CN113821592A

CN113821592A - Data processing method, device, equipment and storage medium

Info

Publication number: CN113821592A
Application number: CN202110702133.3A
Authority: CN
Inventors: 罗锦文; 郭伟东
Original assignee: Tencent Technology Shenzhen Co Ltd
Current assignee: Tencent Technology Shenzhen Co Ltd
Priority date: 2021-06-23
Filing date: 2021-06-23
Publication date: 2021-12-21
Anticipated expiration: 2041-06-23

Abstract

The embodiment of the application discloses a data processing method, a device, equipment and a storage medium, wherein the method comprises the following steps: acquiring multimedia data associated with a target object identifier; calling an entity recognition model to perform entity recognition on the multimedia data, and determining entity words in the multimedia data and entity word types of the entity words according to recognition results of the entity recognition model; according to the entity words and the entity word types, object data processing is carried out on the target object identification; the entity recognition model comprises a coding module and a decoding module, wherein the coding module is used for determining a feature vector of a target character in the multimedia data, the feature vector of the target character is used for representing character features of the target character when the target character is in the multimedia data, the corresponding feature vectors of the target character under different multimedia data are different, and the decoding module is used for carrying out entity recognition on the target character according to the feature vector of the target character. The recognition of fine-grained entity words can be realized, and therefore the recognition precision is improved.

Description

Data processing method, device, equipment and storage medium

Technical Field

The present application relates to the field of computer technologies, and in particular, to a data processing method, apparatus, device, and storage medium.

Background

Named entities refer to words in text that describe entities, such as person names, place names, organizational names, product names, etc., and are referred to as named entities (Name entities). The named entities are often the most concerned words of people in sentences, and the entity identification of the named entities is an important basic tool in the application fields of information extraction, question-answering systems, syntactic analysis, machine translation and the like, and plays an important role in the process of bringing the natural language processing technology into practical use. The definition of the named entity is different from person to person and from scene to scene, for example, in the E-commerce field, the name of a commodity can be defined as the named entity, and in the medical field, the name of a disease can be defined as the named entity.

At present, the method for determining the named entity is generally performed through an existing dictionary, the existing dictionary comprises a large number of entity words of various types, and the current method for determining the entity words through the dictionary has certain limitations, for example, the content of the dictionary is limited, the quantity of the entity words is very large, and the method for determining the entity words in the sentence based on the dictionary is not comprehensive and accurate enough.

Disclosure of Invention

The embodiment of the application provides a data processing method, a data processing device, data processing equipment and a storage medium, and entity word recognition can be comprehensively and accurately realized.

In one aspect, an embodiment of the present application discloses a data processing method, where the method includes:

acquiring multimedia data associated with a target object identifier;

calling an entity recognition model to perform entity recognition on the multimedia data, and determining entity words in the multimedia data and entity word types of the entity words according to recognition results of the entity recognition model;

according to the entity words and the entity word types, object data processing is carried out on the target object identification;

the entity identification model comprises an encoding module and a decoding module, wherein the encoding module is used for determining a feature vector of a target character in the multimedia data, the feature vector of the target character is used for representing character features of the target character when the target character is in the multimedia data, the corresponding feature vectors of the target character under different multimedia data are different, and the decoding module is used for carrying out entity identification on the target character according to the feature vector of the target character.

On the other hand, the embodiment of the present application discloses a data processing apparatus, the apparatus includes:

the acquiring unit is used for acquiring multimedia data associated with the target object identifier;

the recognition unit is used for calling an entity recognition model to perform entity recognition on the multimedia data and determining entity words in the multimedia data and entity word types of the entity words according to recognition results of the entity recognition model;

the processing unit is used for carrying out object data processing on the target object identification according to the entity words and the entity word types;

Correspondingly, the embodiment of the application also discloses an intelligent device, which comprises a processor, a memory and a network interface, wherein the processor, the memory and the network interface are connected with each other, the memory is used for storing a computer program, the computer program comprises program instructions, and the processor is configured to call the program instructions and execute the method.

Accordingly, the present application also discloses a computer readable storage medium storing a computer program comprising program instructions which, when executed by a processor, cause the processor to perform the above-mentioned method.

Accordingly, the embodiment of the present application also discloses a computer program product or a computer program, where the computer program product or the computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium, a processor of the smart device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to implement the method described above.

In the embodiment of the application, multimedia data associated with the target object identifier can be acquired, the entity recognition model is called to perform entity recognition on the multimedia data, so that the entity words and the entity word types of the entity words in the multimedia data are determined according to the recognition result of the entity recognition model, and further, the target object identifier can be subjected to object data processing according to the entity words and the entity word types. By implementing the method, the fast recognition of the entity words can be realized through the mode of model recognition, so that the recognition of the entity words is more comprehensive, and the recognition accuracy of the entity words can be improved.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present application, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

Fig. 1a is a schematic flowchart of a data processing method according to an embodiment of the present application;

FIG. 1b is a block diagram of a data processing system according to an embodiment of the present application;

fig. 2 is a schematic flowchart of a data processing method according to an embodiment of the present application;

FIG. 3a is a schematic flow chart illustrating identification using an entity identification model according to an embodiment of the present application;

FIG. 3b is a schematic interface diagram of a search interface provided by an embodiment of the present application;

FIG. 3c is a schematic interface diagram of a search result browsing interface provided by an embodiment of the present application;

fig. 4 is a schematic flowchart of a data processing method according to an embodiment of the present application;

FIG. 5a is a schematic flow chart illustrating classification using an entity classification model according to an embodiment of the present application;

FIG. 5b is a schematic flowchart of entity classification model training according to an embodiment of the present disclosure;

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

fig. 7 is a schematic structural diagram of an intelligent device according to an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are some, but not all, embodiments of the present application. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.

Artificial Intelligence (AI) is a theory, method, technique and application system that uses a digital computer or a machine controlled by a digital computer to simulate, extend and expand human Intelligence, perceive the environment, acquire knowledge and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the realization method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.

The artificial intelligence technology is a comprehensive subject and relates to the field of extensive technology, namely the technology of a hardware level and the technology of a software level. The artificial intelligence infrastructure generally includes technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and the like.

Machine Learning (ML) is a multi-domain cross discipline, and relates to a plurality of disciplines such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach for computers to have intelligence, and is applied to all fields of artificial intelligence. Machine learning/deep learning generally includes techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and formal education learning.

Natural Language Processing (NLP) is an important direction in the fields of computer science and artificial intelligence. It studies various theories and methods that enable efficient communication between humans and computers using natural language. Natural language processing techniques typically include text processing, semantic understanding, machine translation, robotic question and answer, knowledge mapping, and the like.

The scheme provided by the embodiment of the application relates to the technologies of machine learning, natural language processing and the like mentioned in the artificial intelligence technology, and is specifically explained by the following embodiment.

Named Entity Recognition (NER) refers to the Recognition of special objects from text, the semantic categories of which can be generally predefined before Recognition, and the predefined categories can be names of people, places, organizations, etc. Named entity recognition can be an independent information extraction task, and can also be applied to natural language processing technology, for example, the named entity recognition can be applied to the fields of information retrieval, automatic text summarization, syntactic analysis, question-answering system, machine translation, knowledge base (knowledge graph) building and the like.

The named entity can be understood as a word describing the entity in the text, and for example, a name of a person, a name of a place, a name of an organization, a name of a product, and the like can be called the named entity. Named entities tend to be the most interesting words in sentences. The definition of the named entity can be different from person to person and from scene to scene, for example, in the E-commerce field, the name of a commodity can be defined as an entity, and in the medical field, the name of a disease can be defined as an entity.

It will be appreciated that difficulties with entity identification tasks may generally include the following three points:

one is that the number of entities is infinite. The entity is an open set, for example, for a strongly structured entity, for example, the entity type of "1929-11-10" is time, the entity type of "good neighbor company" is organization name, and entity identification can be performed by means of template matching. However, for a weakly structured entity, if "all good" is hard to recognize by means of template matching or dictionary, because "all good" can be a work or other types, for example, "all good" can be the name of a restaurant, and the type of entity "all good" can be the name of an organization.

Secondly, the word boundaries of the entity are flexible. For example, in the case of "see again the san zheng when eating", the sentence may be recognized as "san zheng" or "san zheng" as a name of a person.

Thirdly, the entity types are easy to be confused. For example, taking the example of "tv drama is tall and getting on fire and the restaurant is tall and tall" two "in the sentence express two different types of entities, the first" tall and tall "type is the work and the second" tall and tall "type is the organization name. In a scenario like a food recommendation, it is more important to identify the second "all good" as the organization name, and the type of "all good" may also be a work, then in this case, it is likely that the entity types are confused.

In the embodiment of the application, an entity recognition model is specifically adopted to realize the analysis and recognition of entity words in a sentence or a paragraph, and the entity recognition task can be regarded as a sequence labeling task and the entity recognition is carried out through an end-to-end model. The entity recognition model can be generally divided into an input layer, an encoding layer and a decoding layer. As shown in fig. 1a, which is a schematic flow chart of a data processing method provided by the present application, it can be seen from fig. 1a that an input layer of a model can be input in a word granularity manner, or a word mixing manner, and the effect of the word mixing manner is better than that of the word granularity manner and the word granularity manner. The coding layer of the model may generally adopt a Recurrent Neural Network (RNN), a Convolutional Neural Network (CNN), or a transform Network. The decoding layer may typically employ Conditional Random Fields (CRF) and SoftMax.

In one implementation, when entity recognition is performed by using the model in the above description, the model needs to be trained first. In one implementation, the corpus needs to be obtained and labeled before the model is trained. The corpus includes a plurality of sentences, each sentence may include entity words, and the corpus may be obtained from the whole domain or from a specific domain. The whole domain may be a domain including various domains such as a medical domain, a sports domain, a food domain, and the like, and the specific domain may be a domain designated according to a demand, for example, if a name of a person in sports is to be recognized, the specific domain may be a sports domain. Labeling means to determine the entity words included in each sentence in the corpus and the entity word types corresponding to the entity words. The labeling mode can be a manual labeling mode or a remote supervision mode.

In an implementation manner, the present application provides a data processing method, which may obtain multimedia data associated with a target object identifier, and invoke an entity recognition model to perform entity recognition on the multimedia data, so as to determine entity words included in the multimedia data and entity word types of the entity words according to a recognition result of the entity recognition model, and further may perform object data processing on the target object identifier according to the entity words and the entity word types. It should be noted that the objects mentioned in the embodiments of the present application mainly refer to some users who can browse various information, send and receive text messages, and the like, while the multimedia data mentioned in the embodiments of the present application mainly refer to some text data, such as a text type of news, characters sent and received between users, and the like, and in some embodiments, the multimedia data may also be various audio and video data, image data, and the like, and after extracting corresponding text contents from audio data, video data, and image data, the processing procedures mentioned in the embodiments of the present application below can be executed.

In one implementation, the entity recognition model may include an encoding module and a decoding module, where the encoding module may be configured to determine a feature vector of a target character in multimedia data, where the feature vector of the target character is used to represent a character feature of the target character when the target character is in the multimedia data, and the feature vectors corresponding to the target character under different multimedia data are different. The decoding module may be configured to perform entity recognition on the target character according to the feature vector of the target character. By the method and the device, entity identification with finer granularity can be realized. For example, current entity identification methods can generally address coarse-grained entity identification, common entity types such as person names, place names, and organization names.

If the current entity identification method is applied to scenes such as news recommendation and the like, the entity type of the entity can be generally only roughly identified, and the entity type needs to be more fine-grained to characterize the entity in the scenes such as news recommendation and the like, so that more specific semantic information needs to be provided to enhance the indicativity of the entity type. For example, in recognizing a person's name, the person's name may be further classified into a sports person name, an entertaining person name, and the like. Fine-grained entity recognition may define the type of an entity by the scenario that needs to be resolved, and may specify the type of entity that needs to be recognized. For example, in the application, for a scene such as news recommendation, multiple entity types can be redefined to realize fine-grained entity identification. For example, the plurality of entity types may include: name of person, place name, organization name, works, gourmet, commodity, creature, time, event, medical treatment. Other entity types may also be included, and are not limited in this application.

In one embodiment, the fine-grained entity identification has a wide application range, and taking a scene of a portal as an example, the fine-grained entity identification can be applied to a plurality of aspects such as searching, recommending, advertising and the like. For example, as shown in fig. 1b, after an object browses an article about a restaurant in "good standing" state by using a terminal, a server may detect a browsing operation of the object and acquire the article, so as to identify entity words contained in the article and entity word types corresponding to the entity words by using the data processing method in the present application. After the recognition result (the recognition result is the entity word and the entity word type corresponding to the entity word) is obtained, the method can be applied to various actual scenes according to the recognition result, for example, the actual scenes can be user images, advertisement delivery, search, interest point generalization, peripheral recommendation and the like.

Specifically, with the present application, entity words can be identified from the article as shown in FIG. 1 b: the 'good' and the entity word type corresponding to the entity word belong to the works and the organization names respectively. Meanwhile, the entity word with the place name as the entity word type in the article can be identified: hangzhou and New Youyuan hotels, and the entity word type is the entity word of the food: young pigeons. After the above-mentioned entity words and entity word types are recognized, in the user portrait scene, it can be determined which entity words in the article can be used as the user portrait of the object through the entity word types, so as to express the real interest of the object, as shown in fig. 1 b. For example, if the entity word type required for the user to portrait is works and gourmet, the user can portrait both the young pigeon and the good pigeon. In the advertisement putting scene, the entity word type is the entity word of the organization name: are good and the type of the entity word is the entity word of the food: young pigeons are all good advertisement triggering words. In a search scenario, the recognition result can better help to realize the purpose recognition and the slot extraction of the search so as to optimize the search result. The generalization of the interest points can also be realized, for example, after the type of the entity word "young pigeon" is judged to be the food, the specific entity word is generalized through the interest points by using a related algorithm, for example, the "young pigeon" can be generalized into the classic cantonese, so that the interest points of the object can be better depicted, and the diversity of recommendation can be improved. Peripheral recommendations may also be implemented, for example, based on an entity word whose type is place name: hangzhou and new youth hotels can recommend local related contents of Hangzhou to objects, for example, can recommend national wetland parks of scenic spots xi river near the new youth hotels.

The server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as cloud service, a cloud database, cloud computing, a cloud function, cloud storage, network service, cloud communication, middleware service, domain name service, security service, CDN, big data and artificial intelligence platform. The terminal can be a game terminal in intelligent equipment such as a Mobile phone, a tablet computer, a notebook computer, a palm computer, Mobile Internet Device (MID, Mobile Internet Device) and the like. The terminal and the server may be directly or indirectly connected through wired or wireless communication, and the application is not limited herein.

The implementation details of the technical solution of the embodiment of the present application are set forth in detail below:

referring to fig. 2, fig. 2 is a schematic flowchart of a data processing method provided in an embodiment of the present application, where the method of the embodiment of the present application may be applied to a smart device, for example, the smart device may be a smart terminal, such as a smart phone, a tablet computer, a smart wearable device, and the like, and the smart device may also be a server. The data processing method of the embodiment of the application comprises the following steps.

S201: and acquiring multimedia data associated with the target object identifier. The target object identifier may be an identifier corresponding to a target object, and the target object may be an object for performing web browsing or web searching by using a terminal, or an object for performing social contact by using social contact software. The multimedia data associated with the target object identifier may be a sentence browsed by the target object on a web page, or a search sentence input by the target object in a search engine, or a dialog sentence of the target object in a chat session using social software, or a reply sentence of the target object in a social circle such as a friend circle for a browsed message, etc.

S202: and calling the entity recognition model to perform entity recognition on the multimedia data, and determining entity words in the multimedia data and entity word types of the entity words according to recognition results of the entity recognition model. The entity recognition model may include an encoding module and a decoding module. In the entity identification task, the encoding module of the entity identification model can be RNN or CNN or Transformer, and the decoding module can be CRF or SoftMax. Regarding the coding module, it is considered that the RNN-based coding module has a weak ability to acquire semantic information, the CNN-based coding module may easily lose long-distance information, or the Transformer-based coding module has a weak ability to learn the boundaries of words.

In order to improve the model effect of the model, the pre-training model with strong semantic representation can be used as the coding module, the coding module can well learn the context information of each character in the input sentence, and the coding module can be BERT, ERNIE, LICHEE or the like. For the decoding module, CRF or SoftMax may be used for decoding, and considering that a structure with markov property such as CRF needs to be decoded according to a time step and decoding parallelization cannot be realized, the inference speed of the model may be affected, that is, the decoding efficiency of the entity identification model may be reduced. And directly classifying the output using SoftMax may result in the model missing the inter-tag constraints when decoding. Then, in the present application, the decoding module may further implement decoding in a pointer labeling (MS) manner, the MS decoding may improve a problem that CRF decoding cannot implement parallelism, and the MS decoding may further identify an entity word type corresponding to an entity word according to context information of each character included in the multimedia data, so that a problem of missing inter-tag constraint in SoftMax decoding may be solved. The MS decoding may also use the MRC framework to split the identification of the entity word type corresponding to the entity word into two steps, i.e., a first word (which may be understood as start) in the entity word and a last word (which may be understood as end) in the entity word, so as to determine the entity word according to the identified first word and last word, and determine the entity word type corresponding to the entity word.

For example, table 1 shows the decoding speed and the model effect when the encoding modules are the same and the decoding modules are different in the entity recognition model. The model effect can be expressed by using a parameter F1 value, the F1 value is a harmonic mean of the precision rate and the recall rate, the value range of the F1 value is 0 to 1, and the larger the value is, the better the model effect is. As can be seen from table 1, in the case where the encoding modules are the same, the model effect using the MS-based decoding module is close to that of the CRF-based decoding module, but the decoding speed of the MS-based decoding module is 2.5 times that of the CRF-based decoding module.

Table 1:

model structure	Decoding speed (ms)	F1
			LICHEE+CRF	32.5	0.8064
LICHEE+SoftMax	12.5	0.7938
			LICHEE+MS	12.9	0.8026

In one implementation, the encoding module may be configured to determine a feature vector of a target character in the multimedia data, and the feature vector of the target character may be used to represent character features of the target character when the target character is in the multimedia data. For example, the target character may refer to a first word in the physical words and a last word in the physical words, and the character characteristics may include context characteristics of the character in the multimedia data. It should be noted that the feature vectors corresponding to the target character under different multimedia data are different, for example, the feature vector of the character "text" in the sentence "a author published an article" is different from the feature vector of the character "text" in the sentence "article a introduced the development history of tennis". The decoding module may be configured to perform entity recognition on the target character according to the feature vector of the target character to determine whether the target character is an entity word and an entity word type corresponding to the entity word.

In one implementation, the input of the encoding module may be characters in the multimedia data arranged according to the language order, and the output of the encoding module may be an L-dimensional feature vector corresponding to each character, where L is an integer greater than 1. For example, the feature vector may be a 768-dimensional vector, and the feature vector may include context features corresponding to characters. The input of the decoding module may be the L-dimensional feature vector corresponding to each of the characters, and the output of the decoding module may be a probability that the character corresponding to the L-dimensional feature vector belongs to a first word in the entity words or a probability that the character corresponding to the L-dimensional feature vector belongs to a last word in the entity words.

For example, fig. 3a illustrates the recognition process of the entity recognition model, which is described by taking multimedia data as a sentence "singing in autumn and cloud in the evening". As shown in fig. 3a, the input of the encoding module is each character in "autumn cloud contributes late", and the output of the encoding module is a feature vector corresponding to each character in "autumn cloud contributes late". Then, the feature vector corresponding to each character may be used as an input of the decoding module, so that the decoding module outputs a probability that the character corresponding to the feature vector belongs to a first word in the entity words or outputs a probability that the character corresponding to the feature vector belongs to a last word in the entity words according to the feature vector corresponding to each character. Further, the entity words included in the sentence can be determined according to the probability of the first word and the probability of the last word. If a certain character is the first word or the last word belonging to the entity words, the probability corresponding to the character is not 0, and if a certain character is the first word or the last word not belonging to the entity words, the probability corresponding to the character is 0, for example, t1 and t2 in fig. 3a are not 0, and t1 to t6 are 0. Where the identity corresponding to the first word and the identity of the last word may be present in the output of the decoding module. Then the first word and the last word may be determined based on the corresponding identifier of the first word and the identifier of the last word. After the first word and the last word are determined, the entity words included in the training sentence may be determined. The first word, the last word, and words included between the first word and the last word may be determined as one entity word. For example, "autumn" may be recognized as the first word, "cloud" may be recognized as the last word, and then the physical word is "autumn cloud". As another example, if the sentence is "a new world hotel," then "big" may be identified as the first word and "shop" as the last word, i.e., the entity word is "a world hotel. The output of the decoding module may further include an entity word type corresponding to each character in the sentence, and then, after the entity word "autumn cloud" included in the sentence is determined, the entity word type corresponding to the entity word may be further determined according to the entity word type corresponding to each character included in the entity word "autumn cloud". For example, if the entity word types of the character "autumn" and the character "cloud" are both names of people, the entity word type corresponding to the entity word "autumn cloud" is a name of people.

S203: and carrying out object data processing on the target object identification according to the entity words and the entity word types. In one implementation, the object data processing may be performed on the target object identifier according to the entity word and the entity word type. For example, the data processing may be updating a user image of the object, recommending information for the object, generalizing an interest feature of the object, and the like. For example, in the case of a data website, after an object browses an article on the data website, user image update, interest feature generalization, and information recommendation can be performed according to the entity words contained in the article and the entity word types corresponding to the entity words.

In one implementation, information recommendation can be performed on an object corresponding to the target object identifier according to the entity word and the entity word type. Specifically, recommendation feature analysis can be performed on the object corresponding to the target object identifier according to the entity words and the entity word types. The recommendation feature analysis can determine the information recommendation type to be recommended by the target object identification based on the entity word type. For example, assuming that the entity word type is a food, it may be determined that the information recommendation type corresponding to the target object identifier is a food type. After the recommendation characteristic analysis is performed, a corresponding recommendation characteristic analysis result can be obtained, so that information recommendation can be performed on the object corresponding to the target object identification according to the recommendation characteristic analysis result. For example, as can be seen from the above description, the recommended feature analysis result may be an information recommendation type, and the information recommendation type may specifically be a food type, and then, information of the recommended food type may be identified for the target object.

For example, assuming that the multimedia data is an article about "good all" restaurant shown in fig. 1b, the entity words and the entity word types corresponding to the entity words in the article may be described as follows. The entity word: the entity words are all well-shaped, and the entity word types corresponding to the entity words respectively belong to works and organization names; the entity word: in Hangzhou and New Youyuan hotels, the type of the entity word corresponding to the entity word is a place name; the entity word: the type of the entity word corresponding to the entity word is the delicious food. Then, after determining the entity words and the entity word types in the article, the target object may be determined to identify the information recommendation types to be recommended based on the entity word types. For example, if the entity word type in the above description has a food, an article recommending the food type may be identified to the target object. In one implementation, the information recommendation may specifically be a peripheral recommendation, for example, based on the entity word whose type is a place name: the Hangzhou and the new-pioneer hotel can recommend the peripheral information to the target object identification. Specifically, recommendation feature analysis can be performed according to the entity words and the entity word types, and it can be further determined that the information recommendation type to be recommended by the target object identifier can be a hotel or a food or a tourist attraction near the newly-opened hotel. In one implementation, the information recommendation may specifically be an advertisement delivery, such as for a physical word with a physical word type of organization name: are good and the type of the entity word is the entity word of the food: if the young pigeons are good advertisement trigger words, recommendation characteristic analysis is performed according to the entity words and the entity word types, and then the type of the advertisement to be recommended by the target object identification can be determined to be similar to a restaurant with the young pigeons.

In one implementation, a user representation of an object corresponding to the target object identification may be updated based on the physical words and the physical word types. For example, it may be determined by the entity word type which entity words may determine the user portrait characteristics that identify the corresponding object for the target object. Taking the article shown in fig. 1b as an example, if the user portrait feature types required by the user portrait include works and gourmets, the user portrait features of the target object corresponding to the target object identification can be determined as the stiff and squab, so as to update the user portrait of the target object corresponding to the target object identification. And after the user portrait corresponding to the target object identification is updated, if the user portrait is continuously applied to an information recommendation scene, information recommendation can be performed on the object corresponding to the target object identification according to the updated user portrait. In an implementation manner, the user portrait characteristics of the object corresponding to the target object identifier can be analyzed according to the entity words and the entity word types, so as to obtain a user portrait characteristic analysis result. For example, a user representation feature type in the user representation identified by the target correspondence may be determined based on the entity word type, such as the entity word type may be determined as the user representation feature type. And after the user portrait feature analysis result is obtained, the user portrait of the user corresponding to the target user identification can be updated according to the user portrait feature analysis result. For example, if the user portrait feature analysis result may indicate that the user portrait feature type in the user portrait includes a food, an entity word with the entity word type of food may be added to the user portrait. Or the user portrait of the object corresponding to the target object identification can be continuously updated according to the user portrait characteristic type determined by the entity word type. For example, after the object corresponding to the target object identifier browses any article, the user portrait may be updated according to the entity words and the entity word types included in the article.

In one implementation, the object corresponding to the target object identifier may be generalized in terms of interest characteristics according to the entity words and the entity word types. Specifically, interest feature generalization processing can be performed on the target object identifier according to the entity words and the entity word types to obtain generalized object features corresponding to the target object identifier. For example, taking the article shown in fig. 1b as an example, after determining that the type of the entity word "squab" in the article is a food, it may be determined that the interest feature corresponding to the target object identifier is a squab in the food, so as to better depict the interest point of the target object identifier, thereby improving the diversity of recommendations. The interest feature may be subjected to an interest feature generalization process, for example, the entity word "young pigeon" may be subjected to an interest feature generalization process using a correlation generalization algorithm to generalize the "young pigeon" into a classic cantonese. Then the generalized object feature corresponding to the target object identification is classic yue dish. After the generalized object features corresponding to the target object identifier are obtained, if the generalized object features are continuously applied to the information recommendation scene, information corresponding to the generalized object features can be recommended to the target object identifier in a targeted manner, for example, through the foregoing description, information related to classic yue dish can be recommended to the target object identifier.

In one implementation, the method can also be specifically applied to search scenes. Specifically, when a search event initiated based on multimedia data is detected, the entity words and the entity word types can be used as search keywords to perform search processing according to the entity words and the entity word types of the entity words included in the multimedia data, and search results obtained through the search processing are displayed on a search result browsing interface. For example, when the target object identifies that the corresponding object is inputting multimedia data in the search interface, it may be determined that a search event initiated based on the multimedia data is detected.

For example, after the object corresponding to the target object identifier inputs multimedia data on the search interface, the terminal may obtain the multimedia data and may send the multimedia data to the server, so that the server determines entity words and entity word types included in the multimedia data, and performs search processing on the database using the entity words and the entity word types as search keywords. For example, as shown in fig. 3b, the object may input data to be searched in a search area marked by 301 in fig. 3b, and assuming that the multimedia data input by the object is what is played in kyakagou, after the server acquires the multimedia data, the server may recognize that the entity word in the multimedia data is kyakagou and the entity word type is a place name. After the entity words and the entity word types are determined, the entity words (jizhai gou) and the entity word types (place names) can be searched in the database as search keywords. For example, the database may be searched for weather about kyani gutter, travel of kyani gutter, and the like. After the server obtains the search result, the server may return the search result to the terminal, and the terminal may jump from the search interface of fig. 3b to a search result browsing interface as shown in fig. 3c, so that the search result returned by the server may be displayed on the search result browsing interface. The search result may be specifically displayed in the search result area marked by 302 in fig. 3c, for example, whether or not the search result area displays brick-and-mortar weather or brick-and-mortar tourism is the search result.

In the embodiment of the application, multimedia data associated with the target object identifier can be acquired, the entity recognition model is called to perform entity recognition on the multimedia data, so that the entity words and the entity word types of the entity words in the multimedia data are determined according to the recognition result of the entity recognition model, and further, the target object identifier can be subjected to object data processing according to the entity words and the entity word types. By implementing the method, the fine-grained entity words can be recognized, and the recognition accuracy is improved. And corresponding data processing can be carried out in various application scenes according to the recognition result, and the data processing efficiency and accuracy can also be improved.

Referring to fig. 4, fig. 4 is a flowchart illustrating a data processing method according to an embodiment of the present disclosure, where the method according to the embodiment of the present disclosure may be applied to a smart device, for example, the smart device may be a smart terminal, such as a smart phone, a tablet computer, a smart wearable device, and the like, and the smart device may also be a server. The data processing method of the embodiment of the application comprises the following steps.

S401: and acquiring multimedia data associated with the target object identifier.

S402: and training through a training sample set to obtain an entity recognition model.

In one implementation, the entity recognition model may be trained by a training sample set, where the training sample set may include a plurality of training sample pairs, and the training sample pairs may include a training sentence and label information corresponding to the training sentence. The labeling information of the training sentence may include: the entity words included in the training sentences and the entity word types corresponding to the entity words. It should be noted that the entity recognition model requires a high-quality training sample set, and if the labeling information has a wrong label or a label missing condition, the recognition effect of the entity recognition model is affected. For example, labeling "phoenix tree" in "very wonderful phoenix tree written by author a" as tree name, the entity word type of the phoenix tree should be a work, which is a wrong labeling. For another example, the "all good" in "all good of last year's fire is not labeled, and in this case, the label is omitted. In the fine-grained entity recognition, as the entity word types corresponding to the recognized entity words become more, the difficulty in labeling the entity words and the entity word types is increased. In an implementation manner, manual tagging can be used for tagging, and it is considered that tagging entity words and entity word types in sentences by using a manual tagging method is quite troublesome, time-consuming and labor-consuming, and error is easily caused in tagging, so that the work efficiency cannot be improved. Then, it can be considered to prepare a training sample set corresponding to the entity recognition model by automatically constructing the training sample. The application may also use the idea of remote supervision to label, for example, a pre-constructed dictionary may be used to label the collected sentences to obtain the entity words and entity word types in the sentences. The specific labeling manner may be understood as performing literal matching by using words and sentences included in the dictionary, for example, if there are three names of people included in the name dictionary, the three names of people in the sentence may be labeled as entity words, and the type of the entity words is name of people.

In one implementation, the specific implementation process of obtaining the entity recognition model through training of the training sample set may be as follows. The structure of the entity recognition model can be as shown in fig. 3 a. For any training sentence in the training sample set, the training sentence may be input to an encoding module in the entity recognition model to obtain a feature vector corresponding to each character in the training sentence, where the feature vector may be a 768-dimensional vector, and the feature vector may include context features corresponding to the characters. After obtaining the feature vector corresponding to each character in the training sentence, the feature vector corresponding to each character in the training sentence may be input to a decoding module in the entity recognition model to determine a probability that each character in the training sentence belongs to a first word in the entity words, or determine a probability that each character in the training sentence belongs to a last word in the entity words, and also determine an entity word type corresponding to each character in the training sentence.

Further, the training entity words included in the training sentence may be determined according to a probability that each character belongs to a first word in the entity words or a probability that each character belongs to a last word in the entity words. For the method for determining the training entity words, reference may be made to the description of the method for determining the entity words, which is not described herein again. For example, taking the training sentence "autumn clouds sing late" as an example, "autumn" can be recognized as the first word, and "cloud" can be recognized as the last word, then the training entity word is "autumn clouds". For another example, taking the training sentence "world hotel in new business" as an example, "big" can be recognized as the first word, and "shop" is recognized as the last word, then the training entity word is "world hotel".

After the training entity words contained in the training sentences are determined, the entity word prediction types corresponding to the training entity words can be determined according to the entity word types corresponding to each character contained in the training entity words. For example, the entity word prediction type corresponding to "autumn cloud" described above is a name of a person, and for another example, the entity word prediction type corresponding to "world hotel" is a name of a place. After the entity word prediction type corresponding to the training entity word is determined, the entity recognition model can be trained according to the entity word in the labeling information of the training sentence, the entity word type corresponding to the entity word, the training entity word and the entity word prediction type corresponding to the training entity word, so that the trained entity recognition model is obtained.

The entity recognition model is trained according to the entity words in the labeling information of the training sentence, the entity word types corresponding to the entity words, the training entity words and the entity word prediction types corresponding to the training entity words, and the specific implementation steps of the trained entity recognition model can be described as follows: for convenience of description, the entity words and the entity word types corresponding to the entity words may be referred to as first information, the training entity words and the entity word prediction types corresponding to the training entity words may be referred to as second information, and the gradient of the loss function may be calculated according to the first information and the second information, where the loss function is not limited in the present application. And then, updating parameters of the model parameters of the entity recognition model according to the gradient of the loss function, detecting whether the loss function meets a preset convergence condition, and stopping updating the parameters of the model parameters when the loss function is detected to meet the preset convergence condition, so that the trained entity recognition model can be obtained. The preset convergence condition may be that the gradient of the loss function is smaller than a preset threshold, or that the weight change between two iterations is already small and smaller than a preset threshold, or that the iteration number of the model reaches a preset maximum iteration number, and when any one of the above conditions is met, the training of the entity recognition model may be stopped.

Specifically, the entity words included in the training sentence may be found and determined from the training sentence through the acquired reference dictionary set, and the entity word type corresponding to the entity words included in the training sentence may be determined according to the type of the reference dictionary used when determining the entity words, where the type of the reference dictionary is related to the aforementioned entity word type, for example, the type of the reference dictionary and the entity word type both include types of a person name, a place name, a mechanism name, a work name, and the like, and for example, the words in the reference dictionary of the person name type are all entity words of a person name. The reference dictionary set may include N types of reference dictionaries, and the types of words included in the reference dictionaries of the same type are all the same. Where N is an integer greater than 1, and the value of N may be set as required, for example, N may be 7, or 10, etc. For example, for an information website, fine-grained entity recognition of 10 entity types of words may be defined, where the 10 entity types of words may include names of people, places, organizations, works, gourmets, commodities, creatures, time, events, medical treatments, and other types, which is not limited in this application. N may be 10, and then 10 types of reference dictionaries may be preset. The words in the reference dictionary can be collected and sorted through the knowledge graph, and the words in each reference dictionary are of an entity word type, for example, the words in the reference dictionary based on the names of people are all names of people, and the words in the reference dictionary based on the names of places are all names of places.

In one implementation, taking the example that the training sentence "five rings are wanted to do before five years and now is implemented by two rings", the entity words in the training sentence can be determined to include "five rings" and "two rings" by referring to the dictionary set (the "five rings" and "two rings" are both the type of road name and the type of organization name). Since "five rings" and "two rings" may exist in the reference dictionary of road names, and also in the reference dictionary of agency names. If the road name is searched by a reference dictionary of the road names, the entity word types of the five rings and the two rings are determined as the road names. If the entity word types of the five rings and the two rings are searched by the institution name reference dictionary, the entity word types of the five rings and the two rings can be determined as the institution names. Alternatively, if the determination of the entity word type is performed with the frequency of occurrence, since the frequency of occurrence of the "five-ring" and the "two-ring" in the reference dictionary of the road name is higher than the frequency of occurrence in the reference dictionary of the organization name, the entity word type of the "five-ring" and the "two-ring" may be determined as the road name, but the entity word type of the "five-ring" and the "two-ring" in the training sentence should be one organization name. It will be appreciated that in the above case, there may be instances of mislabeling. To solve the above problem, the training sentences may be sentence-classified first. For example, the training sentence "five rings are wanted to do before five years, but now the training sentence is realized by two rings" from the finance article, and in the finance article, the probability that the "five rings" and the "two rings" are named entity words of the organization is obviously higher than that the "five rings" and the "two rings" are named entity words of the road. Then, in determining the entity word types corresponding to the "five rings" and the "two rings", the organization name is selected instead of the road name.

As can be seen from the foregoing, determining the entity word type corresponding to the entity word included in the training sentence by using the reference dictionary often leads to a problem of type ambiguity. For example, the training sentence "five rings and five years ago, which is supposed to be done, is now implemented by two rings," because the frequency of occurrence of "five rings" and "two rings" in the reference dictionary of road names is higher than that in the reference dictionary of organization names, and when the entity word type of the entity word is determined, the entity word type of "five rings" and "two rings" may be determined as the road name, and the problem of wrong labeling may occur. In order to solve the problem of entity ambiguity, the type of the reference dictionary and the classification type of the training sentence can be used for jointly determining the entity word type of the entity word so as to enhance the accuracy of the labeling information. Specifically, when M reference dictionaries are used for searching and determining a target entity word from a training sentence, where M is an integer greater than 1, that is, when an entity word is determined and the reference dictionary corresponding to the entity word has multiple types (for example, a target entity word is determined as a five-ring type, but the target entity word is a road name type and a mechanism name type), the entity word type corresponding to the target entity word is determined according to the types of the M reference dictionaries and the classification type of the training sentence. The classification type corresponding to the training sentence can be determined according to the information type of the information article. For example, the classification type of the training sentence can be determined as the financial type according to the information type to which the information article belongs, and then the entity type with the target entity word being "five rings" should be an organization name.

For example, the training sentences may be obtained from information articles published on the target web site, and the information articles published on the target web site may be of different information types. The target website may be a portal website of some internet or other information websites, which is not limited in this application. The portal website has rich information data, which can relate to 14 information types such as movie and television ensembles, sports, digital codes and the like. Then, the training sentence can be obtained from the information article disclosed by the web portal, and when the training sentence is obtained, the information type of the information article corresponding to the training sentence can also be obtained. For example, a training sentence is acquired from a finance article, namely a 'five-ring five-year-ago wanted thing, which is realized by two rings at present', on one hand, the entity word types of the entity words 'five rings' and 'two rings' can be determined as the mechanism name through a reference dictionary of the mechanism name, and the entity word types of the entity words 'five rings' and 'two rings' can be determined as the road name through the reference dictionary of the road name; on the other hand, since the training sentence is from a finance-type article, the classification type corresponding to the training sentence may be determined as a finance type. It is to be understood that the reference dictionary related to the types of finance and economics is a reference dictionary for organization names, and then the entity word types of the entity words "five rings" and "two rings" may be determined as organization names.

In one implementation, when the classification type of the training sentence and the type of the reference dictionary are used together to determine the entity word type of the entity word, so as to enhance the accuracy of the entity word type of the entity word, there may be a problem of entity word ambiguity. For example, the training sentence "lie four next fir trees" is exemplified, wherein "fir tree" is a type of both a person name and a plant. The training sentence is obtained from the entertainment article, the classification type of the training sentence is the entertainment type, the reference dictionary related to the entertainment type can be a name reference dictionary, and when the name reference dictionary is used for searching the entity words in the training sentence, the 'fir tree' in the training sentence can be generally determined as the entity words, and the entity word type corresponding to the 'fir tree' is determined as the name of the person, and actually the 'fir tree' belongs to the plant. In order to solve the problem of the ambiguity of the entity words, it may be considered that the entity words and the types of the entity words are further judged according to the context information of the entity words to determine whether the entity words and the types of the entity words determined by the classification type of the training sentence and the type to which the reference dictionary belongs are correct.

In one implementation, context information of the entity words may be obtained to determine whether the initial entity words in the training sentence and the initial entity word types of the initial entity words are correct according to the context information, and if not, the initial entity words in the training sentence and the initial entity word types of the initial entity words may be adjusted to obtain the entity words in the training sentence and the entity word types of the entity words. The initial entity words and the initial entity word types of the initial entity words can be adjusted by using an entity classification model. For example, the entity classification model may be an entity classification model based on a BERT model, or may be other entity classification models that can obtain context information, which is not limited in this application.

The embodiment of adjusting the initial entity words in the initial training sample set and the initial entity word types of the initial entity words may be specifically described as follows. A plurality of candidate entity word types corresponding to the initial entity word and a target probability that the initial entity word belongs to each candidate entity word type may be determined by the entity classification model. Thus, the initial entity word type of the initial entity word can be adjusted to a target candidate entity word type among the multiple candidate entity word types, so that the target candidate entity word type among the multiple candidate entity word types is used as the entity word type of the initial entity word, wherein the target probability value of the target candidate entity word type is the maximum. When the entity word type of the initial entity word is the reference entity word type, the initial entity word may be used as the entity word, wherein the reference entity word type may be understood as the type to which the reference dictionary belongs.

For example, as shown in fig. 5a, the training sentence "tv series in his collaboration with fir tree and zhang lan" may be input into the entity classification model, from which a plurality of candidate entity word types of the initial entity word "fir tree" and a target probability for each candidate entity word type may be determined. The target probability that the candidate entity word type corresponding to the fir tree is the name of the person is 0.87, and the target probability that the candidate entity word type corresponding to the fir tree is the plant is 0.04. Through the target probability, the candidate entity type of word corresponding to the target probability of 0.87 can be determined as the entity type of the fir tree, namely the entity type of the fir tree is the name of the person. Meanwhile, if the name is also one of the reference entity word types, the entity word type of the fir tree in the recognition result of the entity classification model is the name.

For example, as shown in fig. 5a, the training sentence "lie four next fir trees" is input into the entity classification model, and the target probability that the candidate entity word type corresponding to the initial entity word "fir tree" is the name of the person is determined to be 0.12 by the entity classification model, and the target probability that the candidate entity word type corresponding to the "fir tree" is the plant is determined to be 0.74. The candidate entity word type corresponding to the target probability of 0.74 may be determined as the entity word type of the "fir tree", that is, the entity word type of the "fir tree" is a plant. Meanwhile, if the plant is also one of the reference entity word types, the entity word type of the fir tree is the plant in the recognition result of the entity classification model. If the plant is not one of the reference entity word types, then there is no corresponding entity word type in the recognition result of the entity classification model for the "fir tree".

In one implementation, before the initial entity words in the initial training sample set and the initial entity word types of the initial entity words are adjusted by using the entity classification model, the entity classification model may be trained to perform the above adjustment by using the trained entity classification model. In one implementation, the entity classification model is trained from a sample set of entity words. The entity word sample set can comprise an entity word training sentence and label information, and the label information can comprise a target characteristic entity word and a target characteristic entity word type. The target characteristics may mean that the entity words are not ambiguous, that is, the target characteristics entity words may be understood as unambiguous entity words, and the target characteristics entity word type may be understood as unambiguous entity word types. The target characteristic entity words can be obtained by filtering ambiguous words from the reference dictionary described above. For example, for the name reference dictionary, "zhang san" and "lie si" in the name reference dictionary are target characteristic entity words, and the corresponding target characteristic entity word types are names of people. For another example, the "phoenix tree" in the person name reference dictionary is not the target characteristic entity word, and the "phoenix tree" can be understood as the person name and also as the plant.

In one implementation, as shown in fig. 5b, in order to train the entity classification model, an entity word sample set corresponding to the entity classification model needs to be determined first, where an entity word training sentence included in the entity word sample set may be determined according to a reference dictionary. Specifically, for example, the target characteristic entity word type is used as the name, ambiguous words in the name reference dictionary may be filtered first, that is, ambiguous words in the name reference dictionary are deleted. For example, the name reference dictionary includes phoenix tree, fir tree, lie four and zhang three, and the phoenix tree and fir tree can be filtered out when ambiguous word filtering is performed. In this application, the entity word sample set may be a training sample set obtained by screening the training sample set, that is, it is ensured that the entity word training sentences obtained after screening do not include ambiguous words. The entity words in the training sentences of the name class can be matched with the words in the filtered name reference dictionary, and if a certain entity word in the training sentences is not a word in the filtered name reference dictionary, the training sentences can be filtered. For example, if the physical word "three of" in the training sentence "three of three directors' 3 works all earns" is in the filtered name reference dictionary, the training sentence may be determined as the physical word training sentence. For another example, if the physical word "lie four" in the training sentence "lie four suit with broad leg" is in the filtered name reference dictionary, the training sentence can be determined as a physical word training sentence. As another example, the training sentence "the composition of phoenix," the education i understand "includes the entity word" phoenix, "and" phoenix "is not in the filtered name reference dictionary, the training sentence may be filtered. Then, after the entity word sample set is determined by the method, the entity word sample set can be used to train the entity classification model.

S403: and calling the entity recognition model to perform entity recognition on the multimedia data, and determining entity words in the multimedia data and entity word types of the entity words according to recognition results of the entity recognition model.

S404: and carrying out object data processing on the target object identification according to the entity words and the entity word types.

For specific implementation of steps S401, S403, and S404, reference may be made to the detailed description of steps S201, S202, and S203 in the foregoing embodiment, and details are not repeated here.

In the embodiment of the application, multimedia data associated with the target object identifier can be obtained, the entity recognition model is obtained through training of the training sample set, the entity recognition model is called to perform entity recognition on the multimedia data, so that entity words in the multimedia data and entity word types of the entity words can be determined according to recognition results of the entity recognition model, and further, object data processing can be performed on the target object identifier according to the entity words and the entity word types. The method and the device can realize automatic construction of the training sample set corresponding to the entity recognition model by referring to dictionary remote supervision, context classification of sentences and type classification of entity words. Compared with the manual labeling of the labeling information in the training sample set, the efficiency of labeling and the accuracy of labeling information can be improved. Meanwhile, the entity recognition model can be used for improving the decoding efficiency of the model. The accuracy and efficiency of entity recognition by adopting the entity recognition model can be improved.

Please refer to fig. 6, which is a schematic structural diagram of a data processing apparatus according to an embodiment of the present application. The apparatus of the embodiment of the application can be applied to a smart device, for example, the smart device can be a smart terminal, such as a smart phone, a tablet computer, a smart wearable device, and the like, and the smart device can also be a server.

The data processing apparatus described in this embodiment includes:

an obtaining unit 601, configured to obtain multimedia data associated with a target object identifier;

the identification unit 602 is configured to invoke an entity identification model to perform entity identification on the multimedia data, and determine an entity word in the multimedia data and an entity word type of the entity word according to an identification result of the entity identification model;

a processing unit 603, configured to perform object data processing on the target object identifier according to the entity word and the entity word type;

In an implementation manner, the processing unit 603 is specifically configured to:

updating the user portrait of the object corresponding to the target object identification according to the entity words and the entity word types;

according to the entity words and the entity word types, information recommendation is carried out on the objects corresponding to the target object identifications;

and according to the entity words and the entity word types, performing interest characteristic generalization on the objects corresponding to the target object identifications.

In one implementation, the processing unit 603 is further configured to:

when a search event initiated based on the multimedia data is detected, the entity words and the entity word types are used as search keywords for search processing according to the entity words and the entity word types;

and displaying the search result on the search result browsing interface.

In one implementation manner, the input of the encoding module is each character in the multimedia data arranged according to a language order, the output of the encoding module is an L-dimensional feature vector corresponding to each character, and L is an integer greater than 1;

the input of the decoding module is the L-dimensional feature vector corresponding to each character, and the output of the decoding module is the probability that the character corresponding to the L-dimensional feature vector belongs to the first word in the entity words or the probability that the character corresponding to the L-dimensional feature vector belongs to the last word in the entity words.

In one implementation, the entity recognition model is obtained by training a training sample set, where the training sample set includes a training sample pair, the training sample pair includes a training sentence and labeling information of the training sentence, and the labeling information of the training sentence includes: entity words included in the training sentences and entity word types corresponding to the entity words;

the entity words included in the training sentence are searched and determined from the training sentence through an acquired reference dictionary set, the reference dictionary set comprises N types of reference dictionaries, the types of the words in the same type of reference dictionary are the same, and N is an integer greater than 1;

the entity word type corresponding to the entity words included in the training sentence is determined according to the type of the reference dictionary used when the entity words are determined.

In one implementation, when M reference dictionaries are used for searching and determining a target entity word from a training sentence, an entity word type corresponding to the target entity word is determined according to types of the M reference dictionaries and a classification type of the training sentence, wherein M is an integer greater than 1;

the training sentences are obtained from the information articles disclosed by the target website, the information articles with different information types are disclosed on the target website, and the classification types of the training sentences are determined according to the information types to which the information articles belong.

In one implementation manner, the training sample set is obtained by adjusting initial entity words in an initial training sample set and initial entity word types of the initial entity words by calling an entity classification model;

adjusting the initial entity words in the initial training sample set and the initial entity word types of the initial entity words comprises:

determining a plurality of candidate entity word types corresponding to an initial entity word and a target probability that the initial entity word belongs to each candidate entity word type through the entity classification model;

and adjusting the initial entity word type of the initial entity word to be a target candidate entity word type in the candidate entity word types, wherein the target candidate entity word type in the candidate entity word types is used as the entity word type of the initial entity word, and the target probability of the target candidate entity word type is the maximum value.

In one implementation, the entity classification model is obtained by training an entity word sample set;

the entity word sample set comprises entity word training sentences and labeling information, and the labeling information comprises target characteristic entity words and target characteristic entity word types;

the target characteristic entity words are obtained by filtering ambiguous words of one or more reference dictionaries.

In one implementation, training through the training sample set to obtain the entity recognition model includes:

inputting the training sentence into the coding module aiming at any training sentence in the training sample set, and determining a feature vector corresponding to each character in the training sentence;

inputting the feature vector corresponding to each character in the training sentence into the decoding module, and determining the probability that each character in the training sentence belongs to the first word in the entity words or determining the probability that each character in the training sentence belongs to the last word in the entity words;

determining training entity words included in the training sentences according to the probability that each character belongs to the first word in the entity words or the probability that each character belongs to the last word in the entity words, and determining entity word prediction types corresponding to the training entity words;

and training an entity recognition model according to the entity words in the labeling information of the training sentences and the entity word types corresponding to the entity words, and the entity word prediction types corresponding to the training entity words, so as to obtain the trained entity recognition model.

It is understood that the division of the units in the embodiments of the present application is illustrative, and is only one logical function division, and there may be another division manner in actual implementation. Each functional unit in the embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, and can also be realized in a form of a software functional unit. Based on the same inventive concept, the principle and the advantageous effect of the data processing apparatus provided in the embodiment of the present application for solving the problem are similar to the principle and the advantageous effect of the data processing apparatus in the embodiment of the method of the present application for solving the problem, and for brevity, the principle and the advantageous effect of the implementation of the method may be referred to, and are not described herein again.

Please refer to fig. 7, which is a schematic structural diagram of an intelligent device according to an embodiment of the present application. The smart device described in this embodiment includes: a processor 701, a memory 702, and a network interface 703. Data may be exchanged between the processor 701, the memory 702, and the network interface 703. The smart device of the embodiment of the present application may be, for example, a smart terminal, such as a smart phone, a tablet computer, a smart wearable device, and the like, and the smart device may also be a server.

The Processor 701 may be a Central Processing Unit (CPU), and may also be other general purpose processors, Digital Signal Processors (DSP), Application Specific Integrated Circuits (ASIC), Field-Programmable Gate arrays (FPGA) or other Programmable logic devices, discrete Gate or transistor logic devices, discrete hardware components, and so on. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The memory 702, which may include both read-only memory and random-access memory, provides program instructions and data to the processor 701. A portion of the memory 702 may also include non-volatile random access memory. When the processor 701 calls the program instruction, it is configured to:

acquiring multimedia data associated with a target object identifier;

In one implementation, the processor 701 is specifically configured to:

In one implementation, the processor 701 is further configured to:

and displaying the search result on the search result browsing interface.

Based on the same inventive concept, the principle and the beneficial effect of solving the problem of the intelligent device provided in the embodiment of the present application are similar to the principle and the beneficial effect of solving the problem of the data processing apparatus in the embodiment of the present application, and for brevity, the principle and the beneficial effect of the implementation of the method can be referred to, and are not described herein again.

The embodiment of the present application also provides a computer storage medium, in which program instructions are stored, and when the program is executed, some or all of the steps of the data processing method in the embodiment corresponding to fig. 2 or fig. 4 may be included.

It should be noted that, for simplicity of description, the above-mentioned embodiments of the method are described as a series of acts or combinations, but those skilled in the art should understand that the present application is not limited by the order of acts described, as some steps may be performed in other orders or simultaneously according to the present application. Further, those skilled in the art should also appreciate that the embodiments described in the specification are preferred embodiments and that the acts and modules referred to are not necessarily required in this application.

Those skilled in the art will appreciate that all or part of the steps in the methods of the above embodiments may be implemented by associated hardware instructed by a program, which may be stored in a computer-readable storage medium, and the storage medium may include: flash disks, Read-Only memories (ROMs), Random Access Memories (RAMs), magnetic or optical disks, and the like.

Embodiments of the present application also provide a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. The processor of the smart device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the smart device to perform the steps performed in the embodiments of the methods described above.

The foregoing detailed description is directed to a data processing method, an apparatus, a device, and a storage medium provided in the embodiments of the present application, and specific examples are applied in the present application to explain the principles and implementations of the present application, and the descriptions of the foregoing embodiments are only used to help understand the method and the core ideas of the present application; meanwhile, for a person skilled in the art, according to the idea of the present application, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present application.

Claims

1. A data processing method, comprising:

acquiring multimedia data associated with a target object identifier;

2. The method of claim 1, wherein the object data processing of the target object identifier according to the entity words and the entity word types comprises any one or more of the following steps:

3. The method of claim 1, wherein the method further comprises:

and displaying the search result on the search result browsing interface.

4. The method of claim 1,

the input of the coding module is each character in the multimedia data arranged according to the language order, the output of the coding module is an L-dimensional feature vector corresponding to each character, and L is an integer greater than 1;

5. The method of claim 1,

the entity recognition model is obtained by training through a training sample set, the training sample set comprises a training sample pair, the training sample pair comprises a training sentence and the labeling information of the training sentence, and the labeling information of the training sentence comprises: entity words included in the training sentences and entity word types corresponding to the entity words;

6. The method of claim 5,

when M reference dictionaries are used for searching and determining a target entity word from a training sentence, determining the entity word type corresponding to the target entity word according to the types of the M reference dictionaries and the classification type of the training sentence, wherein M is an integer greater than 1;

7. The method of claim 5 or 6,

the training sample set is obtained by adjusting initial entity words in an initial training sample set and initial entity word types of the initial entity words by calling an entity classification model;

8. The method of claim 7,

the entity classification model is obtained by training an entity word sample set;

9. The method of claim 5, wherein training through the training sample set to obtain the entity recognition model comprises:

10. A data processing apparatus, comprising:

11. An intelligent device comprising a processor, a memory and a network interface, the processor, the memory and the network interface being interconnected, wherein the memory is configured to store a computer program comprising program instructions, and wherein the processor is configured to invoke the program instructions to perform the method of any of claims 1-9.

12. A computer storage medium, characterized in that the computer storage medium stores a computer program comprising program instructions that, when executed by a processor, cause a computer device having the processor to perform the method of any one of claims 1-9.