CN111444334B

CN111444334B - Data processing method, text recognition device and computer equipment

Info

Publication number: CN111444334B
Application number: CN201910040716.7A
Authority: CN
Inventors: 王剑; 蒋卓人; 孙常龙; 刘晓钟
Original assignee: Alibaba Group Holding Ltd
Current assignee: Alibaba Group Holding Ltd
Priority date: 2019-01-16
Filing date: 2019-01-16
Publication date: 2023-04-25
Anticipated expiration: 2039-01-16
Also published as: CN111444334A

Abstract

The embodiment of the application provides a data processing method, a text recognition device and computer equipment. In the embodiment of the application, a plurality of attribute features are obtained by acquiring a first text sample of a target object and first object association information of the target object and segmenting the first text sample. And constructing a first communication graph by taking the plurality of attribute features and the plurality of object association features determined based on the first object association information as nodes and the association relation between each attribute feature and each object association feature as an edge, and determining a first node vector of each node in the first communication graph. And hitting the first text sample to a first node vector of at least one node in the first communication graph and at least one attribute label corresponding to the first node vector, training a classification model of the target object, and determining a recognition result of the text to be recognized of the target object at least based on the classification model and the first node vector. The embodiment of the application further improves the accuracy of classification model attribute identification.

Description

Data processing method, text recognition device and computer equipment

Technical Field

The embodiment of the application relates to the technical field of networks, in particular to a data processing method, a text recognition device and computer equipment.

Background

The commodity attribute identification aims at deeply mining the attention points and interests of users to specific commodities for a large amount of comment text information of the specific commodities, so as to guide the consumption behaviors of the users or guide merchants to determine the research direction, the business direction and the like of the commodities in the field based on the attention points and interests of the users to the commodities.

At present, commodity attribute identification is mainly carried out by manually labeling attribute labels on comment texts of commodities, so that a large number of training samples are obtained. And then, inputting a training sample into the classification model for model training by combining with a supervised machine learning method to obtain the classification model of the commodity. And then decoding the text to be identified (comment text of the commodity which is not marked by manpower) by using the classification model to obtain an identification result.

However, since only text information is used as a training sample, the accuracy of the attribute identification of the classification model obtained by training is not high, and the error of the identification result is large. Therefore, how to further optimize the recognition performance of the classification model and improve the accuracy of classification model recognition become a problem to be solved.

Disclosure of Invention

The embodiment of the application provides a data processing method, a text recognition device and computer equipment, which can further improve the attribute recognition performance of a classification model and greatly improve the accuracy of attribute recognition.

In a first aspect, an embodiment of the present application provides a data processing method, including:

acquiring a first text sample of a target object and first object association information of the target object;

word segmentation is carried out on the first text sample, and a plurality of attribute features are obtained;

taking the attribute features and the object association features determined based on the first object association information as nodes, taking association relations between the attribute features and the object association features as edges, constructing a first communication graph, and determining a first node vector of each node in the first communication graph;

and hitting the first text sample to a first node vector of at least one node in the first communication graph and at least one attribute label corresponding to the first node vector, training a classification model of the target object, and determining a recognition result of a text to be recognized of the target object based on at least the classification model and the first node vector.

In a second aspect, an embodiment of the present application provides a text recognition method, including:

acquiring a text to be identified of a target object;

determining a target node vector of at least one node in the text hit target connected graph to be identified; the target connected graph is constructed by taking a plurality of attribute features obtained by word segmentation based on a text sample and a plurality of object associated features determined based on object associated information as nodes and taking the associated relation between each attribute feature and each object associated feature as an edge; the target node vector of each node is learned and obtained based on the target communication graph;

determining an attribute identification result of the text to be identified based on at least one target node vector corresponding to the text to be identified and a classification model of the target object; the classification model is obtained through training of a target node vector of at least one node of the target connected graph and at least one corresponding attribute label based on the text sample.

In a third aspect, an embodiment of the present application provides a data processing apparatus, including:

the first acquisition module is used for acquiring a first text sample of a target object and first object association information of the target object;

The second acquisition module is used for word segmentation of the first text sample to obtain a plurality of attribute features;

the first communication diagram generation module is used for constructing a first communication diagram by taking the plurality of attribute features and the plurality of object association features determined based on the first object association information as nodes and taking association relations among the attribute features and the object association features as edges;

a first determining module, configured to determine a first node vector of each node in the first connectivity graph;

and the model training module is used for hitting the first text sample to a first node vector of at least one node in the first communication graph and at least one attribute label corresponding to the first node vector, training a classification model of the target object, and determining a recognition result of a text to be recognized of the target object based on at least the classification model and the first node vector.

In a fourth aspect, in an embodiment of the present application, there is provided a text recognition apparatus, including:

the text to be identified acquisition module is used for acquiring the text to be identified of the target object;

the node vector determining module is used for determining a target node vector of at least one node in the text hit target connected graph to be identified; the target connected graph is constructed by taking a plurality of attribute features obtained by word segmentation based on a text sample and a plurality of object associated features determined based on object associated information as nodes and taking the associated relation between each attribute feature and each object associated feature as an edge; the target node vector of each node is learned and obtained based on the target communication graph;

The text recognition module is used for determining an attribute recognition result of the text to be recognized based on at least one target node vector corresponding to the text to be recognized and the classification model of the target object; the classification model is obtained through training of a target node vector of at least one node of the target connected graph and at least one corresponding attribute label based on the text sample.

In a fifth aspect, in an embodiment of the present application, a computer device is provided, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked for execution by the processing component;

the processing assembly is configured to:

In a sixth aspect, in an embodiment of the present application, a computer device is provided, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked for execution by the processing component;

the processing assembly is configured to:

acquiring a text to be identified of a target object;

Compared with the prior art, the method has the following technical effects:

the embodiment of the application provides a data processing method, a text recognition device and computer equipment. According to the method, based on word segmentation of the obtained first text sample of the target object to obtain a plurality of attribute features and a plurality of object association features determined based on first object association information of the target object, the plurality of object association features are used as nodes, and association relations among the attribute features and the object association features are used as edges to construct a first connected graph. And learn a first node vector for each node in the first communication graph. Different from the traditional pure text classification model training method, the method combines the first text sample and the first object association information through the first connection diagram, so that the first node vector learned by each node in the first connection diagram contains more effective information. Therefore, the first text sample is hit on the first node vector of at least one node in the first communication graph and the corresponding at least one attribute label, the classification model of the target object is trained, the recognition performance is further improved, and the accuracy of attribute recognition is greatly improved.

These and other aspects of the present application will be more readily apparent from the following description of the embodiments.

Drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief description will be given below of the drawings that are needed in the embodiments or the prior art descriptions, and it is obvious that the drawings in the following description are some embodiments of the present application, and that other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow diagram illustrating one embodiment of a data processing method according to the present application;

FIG. 2 illustrates a schematic diagram of path sampling provided in accordance with the present application;

FIG. 3 is a flow chart illustrating yet another embodiment of a data processing method provided herein;

FIG. 4 illustrates a flow diagram of one embodiment of a text recognition method provided in accordance with the present application;

FIG. 5 is a flow chart illustrating yet another embodiment of a text recognition method provided in accordance with the present application;

FIG. 6 is a schematic diagram illustrating one embodiment of a data processing apparatus in accordance with the present application;

FIG. 7 is a schematic diagram of a data processing apparatus according to another embodiment of the present application;

FIG. 8 is a schematic diagram illustrating one embodiment of a text recognition device provided in accordance with the present application;

FIG. 9 is a schematic diagram illustrating a structure of a further embodiment of a text recognition device according to the present application;

FIG. 10 illustrates a schematic diagram of one embodiment of a computer device provided in accordance with the present application;

fig. 11 is a schematic structural view of an embodiment of a computer device according to the present application.

Detailed Description

In order to enable those skilled in the art to better understand the present application, the following description will make clear and complete descriptions of the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application.

In some of the flows described in the specification and claims of this application and in the foregoing figures, a number of operations are included that occur in a particular order, but it should be understood that the operations may be performed in other than the order in which they occur or in parallel, that the order of operations such as 101, 102, etc. is merely for distinguishing between the various operations, and that the order of execution is not by itself represented by any order of execution. In addition, the flows may include more or fewer operations, and the operations may be performed sequentially or in parallel. It should be noted that, the descriptions of "first" and "second" herein are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, and are not limited to the "first" and the "second" being different types.

In order to improve accuracy of a classification model of a target object, the inventor proposes a technical scheme of the application through a series of researches, in the embodiment of the application, a plurality of attribute features are obtained based on word segmentation of a first text sample of the obtained target object, a plurality of object association features determined based on first object association information of the target object are taken as nodes, and association relations among all the attribute features and all the object association features are taken as edges, so that a first communication graph is constructed. And learn a first node vector for each node in the first communication graph. Different from the traditional pure text classification model training method, the method combines the first text sample and the first object association information through the first connection diagram, so that the first node vector learned by each node in the first connection diagram contains more effective information. Therefore, the first text sample is hit on the first node vector of at least one node in the first communication graph and the corresponding at least one attribute label, the classification model of the target object is trained, the recognition performance is further improved, and the accuracy of attribute recognition is greatly improved.

The following description of the embodiments of the present application will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all, of the embodiments of the present application. All other embodiments, which can be made by those skilled in the art based on the embodiments herein without making any inventive effort, are intended to be within the scope of the present application.

Fig. 1 is a flowchart of an embodiment of a data processing method according to an embodiment of the present application. The method may comprise the steps of:

101: acquiring a first text sample of a target object and first object association information of the target object;

in practical applications, the target object may be any commodity, for example, may be a commodity in any field such as clothing, electronic products, articles for daily use, industry, agriculture, etc. It may be understood that the target objects according to the embodiments of the present application include, but are not limited to, commodities in any field, service personnel for online or offline consumption, such as drivers of network taxi taking, couriers or takeaway for express and delivery services, news, messages, or videos in the news media industry, and the like, which are not limited specifically herein.

In practical applications, merchants can define specific attribute tag systems for commodities in order to deeply mine attention points and interest points of users for the commodities. Therefore, the merchant carries out deep analysis and mining on comment texts of the users to obtain the attention points and the interest points of the users on the commodities, and can guide the consumption behaviors of the users or guide the merchant to determine the research direction, the business direction and the like of the commodities in the field based on the attention points and the interests of the users on the commodities. The training samples of each target object in the embodiment of the application need to be obtained by manually labeling the attribute tags of the text. Namely, under the condition that the attribute label system of the target object is determined, the attribute labels corresponding to each text sample need to be manually sorted for manual marking.

The first text sample can be comment information of a user shopping through a shopping website on the Internet and comment information of a purchased commodity or comment information of express delivery; the user can also ask questions and leave messages of the user who has purchased the commodity according to the online commodity, or communicate with comment users; for example, the user may comment information on the driver or the meal delivery person such as the network about car, network about meal delivery consumption, etc. For example, the user may use offline consumption of the online platform to review information of offline stores and consumption experiences. Of course, the method can also be applied to news media industry, such as message information and comment information of users such as internet news media, video websites, technical forums and the like. In practice, text information of the type described above may be included, but is not limited to.

For example, the comment text of a piece of clothing purchased by a user is 'the piece of clothing is fashionable, good in quality and high in cost performance', and the user can know that the attribute label corresponding to the comment text can comprise at least three attribute labels of fashion, quality, cost performance and the like through manual analysis based on a set attribute label system. The user can be considered to be relatively concerned about three dimensions of fashion, quality and cost performance of the commodity. Therefore, when the first text sample is generated, the comment text is marked with at least three attribute tags such as fashion, quality and cost performance.

However, since the text information of the target object is only used as the first text sample, the accuracy of the attribute identification of the classification model obtained by training is not high, and the error of the identification result is large, the inventor proposes to fuse the object association information of the target object with the training sample so as to obtain more effective information, thereby further improving the identification performance of the classification model.

Alternatively, the object association information of the target object may be user information, commodity information, and object information of the target object associated with the target object. For example, commodity information of a commodity, a store selling the commodity, user information for purchasing or browsing the commodity, and the like. The object association information based on the target object can be further mined to the deep effective information of the training sample, for example, a user purchases unified commodities of different shops and writes different comment text information respectively, so that more effective information of the user on the commodity attention points and the interest points can be obtained. For example, the commodity information, whether or not there is cooperative binding use or binding promotion information between commodities, etc. can be obtained according to the commodity set in the store, and effective information can be mined to more dimensions, etc.

Therefore, in the embodiment of the application, the increase of the object association information not only greatly improves the richness expected by the system training, but also can deeply mine and acquire the attention points and the interest points of the user for more dimensionalities of the commodity, and enriches the effective information of the training sample to a greater extent.

It should be understood that the descriptions of "first," "second," and the like in this application are for distinguishing between different messages, devices, modules, etc., and not for indicating a sequential order, nor are the descriptions of "first" and "second" different types.

102: and segmenting the first text sample to obtain a plurality of attribute features.

In practical applications, in order to obtain more effective information, the first text sample needs to be divided into smaller word units, and the first text sample is based on semantic relationships, association relationships and the like among the word units. The method comprises the steps of screening a plurality of words obtained by word segmentation of a first text sample by preprocessing such as part-of-speech tagging and syntactic analysis, removing non-actual meaning words such as o, ya, mock and the like, merging the same words, and obtaining a plurality of attribute features in the first text sample. The attribute features are words which can characterize the characteristics of the text sample, and include the attribute features extracted from the attribute tags marked by the text sample. In the embodiment of the present application, the word segmentation method in the prior art may be used to perform word segmentation processing to obtain a plurality of words, or other existing word extraction techniques may be used, which is not limited herein. And if the text sample is a comment text marked with an attribute tag, the comment text is subjected to word segmentation, and each attribute tag is also formed by at least one word, so that the attribute tag also needs to be subjected to extraction of attribute features.

103: and constructing a first connected graph by taking the attribute features and the object association features determined based on the first object association information as nodes and taking association relations among the attribute features and the object association features as edges.

It can be understood that the words corresponding to the attribute tags may also appear in the comment text, for example, when the comment text is "the garment is fashionable, good in quality and high in cost performance", the comment text contains words corresponding to at least three attribute tags such as "fashionable", "quality", "cost performance", and the like, and therefore, the words such as "the garment", "fashionable", "quality", "cost performance", "good", "high" and the like can be obtained by extracting the attribute features of the comment text. Because the plurality of attribute features obtained based on the first text sample can have part of the same attribute features, when constructing the connected graph, the same words are required to be combined to obtain the plurality of attribute features, each attribute feature is obtained after combination and is used as a node, and the connected graph is constructed according to the association relationship among the attribute features as an edge.

In practical applications, the first object association information may include the target object information, user information and store information associated with the target object, and the like. In the embodiment of the application, the object association information includes, but is not limited to, the above-mentioned object association information, and may also include, for example, commodity category information, binding sales information of commodities and other commodities, and the like.

As an implementation manner, taking the plurality of attribute features and the plurality of object association features determined based on the first object association information as nodes and taking association relations between the attribute features and the object association features as edges, constructing a first connection graph includes:

taking the attribute characteristics as word nodes, taking target object characteristics determined based on the target object information as object nodes, taking user characteristics determined based on the user information as user nodes and taking store characteristics determined based on the store information as store nodes;

and respectively taking the co-occurrence relation among the word nodes and the plurality of attribute characteristics as edges, taking the comment quantity relation between the object node and the word nodes as edges, taking the selling relation between the store node and the commodity node as edges, and taking the user behavior relation among the user node and the commodity node, the user node and the store node and the user node and the word nodes as edges to construct the first communication graph.

The store information associated with the object association information may be a store selling the object, including store information of a store being sold, being prepared for sale, or being a store of the following shelf; the user information may be user information of a user who browses the object or store, purchases the object or store, clicks the object or store, or writes comment text of the object, and the like, and is not particularly limited herein.

In addition, the edges of the connection failure type node may further include class information of the object, whether information such as collaborative use, association, binding promotion and the like is provided between the object nodes as the edges between the object nodes, and information such as upper and lower positions and synonyms between words as the edges between the word nodes, etc., and it is understood that the types of the edges include, but are not limited to, the above-mentioned types, the specific types of the edges can be set according to actual requirements, and the richer the types of the edges are, the higher the corresponding connectivity is.

The information is not particularly limited, and may be appropriately selected according to the validity degree and the association degree of the information.

It may be appreciated that, as described in embodiments of the present application, the user behavior may include, but is not limited to, purchasing, clicking, browsing, writing comments, etc., and thus the user behavior relationship may include purchasing relationship, clicking relationship, browsing relationship, writing comments relationship, etc. The setting may be specifically performed according to actual conditions, and is not particularly limited herein.

By defining the types of various nodes and edges, the user behavior information can be fused into the connected graph. As the user behavior and information such as shops and commodities are fused, more effective information can be contained in the node vector learned later, and the richness of the information contained in the node vector can be further improved. For example, a user who pays more attention to the "fashion" attribute such as the version and style when buying clothes will often pay more attention to the "fashion" attribute when buying other fields of merchandise. For example, in the classification model integrating user information, when the user purchases the same commodity or commodities in different fields again, the degree of enrichment of the historical information can also improve the classification performance of the system. For example, a store may sell multiple products in different areas at the same time, and the same user may purchase different products in multiple areas at the same time. Therefore, connectivity is greatly improved by fusing the connected graphs of the multi-level feature information such as shops, commodities, users and user behaviors.

In the embodiment of the application, the established connectivity graph can realize the establishment of association between different types of nodes, and further excavate user behavior information based on the association relationship between the different types of nodes to obtain richer effective information so as to further improve the identification performance of the classification model.

103: a first node vector is determined for each node in the first connectivity graph.

In the embodiment of the application, the node vector of each node in the connected graph can be determined by the existing graph representation learning method or graph mapping learning method, for example, by adopting a node vector learning model such as works 2vec and skip gram, and the specific process is as follows.

In one implementation manner, the determining the first node vector of each node in the first connectivity graph may include:

taking each node in the first communication graph as a starting point to sample paths respectively, and determining a plurality of node paths;

the first node vector for each node is learned based on the plurality of node paths.

In practical applications, each node vector learning model may be neural network based learning, by traversing each node in the connected graph with each node as a starting point, and determining a node path corresponding to each node. In practical application, the initial node vector of each node is preset, and model training is carried out based on the determined node paths and the initial node vector of each node path, so that the node vector of each node is optimized and learned step by step. The initial node vector may be set randomly or according to a certain preset rule.

As shown in fig. 2, a path sampling schematic diagram is shown, where the connectivity diagram includes A, B, C, D, E, F, each node is connected by an edge, and at least six node paths can be determined with each node as a starting point, for example, the six node paths can be ABC, CBD, DBC, EBF, FBA, BA respectively.

Based on the nodes included in each node path, the node vectors of other nodes on the path can be respectively predicted based on the node vector learning model, for example, in the path ABC, the node vectors of the node a and the node C can be respectively predicted by using the node vector of the node B, and meanwhile, the node a can also predict the node vectors of the node B and the node C, so that the node vector of each node is continuously learned based on the prediction result of each time. The foregoing node vector learning process is the prior art, and will not be described in detail herein.

In practical application, when the node vector is learned, the dimension of the node vector is set according to the richness of the required information quantity. For example, the dimensions of each node vector are set to be 128 dimensions, each of which may be represented by a real value, and are not particularly limited herein.

Optionally, in an implementation manner, the determining a plurality of node paths by taking each node in the first connectivity graph as a starting point to sample the paths includes:

Learning probability distribution of each side in the first communication graph to determine probability weight of each side;

and respectively taking each node in the first communication graph as a starting point, sequentially selecting the next node based on the probability weight of each edge, and determining a plurality of node paths.

As shown in fig. 2, there are 5 sides actually connected to the node B, and each side is connected to one node. In path selection, embodiments of the present application may learn the probability distribution of each edge, thereby determining the probability weight of each edge. In actual learning, an initial probability weight needs to be set for each edge first, that is, if node a only has one edge, then node a will directly travel to node B along the edge, but since node B includes multiple edges, how to determine the next node, in the prior art, random walk algorithm is adopted to determine, that is, randomly determine that any edge connected to node B travels to the next node. But the embodiment example of the application uses the probability weight of each edge to select the next node. For example, the node path corresponding to each node may be determined by selecting the edge with the largest probability weight, or the edge with the smallest probability weight, or the edge approaching the average probability weight, or the like, to determine the next node.

In the embodiment of the application, the node path of each node is determined by combining the existing path sampling algorithm with the probability division of the corresponding edge of the node category. For example, the probability weights of different sides are obtained by combining the classical path sampling algorithms such as the existing radomdalk algorithm, the LINE algorithm, the deep walk algorithm, the MetaPath algorithm and the like and a probability distribution type learning method. In practical applications, the initial probability weight of each edge may be determined according to the probability distribution of the node type or may be set randomly. The different types of edges are also arranged between the different types of nodes, for example, the co-occurrence probability of words between word nodes corresponding to comment texts is used as the initial probability weight of the edges, the purchase frequency is used as the processing probability weight of the edges between commodity nodes and user nodes, the proportion of commodities sold between store nodes and commodity nodes is used as the sales probability weight, and the comment quantity between commodity nodes and word nodes and the commodity quantity is used as the initial probability weight. And after the node vector of each node is actually learned based on the determined node path, the probability distribution among the nodes is changed, and the probability weights of the edges are correspondingly adjusted, so that the probability weights of each edge are learned and obtained.

Optionally, in an implementation manner, each node in the first connection graph is taken as a starting point, a next node is sequentially selected based on the probability weight of each edge, and determining the multiple node paths may include:

taking each node in the first communication graph as a path starting point, and preferentially selecting an edge with the largest probability weight to walk to the next node;

judging whether the number of the steps of any path to be travelled meets a step number threshold value or not;

if the step number threshold is met, determining the current node as a path end point;

and determining a plurality of node paths based on the path starting point and the path ending point corresponding to the path starting point.

Because the data volume of the training samples is large, nodes for constructing the connected graph are also abundant. When the data quantity of the nodes contained in the connected graph is large, a step number threshold value of the node path needs to be set so as to avoid overlarge data calculation quantity caused by overlong step number. For example, the step number threshold is set to 100 steps, and when any node is taken as a starting point and a first hundred nodes are walked to, the first hundred nodes are taken as path end points, so that node paths are obtained. Of course, for the node with poor connectivity, a maximum of 50 steps may actually be walked, and the 50 nodes are used as the path end points to determine the corresponding node paths.

The step number threshold can be set according to the precision requirement of the classification model, and each node actually learns more effective information when the path is longer, so that the accuracy of the classification model obtained in model training is higher. However, if the actual connected graph is large, setting the step number threshold is a preferable scheme for achieving a balance between the operation efficiency and the accuracy.

Optionally, the determining the first node vector of each node in the first connectivity graph may include:

and respectively determining a first semantic vector and a first topic semantic distribution vector corresponding to each node in the first connected graph.

In practical application, it can be known from the foregoing that, based on word vector learning models such as words 2vec and skip gram, a semantic vector of a node can be learned and obtained, and meanwhile, based on a topic distribution learning method of LDA (Latent Dirichlet Allocation, document topic distribution model), a topic semantic distribution vector of each node can be obtained by learning based on a determined node path corresponding to each node, and the semantic vector and the topic distribution learning method are not described in detail herein in the prior art.

Based on the category of the object association information, after learning to obtain 128-dimensional semantic vectors R, each node commodity vector I, user vector U and store vector S can be learned at the same time, and each topic semantic distribution vector can be set to be a 128-bit real number vector at the same time. Each node corresponds to a semantic vector and a topic semantic distribution vector, the topic semantic branch vector can be [ I, U, S ], and the arrangement sequence of the topic semantic branch vector corresponding to each topic vector is not limited. This is only a schematic description and is not specifically limited herein.

105: and hitting the first text sample to a first node vector of at least one node in the first communication graph and at least one attribute label corresponding to the first node vector, training a classification model of the target object, and determining a recognition result of a text to be recognized of the target object based on at least the classification model and the first node vector.

In practical application, if the first text sample is a comment text labeled with an attribute tag, at least one node corresponding to each comment text can be determined. For example, when the comment text is "the clothing is fashionable, good in quality and high in cost performance", words such as "the clothing", "fashionable", "quality", "cost performance", "good", "high", "very" and the like can be obtained by extracting words from the comment text, and nodes corresponding to the extracted words are nodes corresponding to the comment text. It is known that the comment text may hit a node vector of at least seven nodes in the first connected graph.

As an implementation manner, the training the classification model of the target object to determine the recognition result of the text to be recognized of the target object based on at least the classification model and the first node vector by hitting the first text sample on the first node vector of at least one node in the first communication graph and the corresponding at least one attribute tag may include:

Determining a first node vector corresponding to at least one node in the first communication graph hit by the first text sample;

vector fusion is carried out on at least one first node vector corresponding to the first text sample, and a first training text vector of the first text sample is obtained;

and training a classification model of the target object based on the first training text vector and at least one attribute tag corresponding to the first text sample, so as to determine a recognition result of a text to be recognized of the target object based on at least the classification model and the first node vector.

For example, taking the foregoing node vector as 128-dimensional example, vector fusion is performed on the seven node vectors to obtain the training text vector corresponding to the comment text. The actual vector fusion mode can adopt various methods, such as a method of taking the maximum value, the minimum value or the average value of corresponding dimensions of seven node vectors, and the like, and can also be flexibly used based on the combination of the at least two methods. Taking the maximum value for vector fusion as an example, if the maximum value is a 128-dimensional node vector, sequentially determining that the maximum value of the first dimension in the 7 node vectors is the first dimension of the training text vector, determining that the maximum value of the second dimension in the 7 node vectors is the second dimension of the training text vector, … …, and determining that the maximum value of the nth dimension in the 7 node vectors is the nth dimension of the training text vector until the 128-dimensional training text vector is obtained. The vector fusion method of taking the minimum value or taking the average value is similar to the vector fusion method of the maximum value, and is not described herein.

As an implementation manner, the vector fusing the at least one first node vector corresponding to the first text sample, and obtaining the first training text vector of the first text sample may include:

determining at least one first semantic vector of a corresponding node of the first text sample and at least one first topic semantic distribution vector of the corresponding node;

averaging the corresponding dimension values of the at least one first semantic vector to obtain a first training text first sub-vector;

taking the maximum value of the dimension value corresponding to the at least one first topic semantic distribution vector to obtain a first training text second sub-vector;

and splicing the first sub-vector of the first training text and the second sub-vector of the first training text to obtain the first training text vector.

The first sub-vector R 'and the second sub-vector [ I', U ', S' ] are fused into one vector [ R ', I', U ', S' ] through vector splicing, so that the training text vector obtained by the dimension of the training text vector is enlarged.

Each first text sample is manually marked with an attribute tag, so that model training is performed on the classification model of the target object based on the attribute tag corresponding to each first text sample as preset output data of the classification model and the corresponding first training text vector as input data of the classification model. The training process of the classification model is realized by adopting the existing supervised machine learning method, and the classification model can be realized by adopting any existing classification model, such as fitting of decision tree classification models, selection of tree classification models and the like. In addition to deep learning based methods, traditional classification statistical models such as maximum entropy, SVW, random forests, etc. may be used. The selection may be specifically selected according to actual needs, and is not specifically limited herein.

In the embodiment of the application, the first object association information of the target object is combined with the first text sample, and the effective information of the first text sample and the effective information of the first object association information are fused through the constructed first connected graph, so that richer effective information such as effective fused user information, commodity information, store information and the like can be obtained from the first node vector obtained based on the first connected graph learning. Therefore, the first training text vector of the target object and at least one attribute label corresponding to the first training text vector are obtained based on the first node vector, the classification model of the obtained target object is trained, the recognition performance is further improved, and the accuracy of attribute recognition is greatly improved.

In addition, in the embodiment of the application, the probability weights of different edges in the graph are obtained by utilizing the existing path sampling algorithm and combining a probability distribution type learning method, so that the next wandering node is selected based on the probability weights of the edges, and the node path is determined. The effectiveness of the node vector obtained by each node through learning is greatly improved, and therefore the accuracy of attribute identification of the classification model of the target object is further improved.

As an implementation manner, the training the classification model of the target object based on the first training text vector and at least one attribute tag corresponding to the first text sample to determine the recognition result of the text to be recognized of the target object based at least on the classification model and the first node vector may include:

determining a preset output attribute tag vector of at least one attribute tag corresponding to the first training text vector;

based on a first training text vector and a corresponding preset output attribute tag vector, training a classification model of a target object to determine a recognition result of a text to be recognized of the target object based at least on the classification model and the first node vector.

The first training text vector is used as an input value of the classification model, and a preset output label vector of at least one attribute label corresponding to the actual first training text vector can be determined according to the number of image vectors in the set attribute label system.

As an optional implementation manner, the determining the preset output attribute tag vector of the attribute tag corresponding to the first training text vector may include:

Determining an attribute tag vector taking the attribute dimension as a vector dimension based on the attribute dimension of the attribute tag corresponding to the target object;

determining the vector dimension corresponding to each attribute tag in the attribute tag vector;

and generating a preset output attribute label vector corresponding to each first text sample according to the vector dimension of the attribute label corresponding to each first text sample.

For example, if there are 10 tag dimensions of the attribute tag, a 10-dimensional attribute tag vector is determined, and each dimension corresponds to an attribute tag. If a comment text is marked with 3 attribute tags, determining the vector dimension of the corresponding attribute tag, and obtaining a preset output attribute tag vector can be represented as [1,0,1,0,0,0,0,1,0,0]. Here, the preset output attribute tag vector is only schematically described, and is not specifically limited herein.

Optionally, training the classification model of the target object based on the first training text vector and at least one attribute tag corresponding to the first text sample, to determine the recognition result of the text to be recognized of the target object based at least on the classification model and the first node vector may include:

Inputting the first training text vector into the classification model, and outputting at least one prediction attribute label;

judging whether the at least one predicted attribute tag is matched with at least one attribute tag corresponding to the first training text vector;

if yes, a classification model of the target object is obtained, and a recognition result of a text to be recognized of the target object is determined at least based on the classification model and the first node vector;

if not, optimizing model parameters of the classification model based on the difference between the at least one predicted attribute tag and the at least one attribute tag until the at least one predicted attribute tag matches the at least one attribute tag.

In order to optimize the model parameters, a difference between an actual output value of the classification model and a preset output value is required to be compared, and after a first training text vector is used as an input of the classification model to the classification model, the judging whether the at least one predicted attribute label is matched with at least one attribute label corresponding to the first training text vector may be:

judging whether an output difference value of the actual output attribute label vector of the classification model and the preset output attribute label vector corresponding to the first training text vector meets a difference value threshold;

If so, determining that the two images are matched; if not, a mismatch is determined.

The difference threshold may be set according to the actual accuracy requirement. Thus, the difference between the at least one predicted attribute tag and the at least one attribute tag may be an output difference of the preset output attribute tag vector corresponding to the first training text vector and the actual output attribute tag vector of the classification model. And gradually optimizing model parameters of the classification model based on the output difference value. And the actual output value of the classification model is enabled to be more and more close to a preset output value through optimizing model parameters until the output difference value meets a difference value threshold set based on the system precision requirement, and the classification model obtained through training is determined to be the classification model of the target object.

In practical application, the system provided by the embodiment of the application optimizes the model parameters of the classification model and further optimizes the node vector of each node in the connected graph to obtain the optimal training text vector value. And obtaining a classification model of the target object through repeated training. The optimizing model parameters of the classification model based on the difference of the at least one predicted attribute tag and the at least one attribute tag until the at least one predicted attribute tag matches the at least one attribute tag may therefore comprise:

Optimizing a first node vector of each node in the first connected graph based on a difference value between the at least one predicted attribute tag and the at least one attribute tag;

optimizing the first training text vector based on the first node vector after node optimization corresponding to the first text sample;

and gradually optimizing model parameters of the classification model based on the optimized first text sample vector and at least one attribute label corresponding to the first text sample vector.

In an actual application, optimizing the first node vector of each node in the first connected graph based on the difference between the at least one predicted attribute tag and the at least one attribute tag may include:

optimizing probability weights of each edge in the first connected graph based on the difference value between the at least one predicted attribute tag and the at least one attribute tag;

respectively taking each node in the first communication graph as a starting point, sequentially selecting the next node based on the probability weight optimized by each side, and updating the paths of the nodes;

and optimizing the first node vector of each node based on the updated plurality of node paths.

It can be understood that, while optimizing each node vector, the probability weight of each edge is further learned, and through repeated optimization learning, the node vector with optimal each node is obtained, until at least one predicted attribute label matches with the at least one attribute label, the system will store the first communication graph as a target communication graph, the node vector determined by current optimization as an optimal target node vector, and store the trained obtained model parameters, so as to obtain the classification model and the node vector corresponding to the target object.

When the classification model is subjected to attribute identification application, further real-time optimization or dynamic expansion of attribute labels of the classification model is realized in order to continuously improve the identification performance of the classification model. According to the technical scheme, the newly added training data and the newly added attribute labels can be further expanded into the existing target connected graph, and the online optimization training of the classification model is triggered. The training the classification model of the target object to determine the recognition result of the text to be recognized of the target object based at least on the classification model and the first node vector may further include:

acquiring a second text sample newly added by the target object;

and optimizing and training the classification model of the target object based on the second text sample.

Of course, it can be understood that not only the text information of the target object is updated in real time with the lapse of time, but also the object association information of the target object is updated in real time. Thus, the method may further comprise:

Acquiring second object association information newly added by the target object;

the optimizing training of the classification model of the target object based on the second text sample may include:

and carrying out optimization training on the classification model of the target object based on the second text sample and the second object association information.

Or if only updating of object association information occurs, the optimizing training of the classification model of the target object based on the second text sample may include:

and carrying out optimization training on the classification model of the target object based on the second object association information.

It can be understood that if the first text sample is comment text of the attribute tag for the target object; the second text sample may include:

one or more of a new comment text marked with an attribute tag, a new comment text marked with the new attribute tag and a comment text marked with the new attribute tag.

The second object association information may include one or more of a newly added user, a newly added store, and updated target object information, which is not particularly limited herein.

In practical application, the optimization training is performed on the classification model of the target object based on the second text sample and the second object association information, and the mode of incremental training or non-incremental training can be selected for the optimization training of the model according to whether the second text sample is a quality sample.

The incremental training is to expand the newly added text information and/or the newly added attribute labels into the existing target connected graph, and automatically incremental training node vectors and classification models are performed. However, the training mode has the advantages of high model training efficiency and low data processing capacity; however, if the training sample is a high-quality sample, the existing structure of object communication is only further expanded, but the actual original connection structure is not changed, so that part of effective information is lost, and the training effect of the classification model is not ideal. At this time, a non-incremental training mode may be considered, where the non-incremental training is not dependent on the existing target connected graph, in order to sufficiently fuse the effective information, reconstruct the connected graph based on the original training sample and the newly added training sample, and train the classification model based on the newly constructed connected graph, and this training process is the same as the data processing method in the embodiment of fig. 1, and will not be repeated.

Of course, whether the second text sample is good or not may be determined according to the number of second text samples, and if the number threshold is met, it is determined whether the second text sample is a good training sample or not.

However, it will be appreciated that the distinguishing rules for the quality samples are not the same according to different application requirements, and may also be determined based on whether a new attribute tag exists, whether more valid information is contained, and the like, which is not particularly limited herein.

However, in practice, when the classification model is optimally trained, an optimal training mode may be selected according to an expected training effect, if the training speed is fast and the updating efficiency is high, an incremental training mode may be preferentially selected, and if the model identification performance is better, a non-incremental training mode may be preferentially selected, which is not particularly limited herein. Specific embodiments of model optimization training are provided below.

Fig. 3 is a flowchart of another embodiment of a data processing method according to an embodiment of the present application. The method may comprise the steps of:

301: and acquiring a first text sample of a target object and first object association information of the target object.

302: and segmenting the first text sample to obtain a plurality of attribute features.

303: and constructing a first connected graph by taking the attribute features and the object association features determined based on the first object association information as nodes and taking association relations among the attribute features and the object association features as edges.

304: a first node vector is determined for each node in the first connectivity graph.

305: and hitting the first text sample to a first node vector of at least one node in the first communication graph and at least one attribute label corresponding to the first node vector, training a classification model of the target object, and determining a recognition result of a text to be recognized of the target object based on at least the classification model and the first node vector.

306: and acquiring a second text sample newly added to the target object and second object association information newly added to the target object.

307: and fusing a plurality of attribute features obtained by word segmentation of the second text sample and a plurality of object association features determined based on the second object association information into the first connected graph as nodes to construct a second connected graph.

The fusion process of the connected graph is actually to use a plurality of attribute features obtained by word segmentation of the second text sample and a plurality of object association features determined based on the second object association information as nodes, and to fuse association relations between newly added nodes and original nodes as edges, so that each node in the second connected graph comprises the determination of edges of the newly added nodes and the original nodes and the determination of edges between the newly added nodes. However, the added semantic information may contain the change of the association relation between the original nodes, but the fusion mode cannot be modified, so that the effective information of the node vector learned later is less than the data amount of the effective information of the node vector obtained by the non-incremental training mode, and the training effect may be unsatisfactory.

But the second text sample has less quantity and contains less effective information, so that the method is suitable for the incremental training mode, and can not only improve the recognition performance of the classification model, but also realize the rapid and effective optimization training of the classification model.

308: a second node vector is determined for each node in the second connected graph.

The method for determining the second node vector is similar to that described above, and will not be described here again.

309: and respectively hit the first text sample and the second text sample to a second node vector of at least one node in the second connected graph and at least one attribute label corresponding to the second node vector, training a classification model of the target object to obtain an optimized classification model, and determining a recognition result of a text to be recognized of the target object based on at least the classification model and the second node vector.

In practice, since the node vector of each node in the second connected graph is optimized to obtain the second node vector, when the model is trained, the second node vector of at least one node corresponding to the first text sample and the second text sample can be subjected to vector fusion to obtain the second training text vector, so that a better recognition effect is desired to be obtained.

It will be appreciated that the second training text vector may also be obtained based on only the second node vector of the corresponding node of the second text sample, and that the optimization training of the classification model may also be achieved. The selection may be specifically selected according to actual needs, and is not specifically limited herein.

The training process for training the classification model of the target object based on the second training text vector and the at least one attribute tag corresponding to the second text sample is similar to the foregoing. The foregoing details of the implementation manner of the embodiments of the present application have been described in detail, and are not repeated herein.

As an alternative embodiment, the acquiring the second text sample may include:

receiving feedback information generated by an actual recognition result of performing attribute recognition output on the text to be recognized aiming at the classification model; the feedback information is generated when the actual output result is not matched with the expected recognition result corresponding to the text to be recognized;

and acquiring a second text sample generated based on the feedback information.

In the embodiment of the application, the user can carry out online correction on the attribute tag of the text to be identified according to the actual output result of the classification model, and when the classification model identifies that the actual output result of the text to be identified is different from the expected result, the user can generate feedback information and submit the feedback information so as to realize the optimal training of the classification model based on the feedback information.

In practical application, a user can label the text to be identified according to the attribute label corresponding to the expected output result, and trigger the optimization training of the classification model. The background expert can automatically trigger the optimization training module after submitting the error prompt according to the prompt information provided by the user.

Optionally, the acquiring the second text sample generated based on the feedback information may include:

determining a text to be identified corresponding to the feedback information;

acquiring at least one attribute tag corresponding to an expected recognition result of the text to be recognized;

and acquiring a second text sample generated by labeling the text to be recognized by at least one attribute tag corresponding to the expected recognition result.

In practical application, the text to be identified can be the comment text without manually labeling the attribute tag in the comment text of the target object, and the attribute identification can be performed by using a classification model, or the training sample can be used as the text to be identified to test the training model, and the method is not particularly limited.

The expected recognition result of the text to be recognized is that the user expects the attribute label corresponding to the text to be recognized according to the analysis of the text to be recognized. The actual output result is the predicted attribute label after the classification model decodes and classifies the text to be identified. Therefore, when the actual output result of the classification model is different from the expected result, the attribute label corresponding to the text to be identified can be corrected through the feedback information, and the result expected by the user is marked to obtain a second text sample. For example, the text to be identified is 'the clothes is very fit, good in quality and high in cost performance', and the corresponding attribute labels are 'quality' and 'cost performance' obtained through attribute identification. However, the user expects to "fit" a new tag into the classification model, so that the classification model has the ability to decode the new tag "fit", at which time the user may submit corresponding feedback information, thereby labeling the text to be identified with "quality", "fit", and "cost performance", and obtaining a second text sample.

After the classification model is optimized and trained by the method, the classification model can obtain the performance of 'fitting' of the new decoding label, so that the attribute identification performance of the classification model is further improved.

On the other hand, even if no new label is added in the second text sample, only richer text information and object association information are provided, and after the classification model is optimally trained by the method, the accuracy of attribute identification of the classification model can be further improved.

In practice, besides the manual selection of the optimized training mode for triggering the classification model, whether the second text sample is a high-quality training sample or not can be automatically identified, so that the corresponding optimized training mode is automatically triggered. As an optional implementation manner, the fusing, as nodes, the plurality of attribute features obtained by word segmentation of the second text sample and the plurality of object association features determined based on the second object association information into the first connected graph, before constructing the second connected graph, may further include:

determining whether the second text sample is a premium training sample;

if yes, respectively taking a plurality of words obtained by extracting words from the first text sample and the second text sample, the first object association information and the second object association information as nodes, and determining edges for connecting the nodes according to association relations between the nodes to generate a third communication graph;

Determining whether the second text sample is a premium training sample;

if yes, respectively taking a plurality of attribute features obtained by word segmentation of the first text sample and the second text sample, a plurality of object association features determined based on the first object association information and the second object association information as nodes, and constructing a third communication graph by taking association relations between each attribute feature and each object association feature as edges;

determining a third node vector of each node in the third communication graph;

respectively hit the first text sample and the second text sample to a third node vector of at least one node in the third connected graph and at least one attribute label corresponding to the third node vector, training a classification model of the target object to obtain an optimized classification model, and determining a recognition result of a text to be recognized of the target object based on at least the classification model and the third node vector;

and if not, executing the step of fusing the plurality of attribute features obtained by word segmentation of the second text sample and the plurality of object association features determined based on the second object association information into the first connected graph as nodes to construct a second connected graph.

In the embodiment of the application, the online optimization training method for the classification model is provided, and the incremental training or non-incremental training mode can be selected to perform optimization training on the classification model by acquiring the second text sample and/or the second object association information newly added by the target object, so that the capability of the classification model for identifying new attribute labels is further improved, and meanwhile, the accuracy of classification model identification can also be improved.

In addition, the second text sample implemented by the method can be automatically generated based on feedback information of the user, so that automatic optimization training of the classification model is triggered, participation degree and use experience of the user can be improved, and the classification model obtained through optimization is more suitable for the use requirement of the user.

In practical application, besides the incremental training or non-incremental training mode for the target object classification model based on the existing connected graph, a rule model is added to the classification model based on the mapping relation between the second text sample and the attribute label corresponding to the second training sample, and the attribute identification performance of the classification model is supplemented and the accuracy performance is improved through the rule model.

The mapping relation between the second text sample and the attribute label corresponding to the second training sample needs to be manually summarized and defined by a person skilled in the art, a rule model is obtained based on rule training of manual summarization and definition, and when a new attribute label is generated, the rule can be manually modified in real time, so that the attribute recognition performance and accuracy of the rule model are further improved by combining the rule model updated in real time according to the second text sample.

The training process of the rule model may be implemented by any prior art in the field, and will not be described herein.

Fig. 4 is a flowchart of one embodiment of a text recognition method according to an embodiment of the present application. The method may comprise the steps of:

401: and acquiring a text to be identified of the target object.

The text to be identified of the actual target object is distinguished from the text sample by not being marked with the attribute tag manually. The text to be identified can be the comment information of the purchased commodity or the comment information of the express delivery when the user performs shopping through a shopping website on the Internet; the user can also ask questions and leave messages of the user who has purchased the commodity according to the online commodity, or communicate with comment users; for example, the user may comment information on the driver or the meal delivery person such as the network about car, network about meal delivery consumption, etc. For example, the user may use offline consumption of the online platform to review information of offline stores and consumption experiences. Of course, the method can also be applied to news media industry, such as message information and comment information of users such as internet news media, video websites, technical forums and the like. The selection may be specifically performed according to actual requirements, and is not specifically limited herein.

402: and determining a target node vector of at least one node in the text hit target connected graph to be identified.

The target connected graph is constructed by taking a plurality of attribute features obtained by word segmentation based on a text sample and a plurality of object associated features determined based on object associated information as nodes and taking the associated relation between each attribute feature and each object associated feature as an edge; the target node vector of each node is learned based on the target connectivity graph.

In practical applications, the determining the target node vector of at least one node in the text hit target connected graph to be identified may include:

word segmentation is carried out on the text to be recognized, and at least one attribute feature to be recognized is obtained;

and determining a target node vector of at least one node matched with the at least one attribute feature to be identified in the target communication graph.

The actual extraction process of the attribute features of the text to be identified is the same as that described above, and will not be described in detail here.

403: and determining an attribute identification result of the text to be identified based on at least one target node vector corresponding to the text to be identified and the classification model of the target object.

The classification model is obtained through training of a target node vector of at least one node of the target connected graph and at least one corresponding attribute label based on the text sample.

Optionally, in some embodiments, the determining, based on at least one target node vector corresponding to the text to be identified and the classification model of the target object, the attribute identification result of the text to be identified may include:

vector fusion is carried out on at least one target node vector corresponding to the text to be identified, and a text vector to be detected is obtained;

and inputting the text vector to be detected into a classification model of the target object for attribute identification, and obtaining at least one predicted attribute tag corresponding to the text to be identified.

The vector fusion method adopted by the actual text vector to be tested is the same as the vector fusion method used for generating the training text vector, so that the vector dimension is consistent, and the accuracy of attribute identification of the classification model is improved.

Optionally, the target connectivity graph may include:

taking a plurality of attribute features obtained by word segmentation extraction based on a first text sample and a plurality of object association features determined based on first object association information as nodes, and constructing an obtained first communication diagram by taking association relations between each attribute feature and each object association feature as edges;

Or fusing a plurality of attribute features obtained by word segmentation extraction based on a second text sample and a plurality of object association features determined based on second object association information to a second connected graph constructed in the first connected graph as nodes;

or respectively taking a plurality of attribute features obtained by word segmentation extraction based on the first text sample and the second text sample and a plurality of object association features determined based on the first object association information and the second object association information as nodes, and constructing an obtained third connected graph by taking association relations among all the attribute features and all the object association features as edges.

From the foregoing, the obtaining the target node vector of each node may include:

a first node vector obtained based on the first communication map learning, or a second node vector obtained based on the second communication map learning, or a third node vector obtained based on the third communication map learning;

the target training text vector corresponding to the training sample may include:

a first training text vector corresponding to the first text sample obtained based on the first node vector, a second training text vector corresponding to the first text sample and the second text sample obtained based on the second node vector, or a third training text vector corresponding to the first text sample and the second text sample obtained based on the third node vector.

It can be appreciated that the classification model obtained by the training constructs a functional mapping relationship between the input value and the attribute tag. And outputting a corresponding actual output attribute label vector based on the constructed function mapping relation after the actual text vector to be detected is sent to the classification model. Based on the actually output attribute tag vector and the attribute tags corresponding to the vector dimensions of the attribute tag vector, the attribute tag corresponding to each text to be identified can be obtained.

For example, the attribute labels of the target object are respectively materials, upper body effects, styles, patterns, comfort, touch feeling and the like. Each dimension of the constructed attribute tag vector is respectively [ material, upper body effect, style, model, comfort and touch ]. If a text vector to be detected corresponding to any text to be identified is input, and an actual output attribute tag vector is [1,0, 1], determining that the attribute tag corresponding to the text to be identified includes: texture, comfort and touch.

The attribute labels corresponding to a large number of texts to be tested are predicted, and the attention dimension of the user to the target object can be further obtained through statistics and analysis, so that the consumption behavior of the user can be guided, or a merchant can be guided to determine the research direction, the business direction and the like of the commodities in the field.

For example, the specific gravity of each attribute dimension of interest of the commodity user in the clothing field is respectively material 20, upper body effect 10, style 10, model 10, comfort 30, and touch 20, which are obtained through statistics. The user can be determined whether the material of the commodity is comfortable or not, the touch sense is good or not, and the merchant can bias to select comfort level more when designing the clothing, so that the clothing is made of the material with better touch sense.

Of course, if the user information is added in the text to be identified, the attention dimension of each user to the commodity can be further analyzed to realize commodity recommendation based on different users, so that the purchasing behavior of the user is further guided.

The foregoing details of the implementation manner of the embodiments of the present application have been described in detail, and are not repeated herein.

In the embodiment of the application, the attribute identification is performed on the text to be identified of the target object based on the classification model obtained through the training, so that the attention point and the interest point of the user on the commodity in the target field are further obtained, and a foundation is laid for guiding the consumption behavior of the user or guiding a merchant to determine the research direction, the business direction and the like of the commodity in the field.

Fig. 5 is a flowchart of another embodiment of a text recognition method according to an embodiment of the present application. The method may comprise the steps of:

501: and acquiring a text to be identified of the target object.

502: and determining a target node vector of at least one node in the text hit target connected graph to be identified.

503: and carrying out vector fusion on at least one target node vector corresponding to the text to be identified to obtain a text vector to be detected.

504: and inputting the text vector to be detected into a classification model of the target object for attribute identification, and obtaining at least one predicted attribute tag corresponding to the text to be identified.

505: and matching at least one predicted attribute tag with at least one attribute tag corresponding to the expected recognition result of the text to be recognized.

The at least one predicted attribute tag is actually matched with at least one attribute tag corresponding to the user expected recognition result, the attribute tag can be matched, or converted into the attribute tag vector to be subjected to difference comparison, if the output difference value meets a matching threshold value, the matching is considered, otherwise, the matching is not considered.

506: and if the matching is unsuccessful, generating feedback information to optimally train the classification model based on the feedback information.

In the embodiment of the application, a user can input a desired recognition result expected by the user and match the desired recognition result with an actual output recognition result based on the classification model, and if the desired recognition result is not matched with the actual output recognition result, feedback information is generated, so that optimization training of the classification model is triggered based on the feedback information. Of course, the feedback information may also be generated by the user by judging whether the desired recognition result matches with the actual output recognition result based on the classification model. Therefore, the optimization training can be carried out on the classification model on line, so that the optimized classification model meets the actual requirements of users, and the user experience is further improved while the attribute identification performance of the classification model is improved.

Fig. 6 is a schematic structural diagram of an embodiment of a data processing apparatus according to an embodiment of the present application.

The apparatus may include:

the first obtaining module 601 is configured to obtain a first text sample of a target object and first object association information of the target object.

And the second obtaining module 602 is configured to segment the first text sample to obtain a plurality of attribute features.

The first connection graph generating module 603 is configured to construct a first connection graph by using the plurality of attribute features and the plurality of object association features determined based on the first object association information as nodes and using association relations between each attribute feature and each object association feature as edges.

A first determining module 604 is configured to determine a first node vector of each node in the first connectivity graph.

The model training module 605 is configured to hit the first text sample on a first node vector of at least one node in the first communication graph and a corresponding at least one attribute tag, train a classification model of the target object, and determine a recognition result of a text to be recognized of the target object based at least on the classification model and the first node vector.

Therefore, the target object information is set as the object node, the user information is set as the user node, and the store information is set as the store node. And mining user behavior information, such as commodity information purchased by a user, commodity information reviewed by the user, commodity information browsed by the user or store information, and further can comprise store commodity information sold by the store, store click or commodity information clicked by the user and the like, so that association relations among different types of nodes are obtained.

As an implementation manner, the first connectivity map generating module 603 may include:

In an implementation manner, the first determining module 604 may include:

and the node path determining unit is used for respectively taking each node in the first communication graph as a starting point to sample paths and determining a plurality of node paths.

A first node vector determination unit configured to learn a first node vector of each node based on the plurality of node paths.

In an implementation manner, the node path determining unit may specifically be configured to:

Optionally, in an implementation manner, each node in the first connection graph is taken as a starting point, a next node is sequentially selected based on the probability weight of each edge, and determining multiple node paths may be specifically used for:

As an implementation manner, the model training module 605 may include:

a first node vector determining unit, configured to determine that the first text sample hits a first node vector corresponding to at least one node in the first communication graph;

the first training text vector acquisition unit is used for carrying out vector fusion on at least one first node vector corresponding to the first text sample to obtain a first training text vector of the first text sample;

and the model training unit is used for training a classification model of the target object based on the first training text vector and at least one attribute label corresponding to the first text sample so as to determine a recognition result of the text to be recognized of the target object based on at least the classification model and the first node vector.

As an implementation manner, the first training text vector obtaining unit may specifically be configured to:

As an implementation manner, the model training unit may include:

a preset output vector determining subunit, configured to determine a preset output attribute tag vector of at least one attribute tag corresponding to the first training text vector;

the model training subunit is used for training the classification model of the target object based on the first training text vector and the corresponding preset output attribute label vector.

As an alternative embodiment, the preset output vector determination subunit may specifically be configured to:

Optionally, the model training unit may include:

a first output subunit, configured to input the first training text vector into the classification model, and output at least one prediction attribute tag;

a first judging subunit, configured to judge whether the at least one predicted attribute tag matches with at least one attribute tag corresponding to the first training text vector; if yes, triggering a classification model determining unit; if not, triggering a model parameter optimization unit.

A classification model determining subunit, configured to obtain a classification model of the target object, so as to determine a recognition result of a text to be recognized of the target object based at least on the classification model and the first node vector;

and the model parameter optimization subunit is used for optimizing the model parameters of the classification model based on the difference value between the at least one predicted attribute tag and the at least one attribute tag until the at least one predicted attribute tag is matched with the at least one attribute tag.

In order to optimize the model parameters, a difference between an actual output value of the classification model and a preset output value needs to be compared, and after the first training text vector is used as the input of the classification model to the classification model, the first judging subunit may specifically be used for:

In practical application, the system provided by the embodiment of the application optimizes the model parameters of the classification model and further optimizes the node vector of each node in the connected graph to obtain the optimal training text vector value. Through repeated training, a classification model of the target object is obtained, so that the model parameter optimization subunit can be specifically used for:

In practical applications, the optimizing the first node vector of each node in the first connected graph based on the difference between the at least one predicted attribute tag and the at least one attribute tag may specifically be used to:

When the classification model is subjected to attribute identification application, further real-time optimization or dynamic expansion of attribute labels of the classification model is realized in order to continuously improve the identification performance of the classification model. According to the technical scheme, the newly added training data and the newly added attribute labels can be further expanded into the existing target connected graph, and the online optimization training of the classification model is triggered. After the model training module 605, it may further include:

The second text sample acquisition module is used for acquiring a second text sample newly added by the target object;

and the model optimization training module is used for carrying out optimization training on the classification model of the target object based on the second text sample.

Of course, it can be understood that not only the text information of the target object is updated in real time with the lapse of time, but also the object association information of the target object is updated in real time. Thus, it may further comprise:

and the second object association information acquisition module is used for acquiring second object association information newly added by the target object.

The model optimization training module can be specifically used for:

Or if only an update of the object association information occurs, the model optimization module may be specifically configured to:

Fig. 7 is a schematic structural diagram of another embodiment of a data processing apparatus according to an embodiment of the present application. The apparatus may include:

the first obtaining module 701 is configured to obtain a first text sample of a target object and first object association information of the target object.

And the second obtaining module 702 is configured to segment the first text sample to obtain a plurality of attribute features.

The first connection graph generating module 703 is configured to construct a first connection graph with the plurality of attribute features and the plurality of object association features determined based on the first object association information as nodes and with association relationships between the attribute features and the object association features as edges.

A first determining module 704 is configured to determine a first node vector of each node in the first connectivity graph.

The model training module 705 is configured to hit the first text sample on a first node vector of at least one node in the first communication graph and a corresponding at least one attribute tag, train a classification model of the target object, and determine a recognition result of a text to be recognized of the target object based at least on the classification model and the first node vector.

A third obtaining module 706, configured to obtain a second text sample newly added to the target object and second object association information newly added to the target object.

And a model optimization training module 707, configured to perform optimization training on the classification model of the target object based on the second text sample and the second object association information.

Model optimization training module 707 may include:

and a second connected graph generating unit 711, configured to fuse, as nodes, a plurality of attribute features obtained by word segmentation of the second text sample and a plurality of object association features determined based on the second object association information into the first connected graph, and construct a second connected graph.

A second node vector determining unit 712, configured to determine a second node vector of each node in the second connected graph.

A first model optimizing unit 713, configured to hit the first text sample and the second text sample with a second node vector and at least one attribute tag corresponding to at least one node in the second connected graph, respectively, train the classification model of the target object to obtain an optimized classification model, and determine a recognition result of the text to be recognized of the target object based on at least the classification model and the second node vector.

As an alternative embodiment, the third obtaining module 706 may specifically be configured to:

and receiving feedback information generated by the actual recognition result generated by carrying out attribute recognition output on the text to be recognized according to the classification model.

The feedback information is generated when the actual output result is not matched with the expected recognition result corresponding to the text to be recognized.

And acquiring a second text sample generated based on the feedback information.

Optionally, the acquiring the second text sample generated based on the feedback information may specifically be used to:

Determining a text to be identified corresponding to the feedback information;

After the classification model is optimized and trained by the method, the performance of decoding the new label can be obtained by the classification model, so that the attribute identification performance of the classification model is further improved.

On the other hand, in the instant second text sample, no new label is added, only richer text information and object association information are provided, and after the classification model is optimally trained by the method, the accuracy of attribute identification of the classification model can be further improved.

In practice, besides the manual selection of the optimized training mode for triggering the classification model, whether the second text sample is a high-quality training sample or not can be automatically identified, so that the corresponding optimized training mode is automatically triggered. As an alternative embodiment, before the second communication map generating unit 711, it may further include:

And the second judging unit is used for determining whether the second text sample is a high-quality training sample or not. If yes, triggering a third communication diagram generating unit; if not, the second communication map generation unit 711.

A third connected graph generating unit, configured to respectively use a plurality of words obtained by extracting words from the first text sample and the second text sample, the first object association information and the second object association information as nodes, and determine edges of connecting nodes according to association relations between the nodes, so as to generate a third connected graph;

a third node vector determining unit, configured to determine a third node vector of each node in the third communication graph;

and the second model optimization training unit is used for respectively hit the first text sample and the second text sample into a third node vector of at least one node in the third connected graph and at least one corresponding attribute label, training the classification model of the target object to obtain an optimized classification model, and determining the recognition result of the text to be recognized of the target object based on at least the classification model and the third node vector.

In the embodiment of the application, the online optimization training method for the classification model is provided, and the classification model can be optimized and trained in an incremental or non-incremental mode by acquiring the second text sample and/or the second object association information newly added by the target object, so that the capability of the classification model for identifying the new attribute label is further improved, and meanwhile, the accuracy of classification model identification can also be improved.

Fig. 8 is a schematic structural diagram of another embodiment of a text recognition device according to an embodiment of the present application. The apparatus may include:

a text to be identified obtaining module 801, configured to obtain a text to be identified of a target object;

a node vector determining module 802, configured to determine a target node vector of at least one node in the target connected graph hit by the text to be identified.

The text recognition module 803 is configured to determine an attribute recognition result of the text to be recognized based on at least one target node vector corresponding to the text to be recognized and the classification model of the target object.

Optionally, in some embodiments, the text recognition module 803 may include:

the text vector to be detected determining unit is used for carrying out vector fusion on at least one target node vector corresponding to the text to be identified to obtain a text vector to be detected;

and the text recognition unit is used for inputting the text vector to be detected into the classification model of the target object to perform attribute recognition and obtaining at least one predicted attribute tag corresponding to the text to be recognized.

In practical applications, the node determining module 802 may specifically be configured to:

extracting words from the text to be recognized to obtain at least one word to be detected;

and determining at least one node matched with the at least one word to be detected in the target communication graph.

Optionally, the target connectivity graph may include:

Fig. 9 is a schematic structural diagram of another embodiment of a text recognition device according to an embodiment of the present application. The apparatus may include:

the text to be recognized acquiring module 901 is configured to acquire a text to be recognized of a target object.

A node vector determining module 902, configured to determine a target node vector of at least one node in the target connected graph hit by the text to be identified.

The text recognition module 903 is configured to determine an attribute recognition result of the text to be recognized based on at least one target node vector corresponding to the text to be recognized and the classification model of the target object.

Optionally, in some embodiments, the text recognition module 903 may include:

and the text vector to be detected determining unit 911 is configured to perform vector fusion on at least one target node vector corresponding to the text to be identified, so as to obtain a text vector to be detected.

The text recognition unit 912 is configured to input the text vector to be detected into a classification model of the target object to perform attribute recognition, and obtain at least one predicted attribute tag corresponding to the text to be recognized.

And the matching module 904 is configured to match at least one predicted attribute tag with at least one attribute tag corresponding to a desired recognition result of the text to be recognized.

And the feedback information generating module 905 is configured to generate feedback information to perform optimization training on the classification model based on the feedback information if the matching is unsuccessful.

Fig. 10 is a schematic structural diagram of an embodiment of an electronic device provided in the embodiments of the present application, where the server may include a processing component 1001 and a storage component 1002. The storage component 1002 is configured to store one or more computer instructions, where the one or more computer instructions are provided for execution by the processing component.

The processing component 1001 may be configured to:

Wherein the processing component 1001 may include one or more processors to execute computer instructions to perform all or part of the steps in the methods described above. Of course, the processing component may also be implemented as one or more Application Specific Integrated Circuits (ASICs), digital Signal Processors (DSPs), digital Signal Processing Devices (DSPDs), programmable Logic Devices (PLDs), field Programmable Gate Arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic elements for executing the methods described above.

The storage component 1002 is configured to store various types of data to support operations in a server. The memory component may be implemented by any type or combination of volatile or nonvolatile memory devices such as Static Random Access Memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic or optical disk.

Of course, the computer device may naturally also include other components, such as input/output interfaces, communication components, and the like.

The input/output interface provides an interface between the processing component and a peripheral interface module, which may be an output device, an input device, etc.

The communication component is configured to facilitate communication between the server and other devices, either wired or wireless, such as communication with a terminal.

The embodiment of the present application further provides a computer readable storage medium storing a computer program, where the computer program when executed by a computer may implement the data processing method of the embodiment shown in fig. 1 and fig. 3.

Fig. 11 is a schematic structural diagram of one embodiment of a computer device provided in the embodiments of the present application, where the terminal device may include a processing component 1101 and a storage component 1102. The storage component 1102 is configured to store one or more computer instructions, wherein the one or more computer instructions are provided for execution by the processing component.

The processing component 1101 may be configured to:

acquiring a text to be identified of a target object;

Wherein the processing component 1101 may include one or more processors to execute computer instructions to perform all or part of the steps of the methods described above. Of course, the processing component may also be implemented as one or more Application Specific Integrated Circuits (ASICs), digital Signal Processors (DSPs), digital Signal Processing Devices (DSPDs), programmable Logic Devices (PLDs), field Programmable Gate Arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic elements for executing the methods described above.

The storage component 1102 is configured to store various types of data to support operations in a server. The memory component may be implemented by any type or combination of volatile or nonvolatile memory devices such as Static Random Access Memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic or optical disk.

The embodiment of the application further provides a computer readable storage medium, and a computer program is stored, and when the computer program is executed by a computer, the method for identifying the text in the embodiment shown in fig. 4 and 5 can be realized.

It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described systems, apparatuses and units may refer to corresponding procedures in the foregoing method embodiments, which are not repeated herein.

The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.

From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.

Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present application, and are not limiting thereof; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the corresponding technical solutions.

Claims

1. A method of data processing, comprising:

acquiring a first text sample of a target object and first object association information of the target object, wherein the first text sample comprises comment text of a user on the target object, and the first object association information comprises one or more of the following: user information, merchandise information, or object information of the target object associated with the target object;

2. The method of claim 1, wherein after the training the classification model of the target object to determine the recognition result of the text to be recognized of the target object based at least on the classification model and the first node vector by hit the first text sample on the first node vector of at least one node in the first communication graph and the at least one attribute tag corresponding to the first text sample, further comprises:

acquiring a second text sample newly added by the target object;

3. The method according to claim 2, wherein the method further comprises:

the optimizing training of the classification model of the target object based on the second text sample comprises the following steps:

4. The method of claim 3, wherein the optimally training the classification model of the target object based on the second text sample and the second object association information comprises:

Fusing a plurality of attribute features obtained by word segmentation of the second text sample and a plurality of object association features determined based on the second object association information into the first connected graph as nodes to construct a second connected graph;

determining a second node vector of each node in the second connected graph;

and respectively hit the first text sample and the second text sample to a second node vector of at least one node in the second connected graph and at least one attribute label corresponding to the second node vector, training a classification model of the target object to obtain an optimized classification model, and determining a recognition result of a text to be recognized of the target object based on at least the classification model and the second node vector.

5. The method according to claim 4, wherein the fusing the plurality of attribute features obtained by word segmentation of the second text sample and the plurality of object association features determined based on the second object association information as nodes into the first connected graph, before constructing a second connected graph, further comprises:

determining whether the second text sample is a premium training sample;

Determining a third node vector of each node in the third communication graph;

and if not, executing the step of fusing the plurality of attribute features obtained by word segmentation of the second text sample and the plurality of object association features determined based on the second object association information into the first communication graph as nodes to construct a second communication graph.

6. The method of claim 5, wherein the determining whether the second text sample is a premium training sample comprises:

judging whether the number of the second text samples meets a number threshold;

and if the number threshold is met, determining that the second text sample is a premium training sample.

7. The method of claim 2, wherein the obtaining a second text sample comprises:

and acquiring a second text sample generated based on the feedback information.

8. The method of claim 7, wherein the obtaining a second text sample generated based on the feedback information comprises:

determining a text to be identified corresponding to the feedback information;

9. The method of claim 2, wherein the first text sample is comment text of the target object tagged with an attribute tag; the second text sample includes:

10. The method of claim 1, wherein the first object association information includes the target object information, and user information and store information associated with the target object.

11. The method of claim 10, wherein constructing a first connectivity graph with the plurality of attribute features and the plurality of object-related features determined based on the first object-related information as nodes and with association relationships between the respective attribute features and the respective object-related features as edges comprises:

and respectively taking the co-occurrence relation among the word nodes and the plurality of attribute characteristics as edges, taking the comment quantity relation between the object nodes and the word nodes as edges, taking the selling relation between the store nodes and the object nodes as edges, and taking the user behavior relation among the user nodes and the object nodes, the user nodes and the store nodes and the user nodes and the word nodes as edges to construct the first communication graph.

12. The method of claim 1, wherein the determining a first node vector for each node in the first connectivity graph comprises:

13. The method of claim 12, wherein the determining a plurality of node paths by sampling paths starting from each node in the first connectivity graph, respectively, comprises:

14. The method of claim 13, wherein the determining a plurality of node paths with each node in the first connected graph as a starting point, sequentially selecting a next node based on the probability weight of each edge comprises:

15. The method of claim 1, wherein the training the classification model of the target object to determine the recognition result of the text to be recognized of the target object based at least on the classification model and the first node vector by hit the first text sample to a first node vector and a corresponding at least one attribute tag of at least one node in the first communication graph comprises:

16. The method of claim 15, wherein the training the classification model of the target object based on the first training text vector and at least one attribute tag corresponding to the first text sample to determine a recognition result of the text to be recognized of the target object based at least on the classification model and the first node vector comprises:

17. The method of claim 16, wherein optimizing model parameters of the classification model based on the difference between the at least one predicted attribute tag and the at least one attribute tag until the at least one predicted attribute tag matches the at least one attribute tag comprises:

and gradually optimizing model parameters of the classification model based on the optimized first training text vector and at least one attribute label corresponding to the first training text vector.

18. The method of claim 17, wherein optimizing the first node vector for each node in the first connected graph based on the difference between the at least one predicted attribute tag and the at least one attribute tag comprises:

19. The method of claim 15, wherein the determining a first node vector for each node in the first connectivity graph comprises:

20. The method of claim 19, wherein vector fusing the at least one first node vector corresponding to the first text sample to obtain a first training text vector for the first text sample comprises:

determining at least one first semantic vector corresponding to the first text sample and at least one corresponding first topic semantic distribution vector;

and vector stitching is carried out on the first sub-vector of the first training text and the second sub-vector of the first training text, so that the first training text vector is obtained.

21. A method of text recognition, comprising:

Acquiring a text to be identified of a target object;

determining a target node vector of at least one node in the text hit target connected graph to be identified; the target connected graph is constructed by taking a plurality of attribute features obtained by word segmentation based on a text sample and a plurality of object associated features determined based on object associated information as nodes and taking the associated relation between each attribute feature and each object associated feature as an edge; the target node vector of each node is learned and obtained based on the target communication graph; wherein the text sample comprises user comment text of the target object, and the object association information comprises one or more of the following: user information, merchandise information, or object information of the target object associated with the target object;

22. The method of claim 21, wherein the determining the attribute identification result of the text to be identified based on the at least one target node vector corresponding to the text to be identified and the classification model of the target object comprises:

23. The method according to claim 22, wherein the inputting the text vector to be detected into the classification model of the target object for attribute recognition, after obtaining at least one predicted attribute tag corresponding to the text to be recognized, further comprises:

matching the at least one predicted attribute tag with at least one attribute tag corresponding to the expected recognition result of the text to be recognized;

and if the matching is unsuccessful, generating feedback information to optimally train the classification model based on the feedback information.

24. The method of claim 21, wherein the determining a target node vector for at least one node in the text-to-identify corresponding target connected graph comprises:

25. The method of claim 21, wherein the target connectivity map comprises:

26. The method of claim 25, wherein the obtaining the target node vector for each node comprises:

the target training text vector corresponding to the text sample comprises:

27. A data processing apparatus, comprising:

the first acquisition module is used for acquiring a first text sample of a target object and first object association information of the target object, wherein the first text sample comprises comment text of a user on the target object, and the first object association information comprises one or more of the following: user information, merchandise information, or object information of the target object associated with the target object;

28. A text recognition device, comprising:

the node vector determining module is used for determining a target node vector of at least one node in the text hit target connected graph to be identified; the target connected graph is constructed by taking a plurality of attribute features obtained by word segmentation based on a text sample and a plurality of object associated features determined based on object associated information as nodes and taking the associated relation between each attribute feature and each object associated feature as an edge; the target node vector of each node is learned and obtained based on the target communication graph; wherein the text sample comprises user comment text of the target object, and the object association information comprises one or more of the following: user information, merchandise information, or object information of the target object associated with the target object;

29. A computer device comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked for execution by the processing component;

the processing assembly is configured to:

30. A computer device comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked for execution by the processing component;

the processing assembly is configured to:

acquiring a text to be identified of a target object;