain extent.

In one embodiment of the invention, the training information is labeled with an entity label and an attribute label; accordingly, as shown in fig. 3, generating the corresponding entity feature vector and attribute feature vector for each piece of training information in the training information may include the following steps:

s301: determining entity characteristics and attribute characteristics included in the training information according to the entity labels and the attribute labels;

s302: converting the entity features into corresponding entity feature vectors according to a preset first entity dictionary;

s303: and converting the attribute features into corresponding attribute feature vectors according to a preset first attribute dictionary.

It should be noted that there is no strict sequence between step S302 and step S303.

The first entity dictionary may be an existing entity dictionary, or an entity dictionary composed of all entities in all training data sets.

The first attribute dictionary may be an existing attribute dictionary, or an attribute dictionary composed of all attributes in all training data sets.

The entity label and the attribute label are feature labels set to clearly distinguish the entity and the attribute in the training information. For example, the training information: the name of the tourist attractions in the city A is provided, the corresponding entity is the city A, the attribute is the tourist attraction, the name of the city A can be marked by the mark (T) of the city A, the name of the tourist attraction can be marked by the mark (X) of the tourist attraction, one color can be marked for the city A, and the other color can be marked for the tourist attraction, so that the characteristic marks of the entity and the attribute are realized.

Aiming at a specific implementation mode of converting entity features into corresponding entity feature vectors according to a preset first entity dictionary: and determining the number of elements included in the entity feature vector according to the number of entities included in the first entity dictionary, assigning 1 to the elements at the corresponding positions in the entity feature vector according to the positions of the entity features in the first entity dictionary, and assigning 0 to other elements. For example, if the number of entities included in the first entity dictionary is 50, and the position of the entity feature a in the first entity dictionary is 50 th, the obtained entity feature vector (50 elements in total): (0, 0, 0, … …, 1).

For a specific implementation mode of converting the attribute features into corresponding attribute feature vectors according to a preset first attribute dictionary: and determining the number of elements included in the attribute feature vector according to the number of attributes included in the first attribute dictionary, assigning 1 to the element at the corresponding position in the attribute feature vector according to the position of the attribute feature in the first attribute dictionary, and assigning 0 to other elements. For example, if the number of entities included in the first attribute dictionary is 30, and the position of the attribute feature b in the first attribute dictionary is the 10 th position, the obtained attribute feature vector (30 elements in total): (0, 0, 0, 0, 0, 1, 0, … …, 0).

In one embodiment of the present invention, calculating the semantic feature score for two segments of training information in the training information set may include six implementation manners:

the implementation mode is as follows: calculating semantic feature vectors corresponding to each section of training information in the training information group; and determining a distance score between semantic feature vectors corresponding to the two sections of training information, and taking the distance score as a first semantic feature score.

Calculating semantic feature values corresponding to words in training information by using the following calculation formula (2), and determining semantic feature vectors corresponding to the training information by using the semantic feature values corresponding to all the words included in the training information and a preset word dictionary;

calculating formula (2):

wherein, Y_ijkRepresenting semantic feature values corresponding to words k in the jth section of training information in the training information group i; w is a_kRepresenting the occurrence frequency of a word k in all training information groups; n is_ijRepresenting the total word number included in the jth section of training information in the training information group i; z represents the amount of total training information; z_kThe characterization includes the amount of training information for word k.

The preset word dictionary may be composed of words included in all training data sets, or an existing word dictionary may be selected.

The semantic feature vector corresponding to the training information is determined by utilizing the semantic feature values corresponding to all the words included in the training information and the preset word dictionary, specifically, the number of elements of the semantic feature vector is determined according to the total number of the words included in the preset word dictionary, and the semantic feature value of the word is assigned to the element at the corresponding position in the semantic feature vector according to the position of the word in the training information in the word dictionary. For example, the word dictionary includes 100 words, and the 10 words included in a piece of training information sequentially correspond to the first 10 words in the word dictionary, then the semantic feature vector corresponding to the piece of training information: (Y)_ij1，Y_ij2，…，Y_ij100, 0, …, 0) (i.e. the first 10 elements in the semantic feature vector correspond to the semantic feature values of the words included in the training information, and the last 90 elements are 0).

The specific implementation of determining the distance score between the semantic feature vectors corresponding to the two segments of training information may be:

calculating the distance score between the semantic feature vectors corresponding to the two sections of training information by using the following calculation formula (3);

wherein f is_iRepresenting distance scores corresponding to two pieces of training information in the training information group i, B_i1Representing semantic feature vectors corresponding to a piece of training information in a training information group i, B_i2And representing a semantic feature vector corresponding to another piece of training information in the training information group i.

The implementation mode two is as follows: and calculating a second semantic feature score according to the length and the preset length of each section of training information in the training information group.

The calculating of the second semantic feature score is specifically: and dividing the length of each section of training information by a preset length to obtain a second semantic feature score. For example, if the length of a piece of training information is 10 words and the set maximum length is 30 words, the second semantic feature score is 0.33.

In this implementation, one training information set includes two second semantic feature scores, and accordingly, the corresponding expression manner of the feature score vector may be (second semantic feature score 1, second semantic feature score 2, entity feature score, attribute feature score).

The implementation mode is three: determining the maximum matching length between two sections of training information in a training information group; and calculating a third semantic feature score according to the maximum matching length and the length of each section of training information in the training information group.

The third semantic feature score is calculated specifically: the maximum matching length between two pieces of training information in the training information group is divided by the length of each piece of training information in the training information group. For example, the length of the training information 1 is 10 words, the length of the training information 2 is 15 words, the maximum matching length of the two training information is 5 words, the third semantic feature score corresponding to the training information 1 is 0.5, and the third semantic feature score corresponding to the training information 2 is 0.33. Accordingly, the corresponding expression mode of the feature score vector may be (third semantic feature score 1, third semantic feature score 2, entity feature score, attribute feature score).

The implementation mode is four: the first implementation manner and the second implementation manner are obtained by splicing, and accordingly, the expression manner corresponding to the feature score vector may be (a first semantic feature score, a second semantic feature score 1, a second semantic feature score 2, an entity feature score, and an attribute feature score).

The implementation mode is five: the first implementation manner and the third implementation manner are obtained by splicing, and accordingly, the expression manner corresponding to the feature score vector may be (a first semantic feature score, a third semantic feature score 1, a third semantic feature score 2, an entity feature score, and an attribute feature score).

The implementation mode is six: the first implementation manner, the second implementation manner, and the third implementation manner are obtained by splicing, and accordingly, the expression manner corresponding to the feature score vector may be (a first semantic feature score, a second semantic feature score 1, a second semantic feature score 2, a third semantic feature score 1, a third semantic feature score 2, an entity feature score, and an attribute feature score).

In an embodiment of the present invention, the training method of the information matching model may further include: carrying out normalization processing on original training information; accordingly, determining a plurality of training information sets includes: and determining a plurality of training information groups according to the normalized result. The information consistency can be ensured through the normalization processing, so that the accuracy of the training result is further improved.

The normalization process is to convert several words expressing the same content or the same position into the same word, for example, if the word a, the word B, and the word C all express the same content or the same position, the word B and the word C can be converted into the word a.

In one embodiment of the present invention, determining the plurality of training information sets may comprise: when first training information is received, primarily screening second training information of a preset number of segments for the first training information from a preset information base; for each piece of second training information, performing: and forming a training information group by using the second training information, the first training information and the matching label determined by the first training information and the second training information.

Wherein, the second training information of the preset number of segments is screened out for the first training information, which can be specific: and primarily screening out second training information with the preset segment number closest to the first training information for the first training information by using the reverse index of the preset information base. The predetermined number of stages is, for example, 20 stages.

Through the process, the matching labels corresponding to the training information groups can be effectively distributed in a balanced manner, the number of the non-matching labels is avoided, and the accuracy of the training result is further ensured.

It should be noted that the number of training information sets is generally not less than 5000.

Training a preset specific implementation mode of a model to be trained based on the feature score vectors and the matching labels obtained in the embodiments: inputting a preset model to be trained: training set T { (x)₁,y₁),(x₂,y₂),…,(x_N,y_N) In which x_m∈Rⁿ(x_mRepresents the m-th trainingFeature score vector, R, for information setsⁿA set of feature score vectors corresponding to all sets of training information) are characterized, y_m∈{0,1}(y_mA matching label corresponding to the mth training information group is represented, 0 represents mismatch, and 1 represents match), and m is 1,2, … N; and (3) outputting: and (6) regressing a decision function.

The essence of the training process is as follows: and selecting proper support vector regression parameters for the SVR support vector regression model in an iteration mode to construct a regression decision function. The specific iterative process is consistent with the conventional support vector regression iterative process, and is not described herein again.

As shown in fig. 4, an embodiment of the present invention provides an information matching method implemented by an information matching model, where the information matching method may include the following steps:

s401: when receiving inquiry information, generating corresponding feature score vectors for the inquiry information and a plurality of information to be matched in a preset matching library respectively;

s402: inputting the feature score vector corresponding to the inquiry information and the feature score vector corresponding to the information to be matched into an information matching model;

s403: and determining matching information corresponding to the inquiry information from the information to be matched according to the matching result determined by the information matching model.

The information matching model used in the embodiment of the present invention may be obtained by the training method of the information matching model provided in any one of the embodiments.

The query information may be in a text form or a voice form.

The predetermined matching library may be a database of the question answering system, in which a large amount of question information and answers matching the question information are generally present.

The information to be matched refers to any piece of information stored in the matching library or any piece of information primarily screened from the matching library.

The matching result is generally 1 or 0, wherein 1 represents that the query information is matched with the information to be matched (namely the query information is similar to or the same as the information to be matched); 0 indicates that the query information does not match the information to be matched (i.e., the query information is neither similar nor identical to the information to be matched).

The information matching realized based on the information matching model can effectively improve the information matching efficiency and the accuracy of the information matching.

In an embodiment of the present invention, the information matching method may further include: constructing an inverted index for a preset matching library; screening out a preset amount of information to be matched from a preset matching library by using an inverted index; correspondingly, the step of generating a corresponding feature score vector for the information to be matched in the preset matching library may include: and generating corresponding feature score vectors for the screened preset number of information to be matched.

The information to be matched in the preset number is screened out from the inquiry information in the inverted index mode, so that the primarily screened information to be matched is most similar or close to the inquiry information, unnecessary calculation is avoided, and calculation resources are effectively saved.

In an embodiment of the present invention, the step of generating corresponding feature score vectors for the query information and the information to be matched in the preset matching library respectively may include: respectively calculating corresponding semantic feature scores, entity feature scores and attribute feature scores for the inquiry information and the information to be matched; and generating a corresponding feature score vector by using the entity feature score, the attribute feature score and the semantic feature score.

The training process of the previous information matching model is as follows: the information matching model is trained using a feature score vector generated from a semantic feature score calculated based on semantic features of training information (sentence), an entity feature score calculated based on entity features of the training information (sentence), and an attribute feature score calculated based on attribute features of the training information (sentence). Based on the method, in the information matching process, corresponding semantic feature scores, entity feature scores and attribute feature scores are calculated for inquiry information and information to be matched respectively; and generating a corresponding feature score vector by using the entity feature score, the attribute feature score and the semantic feature score so as to enable the result of information matching realized based on the information matching model to be more accurate.

The entity feature score, the attribute feature score and the semantic feature score are utilized to generate corresponding feature score vectors, and the corresponding feature score vectors are consistent with the feature score vector expression modes corresponding to the training information set in the training process of the information matching model.

In an embodiment of the present invention, the information matching method further includes: generating corresponding entity feature vectors and attribute feature vectors for the query information and the information to be matched respectively according to the entity features and the attribute features included by the query information and the entity features and the attribute features included by the information to be matched; accordingly, the step of calculating the corresponding entity feature score and attribute feature score for the query information may comprise: and calculating entity characteristic scores and attribute characteristic scores corresponding to the inquiry information according to the entity characteristic vector and the attribute characteristic vector corresponding to the inquiry information and the entity characteristic vector and the attribute characteristic vector corresponding to the information to be matched.

The entity characteristics and attribute characteristics included in the query information may be labeled by the user, or labeled by an information matching device in the information matching process.

The generation of the corresponding entity feature vector and attribute feature vector for the query information and the information to be matched can be realized by the existing technology or by the scheme provided by the following embodiments.

The following calculation formula (4) may be adopted to calculate the entity feature score or the attribute feature score corresponding to the query information.

Calculating formula (4):

wherein, when R represents the entity feature score corresponding to the query information, G represents the entity feature vector corresponding to the query information, C_vRepresenting entity feature vectors corresponding to the information v to be matched; when R represents the attribute feature score corresponding to the inquiry information, G represents the attribute feature vector corresponding to the inquiry information,C_vand characterizing attribute feature vectors corresponding to the information v to be matched.

In one embodiment of the invention, the information to be matched is marked with an entity label and an attribute label; correspondingly, generating corresponding entity feature vectors and attribute feature vectors for the query information and the information to be matched respectively comprises the following steps: determining entity characteristics and attribute characteristics included in the information to be matched according to the entity labels and the attribute labels marked by the information to be matched; respectively converting entity features in the query information and the information to be matched into corresponding entity feature vectors according to a preset second entity dictionary; and respectively converting the attribute features in the query information and the information to be matched into corresponding attribute feature vectors according to a preset second attribute dictionary.

Wherein, the second entity dictionary may be: the existing entity dictionary may be: and the entity dictionary is composed of entities contained in a preset matching library.

Wherein the second attribute dictionary may be: the existing attribute dictionary may be: and the attribute dictionary is composed of attributes contained in a preset matching library.

The method for converting the entity features in the query information and the information to be matched into the corresponding entity feature vectors according to the preset second entity dictionary and converting the attribute features in the query information and the information to be matched into the corresponding attribute feature vectors according to the preset second attribute dictionary is consistent with the method for converting the attribute features into the corresponding attribute feature vectors according to the preset first entity dictionary and the preset first attribute dictionary, which are provided in the previous embodiment, and the method is not repeated here.

In one embodiment of the invention, the information to be matched is configured with answer information; accordingly, the information matching method may further include: and outputting the answer information of the matching information corresponding to the inquiry information to the user who initiates the inquiry information. Through the process, automatic question answering service can be realized. Because the information matching method can accurately search the matching information for the inquiry information, the answer information corresponding to the matching information can meet the requirements of the user.

As shown in fig. 5, an embodiment of the present invention provides an information matching model training apparatus 500, where the information matching model training apparatus 500 may include: a determination unit 501, a processing unit 502 and a training unit 503, wherein,

a determining unit 501, configured to determine multiple training information sets, where a training information set includes two pieces of training information and matching labels of the training information sets;

a processing unit 502, configured to generate a corresponding feature score vector for the training information group according to two pieces of training information in the training information group determined by the determining unit 501;

and the training unit 503 is configured to train a preset model to be trained by using the feature score vector and the matching label generated by the processing unit 502 to obtain an information matching model.

In an embodiment of the present invention, the processing unit 502 is configured to calculate semantic feature scores for two segments of training information in a training information group; calculating entity characteristic scores and attribute characteristic scores corresponding to the training information groups; and generating a feature score vector corresponding to the training information group by using the entity feature score, the attribute feature score and the semantic feature score.

In an embodiment of the present invention, the processing unit 502 is configured to generate a corresponding entity feature vector and an attribute feature vector for each piece of training information in the training information; and calculating entity characteristic scores and attribute characteristic scores corresponding to the training information sets according to the entity characteristic vectors and the attribute characteristic vectors corresponding to each section of training information.

In one embodiment of the invention, the training information is labeled with an entity label and an attribute label; a processing unit 502, configured to determine, according to the entity label and the attribute label, an entity feature and an attribute feature included in the training information; converting the entity features into corresponding entity feature vectors according to a preset first entity dictionary; and converting the attribute features into corresponding attribute feature vectors according to a preset first attribute dictionary.

In an embodiment of the present invention, the processing unit 502 is configured to calculate a semantic feature vector corresponding to each piece of training information in the training information set; and determining a distance score between semantic feature vectors corresponding to the two sections of training information, and taking the distance score as a first semantic feature score.

In an embodiment of the present invention, the processing unit 502 is configured to calculate the second semantic feature score according to a length of each piece of training information in the training information set and a preset length.

In an embodiment of the present invention, the processing unit 502 is configured to determine a maximum matching length between two pieces of training information in a training information group; and calculating the three-semantic-characteristic score according to the maximum matching length and the length of each section of training information in the training information group.

In an embodiment of the present invention, the determining unit 501 is configured to perform normalization processing on original training information; and determining a plurality of training information groups according to the normalized result.

In an embodiment of the present invention, the determining unit 501 is configured to, when first training information is received, primarily screen out second training information of a preset number of segments for the first training information from a preset information base; for each piece of second training information, performing: and forming a training information group by using the second training information, the first training information and the matching label determined by the first training information and the second training information.

As shown in fig. 6, an embodiment of the present invention provides an information matching apparatus 600, including: a vector generation unit 601 and a matching unit 602, wherein,

the vector generation unit 601 is configured to generate corresponding feature score vectors for the query information and information to be matched in a preset matching library when the query information is received;

a matching unit 602, configured to input an information matching model to the feature score vector corresponding to the query information generated by the vector generation unit 601 and the feature score vector corresponding to the information to be matched; and according to the matching result determined by the information matching model, determining matching information corresponding to the query information from the information to be matched, wherein the information matching model is obtained based on the training method of the information matching model provided by the embodiment or the information matching model is based on the training device of the information matching model provided by the embodiment.

In one embodiment of the present invention, the information matching apparatus 600 may further include: an index generation unit and a prescreening unit (not shown in the figure), wherein,

the index generating unit is used for constructing an inverted index for a preset matching library;

the preliminary screening unit is used for screening a preset amount of information to be matched from a preset matching library by using the inverted index;

the vector generating unit 601 is configured to generate corresponding feature score vectors for the screened preset number of information to be matched.

In an embodiment of the present invention, the vector generating unit 601 is configured to calculate corresponding semantic feature scores, entity feature scores, and attribute feature scores for the query information and the information to be matched, respectively; and generating a corresponding feature score vector by using the entity feature score, the attribute feature score and the semantic feature score.

In an embodiment of the present invention, the vector generating unit 601 is configured to generate a corresponding entity feature vector and attribute feature vector for the query information according to the entity feature and attribute feature included in the query information; and calculating the entity characteristic score and the attribute characteristic score corresponding to the inquiry information according to the entity characteristic vector and the attribute characteristic vector corresponding to the inquiry information.

In one embodiment of the invention, the information to be matched is marked with an entity label and an attribute label; the vector generating unit 601 is further configured to determine, according to the entity tag and the attribute tag labeled by the information to be matched, an entity feature and an attribute feature included in the information to be matched; converting the entity features in the information to be matched into corresponding entity feature vectors according to a preset second entity dictionary; converting the attribute features in the information to be matched into corresponding attribute feature vectors according to a preset second attribute dictionary; and calculating the entity characteristic score and the attribute characteristic score corresponding to the information to be matched according to the entity characteristic vector and the attribute characteristic vector corresponding to the information to be matched.

In one embodiment of the invention, the information to be matched is configured with answer information; the information matching apparatus 600 further includes: an output unit (not shown in the figure) in which,

and the output unit is used for outputting the answer information of the matching information corresponding to the inquiry information to the user who initiates the inquiry information.

Fig. 7 shows an exemplary system architecture 700 to which the training method of the information matching model or the training apparatus of the information matching model or the information matching method or the information matching apparatus of the embodiment of the present invention can be applied.

As shown in fig. 7, the system architecture 700 may include

terminal devices

701, 702, 703, a network 704, and a server 705. The network 704 serves to provide a medium for communication links between the

terminal devices

701, 702, 703 and the server 705. Network 704 may include various connection types, such as wired, wireless communication links, or fiber optic cables, to name a few.

A user may use the

terminal devices

701, 702, 703 to interact with a server 705 over a network 704, to receive or send messages or the like. The

terminal devices

701, 702, 703 may have installed thereon various communication client applications, such as a shopping-like application, a web browser application, a search-like application, an instant messaging tool, a mailbox client, social platform software, etc. (by way of example only).

The

terminal devices

701, 702, 703 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop portable computers, desktop computers, and the like.

The server 705 may be a server providing various services, such as a background management server (for example only) providing support for a user to initially screen out corresponding second training information to form a training information group by using first training information provided by the

terminal devices

701, 702, and 703, or a background management server (for example only) providing support for a user to match query results with query information provided by the

terminal devices

701, 702, and 703. The backend management server may analyze and perform other processing on the received data such as the training information and the query information, and feed back a processing result (for example, a query result — just an example) to the terminal device.

It should be noted that the training method of the information matching model or the information matching method provided by the embodiment of the present invention is generally executed by the server 705, and accordingly, the training device of the information matching model or the information matching device is generally disposed in the server 705.

It should be understood that the number of terminal devices, networks, and servers in fig. 7 is merely illustrative. There may be any number of terminal devices, networks, and servers, as desired for implementation.

Referring now to FIG. 8, a block diagram of a computer system 800 suitable for use as a server in implementing embodiments of the present invention is shown. The terminal device shown in fig. 8 is only an example, and should not bring any limitation to the functions and the scope of use of the embodiments of the present invention.

As shown in fig. 8, the computer system 800 includes a Central Processing Unit (CPU)801 that can perform various appropriate actions and processes in accordance with a program stored in a Read Only Memory (ROM)802 or a program loaded from a storage section 808 into a Random Access Memory (RAM) 803. In the RAM 803, various programs and data necessary for the operation of the system 800 are also stored. The CPU 801, ROM 802, and RAM 803 are connected to each other via a bus 804. An input/output (I/O) interface 805 is also connected to bus 804.

The following components are connected to the I/O interface 805: an input portion 806 including a keyboard, a mouse, and the like; an output section 807 including a signal such as a Cathode Ray Tube (CRT), a Liquid Crystal Display (LCD), and the like, and a speaker; a storage portion 808 including a hard disk and the like; and a communication section 809 including a network interface card such as a LAN card, a modem, or the like. The communication section 809 performs communication processing via a network such as the internet. A drive 810 is also connected to the I/O interface 805 as necessary. A removable medium 811 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is mounted on the drive 810 as necessary, so that a computer program read out therefrom is mounted on the storage section 808 as necessary.

In particular, according to the embodiments of the present disclosure, the processes described above with reference to the flowcharts may be implemented as computer software programs. For example, embodiments of the present disclosure include a computer program product comprising a computer program embodied on a computer readable medium, the computer program comprising program code for performing the method illustrated in the flow chart. In such an embodiment, the computer program can be downloaded and installed from a network through the communication section 809 and/or installed from the removable medium 811. The computer program executes the above-described functions defined in the system of the present invention when executed by the Central Processing Unit (CPU) 801.

It should be noted that the computer readable medium shown in the present invention can be a computer readable signal medium or a computer readable storage medium or any combination of the two. A computer readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. More specific examples of the computer readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the present invention, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. In the present invention, however, a computer readable signal medium may include a propagated data signal with computer readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, fiber optic cable, RF, etc., or any suitable combination of the foregoing.

The flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams or flowchart illustration, and combinations of blocks in the block diagrams or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

The units described in the embodiments of the present invention may be implemented by software or hardware. The described units may also be provided in a processor, and may be described as: a processor includes a determination unit, a processing unit, and a training unit. Where the names of these units do not in some cases constitute a limitation of the unit itself, for example, a determination unit may also be described as a "unit that determines multiple training information sets".

As another aspect, the present invention also provides a computer-readable medium that may be contained in the apparatus described in the above embodiments; or may be separate and not incorporated into the device. The computer readable medium carries one or more programs which, when executed by a device, cause the device to comprise: determining a plurality of training information sets, wherein each training information set comprises two sections of training information and matching labels of the training information sets; generating corresponding feature score vectors for the training information groups according to the two sections of training information; and training a preset model to be trained by using the feature score vector and the matching label to obtain an information matching model.

The computer readable medium carries one or more programs which, when executed by a device, cause the device to comprise: calculating semantic feature scores for two sections of training information in the training information group; calculating entity characteristic scores and attribute characteristic scores corresponding to the training information groups; and generating a feature score vector corresponding to the training information group by using the entity feature score, the attribute feature score and the semantic feature score.

As another aspect, the present invention also provides a computer-readable medium that may be contained in the apparatus described in the above embodiments; or may be separate and not incorporated into the device. The computer readable medium carries one or more programs which, when executed by a device, cause the device to comprise: when receiving inquiry information, generating corresponding feature score vectors for the inquiry information and information to be matched in a preset matching library respectively; inputting the feature score vector corresponding to the inquiry information and the feature score vector corresponding to the information to be matched into an information matching model; and determining matching information corresponding to the inquiry information from the information to be matched according to the matching result determined by the information matching model.

According to the technical scheme of the embodiment of the invention, the model to be trained is trained by generating the corresponding feature score vector and the matching label of the training information group for the training information group, wherein the feature score vector is obtained based on the training information and can better reflect the training information, so that the matching relation between the information can be reflected through the feature score vector and the matching label of the training information group. The information matching can be realized based on the obtained information matching model, a large amount of human resources are not needed for maintaining and expanding similar sentences, and the operation cost can be effectively reduced.

The above-described embodiments should not be construed as limiting the scope of the invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions can occur, depending on design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A training method of an information matching model is characterized by comprising the following steps:

and training a preset model to be trained by using the feature score vector and the matching label to obtain an information matching model.

2. The method of training an information matching model according to claim 1, wherein generating a corresponding feature score vector for the training information set comprises:

3. The method of training an information matching model according to claim 2,

4. The method of training an information matching model according to claim 3,

the training information is marked with an entity label and an attribute label;

5. The method for training the information matching model according to claim 2, wherein calculating semantic feature scores for the two pieces of training information in the training information set comprises:

6. The method for training the information matching model according to claim 2 or 5, wherein calculating the semantic feature score for the two pieces of training information in the training information set comprises:

calculating a second semantic feature score according to the length and the preset length of each section of training information in the training information group;

and/or the presence of a gas in the gas,

7. The method of training an information matching model according to claim 1,

further comprising: carrying out normalization processing on original training information;

8. The method for training the information matching model according to any one of claims 1 to 5 and 7, wherein the determining a plurality of training information sets comprises:

for each piece of the second training information, performing:

9. An information matching method, comprising:

inputting the feature score vector corresponding to the inquiry information and the feature score vector corresponding to the information to be matched into an information matching model; the information matching model is obtained based on a training method of the information matching model in any one of claims 1 to 8;

10. The information matching method according to claim 9,

further comprising:

constructing an inverted index for the preset matching library;

the method comprises the following steps of generating corresponding feature score vectors for information to be matched in a preset matching library, wherein the steps comprise:

11. The information matching method according to claim 9 or 10, wherein the step of generating corresponding feature score vectors for the query information and the information to be matched in a preset matching library respectively comprises:

12. The information matching method according to claim 11,

further comprising: generating corresponding entity feature vectors and attribute feature vectors for the query information and the information to be matched respectively according to the entity features and attribute features included by the query information and the entity features and attribute features included by the information to be matched;

13. The information matching method according to claim 11,

14. The information matching method according to any one of claims 9, 10, 12, and 13,

the information to be matched is configured with answer information;

the information matching method further comprises the following steps:

15. An apparatus for training an information matching model, comprising: a determination unit, a processing unit and a training unit, wherein,

and the training unit is used for training a preset model to be trained by using the feature score vector generated by the processing unit and the matching label to obtain an information matching model.

16. The apparatus for training an information matching model according to claim 15,

17. An information matching apparatus, comprising: a vector generation unit and a matching unit, wherein,

the matching unit is used for inputting the feature score vector corresponding to the inquiry information generated by the vector generation unit and the feature score vector corresponding to the information to be matched into an information matching model; according to the matching result determined by the information matching model, matching information corresponding to the query information is determined from the information to be matched, wherein the information matching model is obtained based on a training method of the information matching model in any one of claims 1 to 8 or is obtained based on a training device of the information matching model in claim 15 or 16.

18. An electronic device, comprising:

one or more processors;

a storage device for storing one or more programs,

when executed by the one or more processors, cause the one or more processors to implement the method of any one of claims 1-14.

19. A computer-readable medium, on which a computer program is stored, which, when being executed by a processor, carries out the method according to any one of claims 1-14.