CN111368548A

CN111368548A - Semantic recognition method and device, electronic equipment and computer-readable storage medium

Info

Publication number: CN111368548A
Application number: CN201811498049.9A
Authority: CN
Inventors: 王颖帅; 李晓霞; 苗诗雨
Original assignee: Beijing Jingdong Century Trading Co Ltd; Beijing Jingdong Shangke Information Technology Co Ltd
Current assignee: Beijing Jingdong Century Trading Co Ltd; Beijing Jingdong Shangke Information Technology Co Ltd
Priority date: 2018-12-07
Filing date: 2018-12-07
Publication date: 2020-07-03

Abstract

The embodiment of the invention provides a semantic recognition method and device, electronic equipment and a computer readable storage medium, and relates to the technical field of computers. The semantic recognition method comprises the following steps: acquiring input information of a target object; extracting a target slot value of a target slot in the input information through a slot extraction model; the slot position extraction model comprises an input layer, a sharing layer and an output layer which are sequentially connected, wherein the output layer comprises n parallel sub-output layers, and each sub-output layer is used for outputting a slot value prediction result of a slot position; n is a positive integer of 2 or more. The technical scheme of the embodiment of the invention can realize the multi-slot prediction of the same word and improve the accuracy of semantic recognition.

Description

Semantic recognition method and device, electronic equipment and computer-readable storage medium

Technical Field

The present invention relates to the field of computer technologies, and in particular, to a semantic recognition method, a semantic recognition apparatus, an electronic device, and a computer-readable storage medium.

Background

In the prior art, the semantic recognition adopts a Stanford CoreNLP processing tool, the analysis of word segmentation, part of speech tagging and the like on the bottom layer of a basic language is firstly carried out, then regular template matching is written, and specific words in specific dialogues are extracted.

Here, the word is a template, for example, "i want to buy x" is a word, and it is seen that the word indicates that the user's shopping intention is a product query.

The disadvantages of the prior art are as follows:

the regular template in the Stenford CoreNLP processing tool is used for setting good-speech sentence patterns in advance by product personnel and writing program matching by using the regular expression. However, the template matching by the stirfur CoreNLP processing tool is regular, semantic information is extracted, and the template is dead, and the template matching can be performed only in the regular specified words.

For example, an after-market scenario, the dialog template may be set to "i buy x want to return", and the program, upon seeing such a sentence, predicts it to the after-market scenario, but this requires the user to say that the words match exactly, and only a complete match can be identified.

That is, in the prior art, matching can be performed only in the regular specified operation, and with the expansion of the business scene of the e-commerce platform, more and more regular templates need to be written, which wastes manpower and has inflexible effect.

Therefore, a new semantic recognition method, a semantic recognition apparatus, an electronic device, and a computer-readable storage medium are needed.

It is to be noted that the information disclosed in the above background section is only for enhancement of understanding of the background of the invention and therefore may include information that does not constitute prior art that is already known to a person of ordinary skill in the art.

Disclosure of Invention

An object of embodiments of the present invention is to provide a semantic recognition method, a semantic recognition apparatus, an electronic device, and a computer-readable storage medium, which overcome one or more of the problems due to the limitations and disadvantages of the related art, at least to some extent.

According to a first aspect of the embodiments of the present invention, there is provided a semantic recognition method, including: acquiring input information of a target object; extracting a target slot value of a target slot in the input information through a slot extraction model; the slot position extraction model comprises an input layer, a sharing layer and an output layer which are sequentially connected, wherein the output layer comprises n parallel sub-output layers, and each sub-output layer is used for outputting a slot value prediction result of a slot position; n is a positive integer of 2 or more.

In an exemplary embodiment of the present invention, the input information includes input voice information and/or input text information.

In an exemplary embodiment of the present invention, the target slot position includes any one or more of a product word slot position, a brand word slot position, a modifier slot position, a gender slot position, a price slot position, a query range slot position, and a channel number slot position.

In an exemplary embodiment of the present invention, further comprising: n kinds of slots of the target product class are predefined.

In an exemplary embodiment of the present invention, further comprising: generating a product word bank, a brand word bank and a brand product word pairing word bank; and establishing a knowledge graph based on the product word bank, the brand word bank and the brand product word pairing word bank.

In an exemplary embodiment of the present invention, further comprising: retrieving recommendation information from the knowledge graph according to a target slot value of the target slot position; and sending the recommendation information to the target object.

In an exemplary embodiment of the present invention, further comprising: and determining a target business scene corresponding to the input information through a business scene classification model.

In an exemplary embodiment of the present invention, the target service scenario includes any one or more of a commodity query scenario, an after-sales service scenario, a fuzzy offer query scenario, a special commodity offer query scenario, an order query scenario, a total station express scenario, and an unknown scenario.

In an exemplary embodiment of the present invention, further comprising: acquiring portrait information of the target object; the portrait information includes any one or more of purchasing power level, member level, category preference information, and sex information.

In an exemplary embodiment of the present invention, the target slot position includes a product word slot position and a brand word slot position, and the method further includes: and if the target slot value of the product word slot position in the input information is extracted, predicting the target slot value of the brand word slot position of the input information.

In an exemplary embodiment of the present invention, further comprising: negative semantics in the input information are identified.

In an exemplary embodiment of the present invention, further comprising: establishing a wrongly written word library; and automatically correcting the wrongly written characters of the input information based on the wrongly written character lexicon.

According to an aspect of the present invention, there is provided a semantic recognition apparatus including: an input information acquisition module configured to acquire input information of a target object; the slot position extraction model is configured to extract a target slot value of a target slot position in the input information through the slot position extraction model; the slot position extraction model comprises an input layer, a sharing layer and an output layer which are sequentially connected, wherein the output layer comprises n parallel sub-output layers, and each sub-output layer is used for outputting a slot value prediction result of a slot position; n is a positive integer of 2 or more.

According to an aspect of the present invention, there is provided an electronic apparatus including: a processor; and a memory having computer readable instructions stored thereon which, when executed by the processor, implement the semantic identification method as in any one of the above.

According to an aspect of the invention, there is provided a computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements a semantic recognition method as described in any one of the above.

In the technical solutions provided by some embodiments of the present invention, a target slot value of a target slot in input information is extracted through a slot extraction model; the slot position extraction model comprises an input layer, a sharing layer and an output layer which are sequentially connected, wherein the output layer comprises n parallel sub-output layers, and each sub-output layer is used for outputting a slot value prediction result of a slot position; n is a positive integer greater than or equal to 2, so that the multi-slot prediction of the same word can be realized, and the accuracy of semantic recognition is improved.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and together with the description, serve to explain the principles of the invention. It is obvious that the drawings in the following description are only some embodiments of the invention, and that for a person skilled in the art, other drawings can be derived from them without inventive effort. In the drawings:

FIG. 1 illustrates a flow diagram of a semantic recognition method according to some embodiments of the invention.

FIG. 2 illustrates a schematic diagram of a training slot extraction model according to some embodiments of the invention.

FIG. 3 illustrates a schematic diagram of a slot extraction model according to some embodiments of the invention.

FIG. 4 illustrates an interface diagram of a semantic recognition method according to some embodiments of the invention.

FIG. 5 illustrates an interface diagram of a semantic recognition method according to further embodiments of the present invention.

FIG. 6 illustrates an interface diagram of semantic recognition methods according to further embodiments of the invention.

FIG. 7 illustrates an overall architecture diagram of a semantic recognition method according to some embodiments of the invention.

FIG. 8 illustrates an interface diagram for an intelligent assistant journal, according to some embodiments of the invention.

FIG. 9 illustrates a schematic diagram of a portion of user input information according to some embodiments of the invention.

FIG. 10 illustrates a schematic diagram of a semantic recognition method according to some embodiments of the invention.

FIG. 11 illustrates a schematic diagram of knowledge graph construction according to some embodiments of the invention.

FIG. 12 illustrates a schematic diagram of a predicted outcome of user input information, according to some embodiments of the invention.

FIG. 13 is a diagram illustrating predicted results of user input information according to further embodiments of the invention.

FIG. 14 is a diagram illustrating predicted results of user input information according to further embodiments of the invention.

FIG. 15 illustrates a diagram of a predicted outcome of user input information according to still further embodiments of the invention.

FIG. 16 illustrates a diagram of predicted results of user input information, according to still further embodiments of the invention.

FIG. 17 illustrates a schematic block diagram of a semantic recognition apparatus according to some exemplary embodiments of the present invention.

FIG. 18 illustrates a schematic structural diagram of a computer system suitable for use with the electronic device to implement an embodiment of the invention.

Detailed Description

Example embodiments will now be described more fully with reference to the accompanying drawings. Example embodiments may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of example embodiments to those skilled in the art. The same reference numerals denote the same or similar parts in the drawings, and thus, a repetitive description thereof will be omitted.

Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to provide a thorough understanding of embodiments of the invention. One skilled in the relevant art will recognize, however, that the invention may be practiced without one or more of the specific details, or with other methods, components, devices, steps, and so forth. In other instances, well-known methods, devices, implementations or operations have not been shown or described in detail to avoid obscuring aspects of the invention.

The block diagrams shown in the figures are functional entities only and do not necessarily correspond to physically separate entities. I.e. these functional entities may be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different networks and/or processor means and/or microcontroller means.

The flow charts shown in the drawings are merely illustrative and do not necessarily include all of the contents and operations/steps, nor do they necessarily have to be performed in the order described. For example, some operations/steps may be decomposed, and some operations/steps may be combined or partially combined, so that the actual execution sequence may be changed according to the actual situation.

Some terms referred to in the embodiments of the present invention are first defined and explained below.

A knowledge graph is a structured semantic knowledge base that uses symbols to describe concepts and their interrelationships. The embodiment of the invention can accurately position and obtain the required knowledge in depth based on the knowledge map.

NLP: natural Language Processing, NLP, in most cases, refers to various large and small Language Processing applications on computers, and practical application programs constructed by NLP technology, such as word segmentation, part-of-speech tagging, named entity recognition, syntactic analysis, grammatical dependency, and the like.

Entity identification: the method refers to the steps of automatically extracting information entities from text corpora, and extracting valuable entities from a data source through machine learning (for example, product words, brand words, modifiers and the like in the embodiment of the invention can be regarded as entities).

The intelligent assistant: the embodiment of the invention refers to a channel of a mobile terminal of an e-commerce platform, such as a mobile phone APP (application), in which a user can input in a voice form or a text form, and an intelligent assistant can recognize semantic intent of information input by the user and further give corresponding personalized recommendation based on a pre-constructed knowledge graph.

Semantic recognition: the embodiment of the invention can be particularly applied to an intelligent assistant project of an e-commerce platform and is used for identifying the shopping intention of the input information or the input content of the user.

As shown in fig. 1, a semantic recognition method provided by an embodiment of the present invention may include the following steps.

In step S110, input information of the target object is acquired.

In the embodiment of the invention, the target object can be a user of an e-commerce platform. The input information can be collected through a mobile terminal or intelligent equipment of a user, and is sent to a background server for further processing, and then corresponding information is fed back to the target object.

In an exemplary embodiment, the input information may include input voice information and/or input text information.

For example, a product of an intelligent assistant is assumed on a mobile phone APP of an e-commerce platform and can be used for guiding a user to shop online, the user can input a sentence or a plurality of sentences in the voice of the intelligent assistant, the intelligent assistant robot can convert the input voice information into text information, analyze the text information, identify the user intention and an entity slot, search in a search interface, and recommend personalized goods.

In step S120, a target slot value of a target slot in the input information is extracted by a slot extraction model.

The slot position extraction model may include an input layer, a sharing layer, and an output layer connected in sequence, where the output layer includes n parallel sub-output layers, and each sub-output layer is used to output a slot value prediction result of a slot position; n is a positive integer of 2 or more.

In the embodiment of the invention, the slot position extraction model can realize that the same word in the input information is subjected to multi-slot position (slot) prediction.

In the prior art, only one slot position can be predicted for the same word by adopting a named entity recognition method such as BilSTM _ CRF (Bidirectional Long Short-Term Memory _ Conditional Random Field) or CRF + +, and the like. For example, a user inputs "i want to buy a mobile phone with six millets" through a smart assistant voice, the prior art predicts that "millets" are brand words and "six" are product words, wherein the prediction of "six" is that the product words are inaccurate, which affects subsequent knowledge retrieval steps, thereby causing that recommended commodity information fed back to the user may be wrong or not the most desired by the user.

In the embodiment of the invention, a multi-task sharing layer is designed in the slot extraction model, the defect that a word is mapped to a slot in the existing BilSTM _ CRF or CRF + + and the like is overcome, parameters of partial feature layer embedding (word vectors) are shared in the multi-task sharing layer, for example, product words, modifier slot models and brand word slot models can be trained in parallel, respective loss functions of the product words, the modifier slot models and the brand word slot models can be calculated in parallel, and finally, the product words, the modifier slot models and the brand word slot models are spliced together through a generalized full-connection layer (see the following figures 2 and 3 specifically), so that multi-slot prediction of the same word can be realized.

In an exemplary embodiment, the target slot may include any one or more of a product word slot, a brand word slot, a modifier slot, a gender slot, a price slot, a query range slot, a channel number slot, and the like.

It should be noted that the slot position is not limited to the above-mentioned ones, and the design may be adjusted according to the specific application scenario and requirement, and the present invention is not limited to this.

For example, semantic slots that may predefine user input information include:

(1) product, product word slot position of the commodity, represents the commodity name, namely the product word of the commodity.

In the embodiment of the invention, because the commodities have the titles of the commodities, but some commodities have long titles, the product words are the central words of the commodities.

For example, if the user input information is "i want to buy a mobile phone", the product word of the product in the input information is "mobile phone".

(2) And (4) Hunted-demo, a modifier slot of the commodity, and representing the description of the commodity.

For example, if the user inputs "i want to buy a rose gold mobile phone", the commodity description, i.e., the modifier, is "rose gold".

(3) deco _ of _ product, a query range slot for an article, indicates a query range for the article.

For example, if the user inputs information "what offers are in my shopping cart", the inquiry range is "shopping cart".

(4) brand word slot of the goods, which represents the brand of the goods.

For example, if the user inputs "i buy apple cell phone", the brand word of the product is "apple".

(5) channel, channel number slot, indicates the channel number.

For example, a "total station through" may be a special channel of an e-commerce platform, which may include channels specific to the e-commerce platform, such as "XX seconds kill", "XX to home", "XX crowd funding", "XX white bar", "XX wallet", etc. The definition of the channel is designed by the product, for example, the second killing means that a plurality of batches of second killing commodities are updated every day.

For example, if the user inputs information "i buy a selfie stick inside second killer", the channel number is "second killer".

In the following examples, the "product word" and the "brand word" are mostly taken as examples, because the product words and the brand words related to the information input by the user of the smart assistant channel are more and widely distributed. However, the present invention is not limited thereto.

In the embodiment of the present invention, in order to ensure the accuracy of the slot extraction model, a large amount of user input information, for example, 5 thousands of pieces (for example, the present invention is not limited thereto), may be collected in advance, and each slot is manually labeled, and the labeled data is used as a training set of the slot extraction model to train the slot extraction model.

It should be noted that, in the embodiment of the present invention, manual annotation may not be introduced in the data preprocessing stage, and the program uses rule processing or a crawler to obtain the annotation. For example, a rule is manually designed to label data, such as "i want to buy", a business scenario is intended as "commodity query", and the corresponding "", if in a product thesaurus, is a product word; if the brand word library exists, the brand word library is the brand word. As another example, a crawler utilizes existing feedback data as annotations.

In an exemplary embodiment, further comprising: n kinds of slots of the target product class are predefined.

In the embodiment of the present invention, the slot position may be predefined in the intelligent assistant service, for example, the "slot position _ slot value" of the mobile phone category may be defined as shown in table 1 below:

TABLE 1

Slot (KEY)	Slot VALUE (VALUE)
		Number of CPU cores	Eight cores
Price	0-199
		Pixel	Over 1600 thousands
Front camera pixel	Over 1600 ten thousand
		Size of screen	3.0 inches and below
Screen configuration	3D screen
		Characteristics of taking a picture	Optical zoom
Fuselage memory	32GB
		Fuselage color	White colour
Material of	TPU
		Style	Soft shell
Game arrangement	ID differentiation design
		Hot spot	1080P screen
Capacity of battery	1000 milliBelow a
		Network	Double-card single 4G
Old man machine configuration	Thermometer hardware support
		Style of a book	Brief introduction

In table 1, the first column indicates well-defined slot positions, the second column indicates slot values, and each slot value has multiple possible distributions. For example, the brand of the mobile phone may be a plurality of values such as "hua shi", "millet", "apple", and the like, and the plurality of values form the distribution of the brand slot of the mobile phone.

In an exemplary embodiment, the method may further include: generating a product word bank, a brand word bank and a brand product word pairing word bank; and establishing a Knowledge Graph (KG) based on the product word bank, the brand word bank and the brand product word pairing word bank.

In the embodiment of the invention, in order to be better applied to the intelligent assistant service, the knowledge data layer can maintain a characteristic knowledge map database with an electronic commerce platform under the category of mobile phones, and the database can be divided into three layers: the method comprises the steps of brand word bank construction, product word bank construction and brand product word pairing word bank construction. These will be described below.

Firstly, brand word bank construction.

In an embodiment of the present invention, a total brand lexicon may be maintained, and a table of the total brand lexicon may include two fields, a first field being a brand name and a second field being a brand source.

Wherein, the brand source can be divided into: a brand lexicon maintained by the big data part; a proprietary brand lexicon of an e-commerce platform (the proprietary brand lexicon is mainly the brand of an e-commerce, includes all online brands provided on the e-commerce platform, and has better support for e-commerce services); brand words obtained by an external crawler; high-quality brand words of operation maintenance; english brand words, etc.

In the brand lexicon, for example, "mobile phone" and "hua ye" represent entities, and "brand" is a relationship between the two entities, and the two entities are connected through the relationship of "brand".

And secondly, building a product word stock.

In the embodiment of the present invention, a total product thesaurus may be maintained, and a table of the total product thesaurus may include two fields, where the first field is a product word name and the second field is a product source.

Wherein, the product sources can be divided into: a product word maintained by the big data part; and the electronic commerce platform is a proprietary product word bank. Similarly, the proprietary product thesaurus may also be for e-commerce, such as "XX white bar" being a product word, but not necessarily so fine grained product words can be contained in the general product thesaurus.

And thirdly, establishing a brand product word pairing word bank.

In the embodiment of the invention, the matched brand words and product phrases are extracted from the commodity detail table of each commodity of the electronic commerce platform to form the product word matched word bank of the product phrase finished product card.

In an exemplary embodiment, the method may further include: retrieving recommendation information from the knowledge graph according to a target slot value of the target slot position; and sending the recommendation information to the target object.

In the embodiment of the invention, in an intelligent assistant channel of an e-commerce platform mobile phone APP, a user clicks a robot icon beside a search box, the robot can automatically answer the question of the user, and then the user is attracted to an interested module, and personalized recommendation is intelligently made for the user.

For example, the user inputs 'i want to buy a millet mobile phone' at the smart assistant, extracts that the product word is 'mobile phone' and the brand word is 'millet' in the sentence input by the user through the slot extraction model, inputs the target slot value 'mobile phone' of the product word slot and the target slot value 'millet' of the brand word slot into the service interface, and then calls the search interface to return the personalized commodity for the user according to the user portrait.

In the embodiment of the invention, the intelligent assistant can interactively shop with the user, screens out the most desired commodities for the user, and comprehensively improves shopping experience from brand price to performance quality.

For example, if the product word extracted by the slot extraction model into the user input information is "mobile phone", the robot asks the user "ask what brand of mobile phone you want? After the user answers millet, the robot asks the user to ask the user what color the user wants again until the robot knows enough slot values of the user, feels that a commodity can be recommended for the user, the interactive conversation is stopped, and then the robot takes the slot position information obtained through interaction, selects the most suitable commodity from a background knowledge map database and recommends the commodity to the user.

In an exemplary embodiment, the method may further include: and determining a target business scene corresponding to the input information through a business scene classification model.

In the embodiment of the present invention, the target service scenario may include any one or more of a commodity query scenario, an after-sales service scenario, a fuzzy offer query scenario, a specific commodity offer query scenario, an order query scenario, a total station express scenario, and an unknown scenario.

For example, the user inputs information "i want to buy x" and belongs to the commodity query scene; as another example, the user enters information "where did my order? "belongs to the order query scenario; also for example, a user entering the information "I want to return" belongs to an after-sales service scenario; as another example, the user inputs information "is there a preference? "belongs to the fuzzy preferential query scene; as another example, a user inputs information "do a mobile phone work again? "belongs to a special commodity preferential query scenario; for another example, "XX second kill" belongs to a total station through scene; as another example, the user input information "haha" belongs to an unknown scene.

It should be noted that, in different application scenarios and requirements, corresponding service scenarios may be designed, which is not limited in the present invention. In the following embodiments, the user shopping intent of the intelligent assistant refers to: commodity inquiry, order inquiry, after-sale service, fuzzy preference inquiry, special commodity preference inquiry, total station direct, unknown and other 7 service scenes.

In the embodiment of the invention, the business scene classification model can use a Convolutional Neural Network (CNN) to realize a scene classifier, and in the intelligent assistance, the shopping intention of a user is determined by the scene classifier, and then slot values of slots such as product words and brand words are extracted to be used for recommending personalized commodities to the user.

In the embodiment of the invention, the sorted and summarized knowledge is provided for the user through a service scene classification model (or also called a service scene classifier) and a slot extraction model (or also called a semantic recognition model or a natural language understanding module).

In the embodiment of the invention, in a knowledge calculation layer, a user inputs a sentence or a plurality of sentences in an intelligent assistant channel, the service scene classification model (which may also be called as an NLU (natural language understanding) intention classifier) and the slot extraction model are called first to perform intention identification and slot filling (for example, if a slot has a "brand word" and a "product word", the corresponding slot may be filled with a "huashi" and a "mobile phone", respectively), and then a session manager model may be used to identify a relationship between current input information of the user and a previously input preceding text, and judge whether a memory state is reserved.

For example, what kinds of cameras are available in a mobile phone and are listed in the mobile phone, "is a user input voice message? ", the robot will go to the knowledge map database, inquire the corresponding Key and Value, and feed back to the user.

In the embodiment of the invention, the interaction between the intelligent assistant and the user can be a process of a multi-turn conversation, the slot position of the current input information can be predicted by applying the slot position extraction model, if the slot position is overlapped with the slot positions in the previous conversations, the current input information is considered to be related to the previous words, otherwise, the current input information is considered to be not related. In some embodiments, there may be a data structure that records user slots, and at most records a specified slot, the state is cleared.

In the embodiment of the invention, the session manager model can be realized by being divided into two submodels, wherein the first submodel judges whether the information currently input by the user contains slot positions such as product words, brand words and the like extracted from the previous text through service logic, and if the information currently input by the user contains slot positions such as product words, brand words and the like, the information currently input by the user is considered to be related to the previous text; if not, the current input information of the user can be considered to be not related to the former; the second sub-model may be a two-class neural network model, and the two-class neural network model identifies that the associated prediction output is 1 and the non-associated prediction output is 0 (for illustration purposes only, the present invention is not limited thereto).

It should be noted that the commodity question-answering in the embodiment of the present invention is a question-answering system that is adapted to the business of an e-commerce platform and is trained by adding a customized corpus of the e-commerce platform (for example, the above-mentioned proprietary product thesaurus and proprietary brand thesaurus), debugging a network structure and parameters, in combination with a specific scenario of the e-commerce platform.

For example, a user asks "what prices are all available to the cell phone" in the smart assistant? The robot recognizes that the slot position of the current input information is a brand word 'Hua' and a product word 'mobile phone', asks a 'price' slot position, searches a plurality of price sections in a knowledge graph, and returns the price value to the user.

In an exemplary embodiment, the method may further include: acquiring portrait information of the target object; the representation information may include any one or more of purchasing power level, member level, category preference information, sex information, and the like.

In the embodiment of the invention, the slot position extraction model can be combined with a target service scene in user input information identified by the service scene classification model to extract a target slot value of a target slot position in the input information and portrait information of the target object, and personalized recommendation information matched with the target object is retrieved from a knowledge graph constructed under each category of an e-commerce platform constructed in advance and returned to the target object.

In an exemplary embodiment, the target slot may include a product word slot and a brand word slot, and the method further includes: and if the target slot value of the product word slot position in the input information is extracted, predicting the target slot value of the brand word slot position of the input information. See in particular the examples of fig. 15 and 16 below.

In an exemplary embodiment, the method may further include: negative semantics in the input information are identified.

In the embodiment of the invention, the negative slot identification of the knowledge graph can be realized. The recognition of the negative intention and the negative slot position is applied to the intelligent assistant service, and the negative semantics can be recognized through the semantic dependency algorithm.

For example, the user inputs information "i want to buy a millet mobile phone and do not need the screen to be large", and then the robot recognizes that the negative slot value (slot _ value) corresponding to the slot (slot _ key) of the "screen" is large, and the mobile phone with the large screen is not recommended to the user. The knowledge graph is associated with the entity of the large-screen mobile phone, so that the effect of associating the graph entity is achieved on the basis of understanding the semantics (the large is a modifier, and the screen is a specific value of a slot position of the mobile phone, so that the semantic understanding effect is improved together). In the prior art, a negative semantic recognition function is not provided, after the negative semantic recognition function provided by the embodiment of the invention is added, the user experience is improved, the AB test (A/B Testing) effect is superior to that of the prior art, and the model is continuously iterated and continuously optimized according to badcase fed back by a user and testers.

In an exemplary embodiment, the method may further include: establishing a wrongly written word library; and automatically correcting the wrongly written characters of the input information based on the wrongly written character lexicon.

In the embodiment of the invention, on one hand, a user inputs the voice of the intelligent assistant, and wrongly written characters possibly occur in the voice recognition stage; on the other hand, the user can input the text of the intelligent assistant by himself or herself, and the wrongly written characters can also appear. Where the wrongly written word corpus may comprise two columns, the first column being the wrong word and the second column being the correct word, for example "addis ═ adidas". See in particular the examples of fig. 13 and 14 below.

In the following embodiments, n is 2, and the output of the slot value prediction result of the product word slot model corresponding to the first sub-output layer and the output of the slot value prediction result of the brand word slot corresponding to the second sub-output layer are taken as an example for description, but actually, the number of sub-output layers in the slot extraction model may be determined according to the number of preset slots.

As shown in fig. 2, X represents input information input to the input layer of the slot extraction model, Y1 represents a true tag of each product word in the input information training set corresponding to the first task, and Y2 represents a true tag of each brand word in the input information training set corresponding to the second task.

In the embodiment of the present invention, in the process of training a model, when input information in an input information training set is input to the model, the first loss function may be obtained according to the product word prediction result output by the first sub-output layer and the true label Y1 of the first task; when the input information in the input information training set is input into the model, the second loss function can be obtained according to the brand word prediction result output by the second sub-output layer and the real label Y2 of the second task, and the first loss function and the second loss function are optimized by the first optimizer and the second optimizer respectively, so that the training process of the model is completed.

In the embodiment of the invention, each task has a respective output layer, a loss function and an optimizer.

For example, for each task, a loss function (also called an optimization function) is established, which may be in the specific form:

Loss＝α*cross_entropy_loss+β*regularization_term(1)

wherein Loss in the above formula (1) represents a Loss function of a corresponding task, α represents a weight of cross _ entry _ Loss of each task, the weights of different tasks can be different, β is a regularization function weight, different tasks can adopt a uniform value, cross _ entry _ Loss is a cross entropy Loss function, the smaller the difference between a prediction result and a real label is, the smaller the value is, and regularization _ term can adopt an L2 norm (norm) of all trainable parameters, the smaller the absolute value of the parameter is, the smaller the value is.

In the embodiment of the present invention, a random gradient descent method (but the present invention is not limited thereto) may be adopted for training, each time a part of data sets in a training set is input, an optimizer corresponding to this task is optimized, so as to update a shared parameter of an overlapping portion of the two tasks and an independent parameter unique to each corresponding task.

In the embodiment of the invention, a slot position extraction model is established by adopting a method of combining multi-task learning and deep learning, a training set is used as all input information of the whole model, and the prediction accuracy of each task can be enhanced on the premise of the same training data volume.

In the embodiment of the present invention, the multi-task learning is combined with the deep learning, and the new network structure is as shown in fig. 3, and the last layer, i.e., the fully-connected layer, of the BiLSTM _ CRF network structure may be divided into a plurality of parallel fully-connected layers (for example, the fully-connected layer 1 and the fully-connected layer 2 shown in fig. 3, the present invention is not limited thereto, the number of parallel fully-connected layers depends on the number of tasks), and each task has its own output layer parameter. For example, fully connected layer 1 may be used to output a first prediction of a product word slot, and fully connected layer 2 may be used to output a second prediction of a brand word slot.

In the embodiment of the present invention, the sharing layer may be different according to the depth convolutional neural network used, and is not limited to the above-mentioned BiLSTM _ CRF.

For example, in other embodiments, a double-layer LSTM-RNN (where the LSTM is all called Long Short-Term Memory in english, and the Long-Short Term Memory Network is a time-recursive Neural Network; the RNN is all called a secure Neural Network, and is a type of Neural Network for processing sequence data) may be trained, and the first layer is also called an input layer, where the input content includes an embedded word vector and a position vector; then into a second hidden layer and finally an output layer. The first layer LSTM inputs word vectors, position characteristics and parts of speech to identify entities such as product words and brand words, distributed expression of a hidden layer in the LSTM and classification label information of the entities obtained through training are used as input of a second layer RNN model, and dependency paths among the entities are input at the second layer.

In the embodiment of the invention, the multi-mask sharing layer mainly refers to the input information of a user, such as 'I want to buy the mobile phone with six apples', wherein 'apples' are identified as brand words and are predicted once through a BilSTM _ CRF neural network; the method is characterized in that the 'apple six' is recognized as a product word, and one prediction is also made through a BilSTM _ CRF neural network, namely, for two words of the 'apple', the two words appear only once in user input information, but the slot position of the 'product word' and the slot position of the 'brand word' are predicted, the two predictions are called as a 'multi-task' model by using two models which are simultaneously parallel, and the two models are called in parallel on engineering.

In the embodiment of the invention, product word prediction and brand word prediction are respectively made into two models, the input data are the same and are labeled differently, but the design of a feature layer is shared, and a loss function is calculated separately.

Fig. 4 is a schematic diagram of a pre-release interface of a semantic recognition algorithm according to an embodiment of the present invention.

As shown in fig. 5, the user inputs "i want to buy six mobile phones with millet", and the embodiment of the present invention may predict "six millet" as a product word and "millet" as a brand word at the same time through the slot extraction model, that is, the word "millet" appears only once in the user input information, but is predicted to two slots at the same time.

As shown in fig. 6, the user inputs "i want to buy an oppoR15 mobile phone," and the slot extraction model provided by the embodiment of the present invention can predict "oppoR 15" as a product word and "oppo" as a brand word.

The semantic recognition method provided by the embodiment of the invention can realize multi-slot prediction of the same word in the same input information, is more suitable for an electronic commerce scene, and can recommend more accurate personalized commodities to a user in the specific application of an intelligent assistant, for example, in the example shown in the figure 6, the intelligent assistant can finely recommend a mobile phone shown to the user oppo R15, not just an oppo mobile phone.

With the advent of the artificial intelligence era, knowledge-graph technology has received a great deal of attention from both the industry and academia. How to extract useful knowledge from mass data is the key of big data analysis of e-commerce websites. The knowledge map technology provides a means for abstracting structured knowledge from massive texts, so that the method has a wide application prospect in e-commerce websites.

The speech recognition and semantic understanding are trends of shopping without caller websites, users speak a sentence to the robot, and after the speech is converted into characters, how to accurately grasp the intention of the users becomes more and more important. It is against this background that the semantic recognition method proposed in the embodiment of the present invention may be an NLP semantic recognition method based on knowledge-graph improvement, and is used to extract key information of user input content.

The knowledge-graph-based improved NLP semantic recognition method provided by the embodiment of the invention can be applied to an intelligent assistant project of an electronic commerce platform and aims to accurately recognize semantics and extract key information of user voice or manually input information.

The core flow of the semantic identification method based on the knowledge graph provided by the embodiment of the invention is shown in FIG. 7. Among them, the lowermost part in fig. 7 is a part of data preprocessing. Data preprocessing is the cleansing of the business data of the intelligent assistant so that the data can be used as input to the model. The natural language processing NLP part introduces a specific application background of a slot extraction model in an intelligent assistant product; knowledge storage, knowledge fusion, knowledge calculation and knowledge application are data to be processed, and in a specific service background, how to apply a knowledge graph to improve the effect.

In the embodiment of the invention, the unstructured data is initial input information of a user in the intelligent assistant, and the noise is large; the semi-structured data is valuable intelligent assistant data obtained after a product is added into a buried point; structured data is data that is processed with a distributed processing tool and then stored in a background database table.

In the embodiment of the invention, in the aspect of processing data, firstly, a text (namely, a sentence for extracting user voice or text and interacting with an intelligent assistant) is extracted from user data, after the user input is obtained, entities such as product words, brand words, modifiers and the like need to be identified through a natural language processing technology, and in the process of semantic identification, word segmentation and part-of-speech tagging are used (output information obtained after data preprocessing is input to an NLP in a figure 7 for further processing).

It should be noted that the user data mentioned above refers to voice or words input by the user through the intelligent assistant, the interaction between the user and the intelligent assistant is divided into two parts, the first part is the voice input of the user and is converted into text by calling the voice recognition module, the second part is the text directly input by the user, and the user data refers to extracting the two parts of interaction data of the user.

In the embodiment of the invention, the natural language processing NLP module comprises models of named entity recognition, similarity calculation and the like, for example, the whole knowledge graph is communicated under the mobile phone category, and NLP and E-business recommendation are linked together in actual service to generate application value.

In the embodiment of the invention, knowledge calculation mainly obtains more implicit knowledge according to information provided by a knowledge map, for example, the implicit knowledge of implicit product words, brand words, modifiers and the like in the user dialect is presumed by writing a regular expression through word segmentation and part-of-speech tagging. Regular expression programs can be written, product words and brand words of users are matched in a fuzzy mode, and more similar products and brand words are mined from the knowledge graph.

For example, the user inputs information that "i want to buy a glory 8 cell phone," and the knowledge graph associates "glory 8" with a "Huayi" brand, recommending a Huayi series of goods to the user.

In the embodiment of the invention, in the slot position extraction model construction process, noise characteristics are designed and formulated, the noise characteristics can be applied to an input error correction model, unordered and unimportant information in a user input text is filtered, and wrongly written characters and important correction are carried out for use, so that even if the wrongly written characters are input by a user, the model can also identify, for example, the user inputs 'I wants to buy a small-life classmate beverage', the method provided by the embodiment of the invention can identify that the 'hit' characters in the 'small-life classmates' are wrongly written characters, and can predict the words as correct brand words under the current context.

The knowledge application in the embodiment of the invention refers to product words, brand words and the like predicted (namely recognized) based on a knowledge graph in business, and the product words, the brand words and the like serve intelligent assistants to better recognize user intentions.

The embodiment of the invention can also carry out association recommendation on users of the intelligent assistant channel based on entity association relation mining of the knowledge graph. For example, the user inputs "i thirsty", the robot digs out "thirsty" and "drink" in the knowledge map, so as to feed back to the user "dong recommends the following drink for you, then a bottle of cola? ", the application of knowledge maps, enhances the user experience of the intelligent assistant.

In the embodiment of the present invention, the main source of the preprocessed data is an intelligent assistant landing log table (that is, the HIVE voice log table in fig. 7 means that the service data of the user is written into a background data table of the intelligent assistant, and landing means that the service data is written into the data), the log of the intelligent assistant landing into a big data HIVE table, all fields are as shown in fig. 8, the fields used in the embodiment of the present invention are user input contents, and the "input contents" field is used as user data.

It should be noted that, the service scenario mentioned above has a part of data derived from the "service scenario" here, the service scenario data is based on the matching of the speech template, and a part of data is derived from the manual annotation.

In the embodiment of the present invention, after removing the user input without information content of the garbage through the regular matching, part of the user input is as shown in fig. 9.

For example, through regular matching of programs, text that is "haha" or "good" is filtered out if the user input information is within three words, and for example, the user input information is filtered out within three words.

In the embodiment of the invention, word segmentation and part-of-speech tagging are required to be used in the basic part of semantic recognition, and text features are extracted (namely, feature extraction of fig. 7 is a main link of data processing, all the following models need data with well-extracted features as input, and only different models need feature input with different formats) and semantic parsing, syntactic dependency and the like are required to be used.

In the embodiment of the invention, in order to improve the word segmentation effect, a special brand word bank and a special product word bank of an electronic commerce platform are added; the part-of-speech tagging can adopt a CoreNLP tool of Stanford to tag out proper nouns, person name pronouns, verbs and the like in the information input by the user and further serve as text characteristics; semantic parsing and syntactic dependency are to know the sentence grammar structure according to the context, generate different grammar trees, and find the best parse tree in the process of repeated backtracking. That is, in the embodiment of the present invention, the prior art may be used as disaster recovery data, the service scene classification model and the slot extraction model provided in the embodiment of the present invention are first used to predict the user intention, the slot of the user is extracted, and if the model cannot be captured, the prior art is used as a bottom-finding scheme, and the key steps are as shown in fig. 10.

In the embodiment of the invention, the special brand word bank and the special product word bank of the electronic commerce platform are added in the word segmentation stage, so that the word segmentation accuracy can be improved, and the word segmentation method can be used as a pre-processing program for labeling, and pre-label data for a labeling person for reference.

For example, the proprietary product thesaurus may include product words with e-commerce characteristics such as "XX sec killing", "XX blank bar", "blank bar expired", "XXE card", and the like.

For another example, the ordinary Chinese word segmentation of 'more grains' can be divided into two parts of 'grains' and 'more', but after a special product word stock of the electronic commerce platform features is added, the 'more grains' can be taken as a whole word segmentation result.

The syntactic parse in fig. 10 refers to parsing out sentence components, and refers to syntactic dependency in fig. 7.

Part of the semantic level on the right side of fig. 10 refers to semantic parsing of fig. 7, and the other part refers to semantic understanding, namely, slots for outputting product words, brand words and the like from the text. And then displaying product words, brand words and the like, and fusing the product words, the brand words and the like with a rear module.

The meaning of the semantic chunk in fig. 10 is that when the intelligent assistant performs personalized recommendation, product words, brand words, etc. are concatenated together to comprehensively recommend products.

The text classification in fig. 10 refers to a sentence or a plurality of sentences input by the user into the intelligent assistant, and the business scenario classification model classifies the sentences into different business scenarios to identify the shopping intention of the user.

In the embodiment of the present invention, the process of constructing the knowledge graph is shown in fig. 11.

The syntactic analysis-specific feature design in fig. 11 is to extract features using a syntactic dependency model for user input information. Fig. 7 is a macroscopic structure diagram, and fig. 11 is a relatively fine-grained structure.

In the embodiment of the invention, part of the knowledge graph is visualized by associating related entities with the proprietary product thesaurus and the proprietary brand thesaurus; the other part is that in the process of interaction between the intelligent assistant and the user, for example, new product words and brand words extracted by the slot extraction model are continuously added into the existing knowledge map, and the knowledge map library is continuously expanded.

In fig. 11, in the knowledge acquisition stage, in order to improve the quality of the knowledge service and provide satisfactory answers for users, the knowledge graph of the embodiment of the present invention not only includes the proprietary product thesaurus and the proprietary brand thesaurus, but also can discover and add new knowledge in time, and the amount and quality of knowledge determine the extent and depth of the service that can be provided by the knowledge graph, so that the knowledge graph construction needs to be supported by efficient knowledge acquisition. The general knowledge (for example, the mobile phone includes Huashi, millet and other brands, and similar contents known by most users) of the embodiment of the invention is mainly obtained from the structured data of the intelligent assistant log, including the user input information under different service scene classifiers. With the input of the user in the intelligent assistant emerging in a large quantity, the generated content of the user is increased continuously, a large number of users make contributions to the construction of a semantic network, new knowledge can find new product words, brand words and modifiers from the input information of the user, and the coverage rate of the knowledge is expanded continuously.

The knowledge graph of the embodiment of the invention provides related recommendations according to the user interests, so the behavior data of the user is extracted content, including the member level of the user on the e-commerce platform, the purchasing power of the user, the sex of the user and the commodity type preference of the user, and a part of behavior descriptions such as supplementary modifiers are also obtained from the data.

For example, if the user inputs "i want to buy a mobile phone", the purchasing power level of the user is extracted from the user representation of the knowledge graph, and a commodity suitable for the consumption level of the user is recommended.

With continued reference to fig. 11, the user inputting the voice or text information, i.e., the user inputting data, means that the user inputs the original input of the intelligent assistant, e.g., "i want to buy the mobile phone" as the user input, and the structured information, e.g., the product word and the brand word, is predicted from the original input information of the user through the slot extraction model, so as to form the structured data. The process of forming the knowledge graph is a process of associating the predicted entity slot positions, and complex models can be made at the later stage of the knowledge graph, for example, slot positions of product words, brand words and the like are predicted by inputting user input information through a slot position extraction model, then the slot positions are input into an entity relationship extraction model, and the relationship of the entities is predicted.

In the embodiment of the invention, in the knowledge fusion stage, since most of knowledge in the knowledge map is input by the e-commerce website user, the problems of repeated knowledge, unclear relation between knowledge and the like exist, and the knowledge fusion is performed. Knowledge fusion is a high-level abstraction of knowledge. Knowledge fusion mainly comprises two parts: entity linking and knowledge consolidation.

The entity link in the embodiment of the invention is to extract product words, brand words and the like from user input information, and design a characteristic kernel function to perform entity disambiguation, for example, an apple in 'I wants to buy an apple mobile phone' is a brand word, and an apple in 'I wants to eat three jin apples' is a product word.

In the embodiment of the present invention, the "feature kernel function" is a mathematical formula, and specifically, the product words "addis" and "addis" are calculated to be the same category by referring to the prior art, for example, the user "addis's clothes" and "addis's clothes" through the feature kernel function.

In the embodiment of the invention, the semantic features not only include bag-of-word vectors, but also include the relationships between preceding and following words. The specific design of semantic features of the embodiment of the invention is characterized in that a word bag feature is constructed for a user common Chinese character word stock, the frequency of current words in the word stock is a feature, and context relationship features refer to context words of the current words, which jointly form a part of slot position extraction model features. After various semantic features are constructed, the semantic features can be input into a subsequent slot extraction model, and then more accurate slot extraction is performed.

In the embodiment of the invention, knowledge merging (i.e. duplicate removal of repeated knowledge) refers to that existing structured data can be acquired from a third-party knowledge base (for example, a hundred-degree knowledge base) when a knowledge graph is constructed, so that the coverage and the accuracy of a slot extraction model are improved.

For example, if a user asks "who the author of the western book is through an intelligent assistant? "the answers to this question may be included in a third party knowledge base, the addition of which allows the knowledge graph to be expanded from the e-commerce domain to a greater number of domains.

In the embodiment of the invention, the knowledge storage in the knowledge graph is an association set with a large scale, and the orderly and available associated knowledge is formed through unstructured voice information input by an intelligent assistant user through early-stage fusion and processing (for example, specific knowledge such as that a product word is 'mobile phone', a brand word is 'apple', and a modifier word is 'rose' is extracted from 'apple mobile phone i want to buy rose' input by the user), and the knowledge is stored in different modules (such as a brand word module, a product word module, a modifier word module and the like) in a knowledge graph database in a relatively standard form according to the category of the knowledge, wherein the brand word module can be further subdivided into brand of clothing, brand of household appliance and the like).

In the embodiment of the invention, the knowledge category can comprise common knowledge and e-commerce specific field knowledge, wherein the e-commerce specific field knowledge can be divided into three delicacies, clothes, household appliances, electronic products and the like.

The linguistic data (such as logs of an intelligent assistant and text features extracted from the bottom layer) in the invention are stored in a big data Distributed HDFS (Hadoop Distributed File System), and the trained models (the models comprise a business scene classification model and a slot extraction model) are stored in Redis. The corpus is offline and calculated in advance, has no time and timeliness requirements, and is a large amount, so the corpus is stored on the HDFS. The model is small and may be called in real time, and is stored in redis, but the present invention is not limited thereto. The stored linguistic data is used by a follow-up module, and the model is used for predicting real new data on the line.

The knowledge retrieval in the embodiment of the invention is based on a commodity width table, a user width table and a commodity category corresponding relation table which are constructed by an electronic commerce platform, so that the intelligent identification of product words, brand words, modifiers and the like of intelligent assistants is realized, and the corresponding semantic intentions of users are found in a knowledge base through commodity similarity (for example, a vector space under the intention of a 'commodity query' business scene has a plurality of known vectors, new commodities to be predicted on a line are calculated, the similarity of the vector space with each intention vector space is calculated, and the model considers that the most similar one is the intention of the user). And the user inputs information, the information is matched with knowledge in a knowledge base after semantic understanding analysis, and statistics, sorting, reasoning, recommendation, prediction and the like are carried out. Wherein, retrieving valuable information according to user input statements is a data processing stage of the knowledge graph.

In the embodiment of the invention, the commodity width table stores commodity information, such as the color, size, order quantity and the like of a commodity; the user wide table stores user information, such as the gender and purchasing power level of the user; the commodity category corresponding relation table stores specific commodities under three categories, such as Huashi and millet under mobile phone categories.

In the embodiment of the invention, a part of data in the knowledge graph is derived from three tables, namely a commodity width table, a user width table and a commodity class corresponding relation table.

In the embodiment of the invention, the functions of the knowledge graph are mainly embodied in the aspects of knowledge organization, display and personalized recommendation:

first, the user is provided with the correct, ideal answer, to some extent overcoming natural language ambiguity.

It should be noted that overcoming ambiguity in natural language is not exactly equivalent to correcting a wrongly written word, but rather can be related in a knowledge graph based on the context of the current word.

For example, the user inputs information that "i want to buy apple mobile phone", after the machine extracts the "apple" slot position, the machine finds association in the knowledge map in combination with the following "mobile phone", and then can recognize that the "apple" is an electronic product and is not eaten fruit.

And secondly, providing the sorted and summarized knowledge to a user through a service scene classification model and a slot position extraction model.

In the embodiment of the invention, in the intelligent assistance, the shopping intention of a user is determined through a business scene classification model, and then slot positions such as product words and brand words are extracted to be recommended to the user.

Thirdly, through information recommendation, broader and deeper knowledge is provided, the knowledge graph tries to reason about search records related to other users, namely, the relevance between entities is found in the knowledge graph, the reasoning with the maximum similarity is carried out, the user is helped to answer the next question before asking questions, the search interest of the user in shopping of the E-commerce website is stimulated, and therefore a brand new shopping operation is carried out.

Compared with the traditional Stenford CoreNLP regular matching template, the semantic identification method provided by the embodiment of the invention is flexible and wide in coverage.

Fig. 12 shows that the user inputs "i want to own apple phone," identifies "apple" as a brand word, identifies "mobile phone" as a product word, and actcommodity represents a "commodity query" business intention for shopping, using the slot extraction model based on the knowledge graph provided by the embodiment of the present invention.

The embodiment of the invention mainly comprises the steps of constructing a common and specific knowledge base of the e-commerce based on a knowledge graph, identifying brand words and product words in the information currently input by a user according to the information input by the user, and then retrieving the content of the same and/or similar brand words and product words in the knowledge base and feeding back the content to the user.

In some embodiments, the text input by the user or the text converted from speech has a high possibility of having wrong words and discordant sentences, and the method provided by the embodiment of the invention can detect and correct the errors.

As shown in fig. 13, when the user inputs "apple bar phone", the slot extraction model provided in the embodiment of the present invention predicts "apple 8" as a product word, and "apple" as a brand word.

As shown in fig. 14, the user inputs "iPhone difference. "the slot extraction model provided by the embodiment of the invention can predict that" iPhoneX "is a brand word and" applet "is a brand word.

The semantic recognition method provided by the embodiment of the invention can realize automatic correction of wrongly written characters, improves the accuracy of semantic recognition, and can return more accurate recommendation results for users, thereby improving the user experience.

In the embodiment of the invention, the slot extraction model can also realize brand mapping.

For example, under the brands of millet, Hua Yi and the like under the mobile phone category, the mobile phone further comprises a plurality of sub-series mobile phones, the slot extraction model provided by the embodiment of the invention can map the main brands of the series of sub-series mobile phones, can be configured in a program, maintains a mapping table, then runs the program, and the system is automatically updated.

As shown in fig. 15, the user inputs "i want to buy a red rice four-cell phone," and the slot extraction model provided in the embodiment of the present invention can predict that "red rice four" is a product word and "millet" is a brand word.

As shown in fig. 16, the user inputs "glory four cell phone. "the slot extraction model provided by the embodiment can predict that" honor four "is a product word and" huayi "is a brand word.

The slot position extraction model provided by the embodiment of the invention can further realize brand word mapping points, not only keeps the original mobile phone subsystems as product words, but also extracts accurate mobile phone main brands, and increases the acceptance of users to the intelligent assistant algorithm while improving the service click conversion rate.

According to the semantic recognition method provided by the embodiment of the invention, on the aspect of network layer optimization of the model, different optimization techniques such as adding masks for different services are adopted, and a plurality of models of the comprehensive knowledge graph are applied to specific services in the east of Beijing, so that the model effect is improved. In addition, by combining a Stanford CoreNLP natural language processing tool, an electronic commerce platform characteristic commodity library (namely a special brand lexicon and a special product lexicon) and artificial labels (in the scheme, the Stanford CoreNLP can be used for word segmentation and part of speech labeling and also used as labels for identifying product words and brand words), better entity relation classification data is provided for the construction of the knowledge graph (for example, the sequence label model identifies a 'mobile phone' entity and a 'Huawei' entity, the corresponding relation is 'brand', a 'half-sleeve' entity and 'pink' are corresponding to 'color', and the like), and the accuracy of model prediction semantics is improved; on the business level, compared with the traditional idea of template matching, the improved algorithm based on the knowledge graph is flexible and covers more users, the user experience and click conversion rate of the intelligent assistant project are improved, and a method for better screening favorite commodities is provided for the E-commerce website users.

In addition, in the embodiment of the invention, a semantic recognition device is also provided. Referring to fig. 17, the semantic recognition apparatus 1700 may include: an input information acquisition module 1710 and a slot extraction model 1720.

The input information acquisition module 1710 may be configured to acquire input information of a target object.

The slot extraction model 1720 may be configured to extract a target slot value of a target slot in the input information through the slot extraction model.

In an exemplary embodiment, the semantic recognition apparatus 1700 may further include a slot definition module, which may be configured to pre-define n slots of the target category.

In an exemplary embodiment, the semantic recognition apparatus 1700 may further include: the word bank building module can be configured to generate a product word bank, a brand word bank and a brand product word pairing word bank; a knowledge graph establishing module configured to establish a knowledge graph based on the product thesaurus, the brand thesaurus and the brand product word paired thesaurus.

In an exemplary embodiment, the semantic recognition apparatus 1700 may further include: a knowledge retrieval module configured to retrieve recommendation information from the knowledge-graph according to a target slot value of the target slot; a recommendation sending module may be configured to send the recommendation information to the target object.

In an exemplary embodiment, the semantic recognition apparatus 1700 may further include a scene classification module, and the scene classification module may be configured to determine a target business scene corresponding to the input information through a business scene classification model.

In an exemplary embodiment, the target service scenario may include any one or more of a commodity query scenario, an after-sales service scenario, a fuzzy offer query scenario, a special commodity offer query scenario, an order query scenario, a total station express scenario, an unknown scenario, and the like.

In an exemplary embodiment, the semantic recognition apparatus 1700 may further include: a portrait information acquisition module that may be configured to acquire portrait information of the target object; the representation information may include any one or more of purchasing power level, member level, category preference information, sex information, and the like.

In an exemplary embodiment, the target slot may include a product word slot and a brand word slot, and the semantic recognition apparatus 1700 may further include: the brand word prediction module may be configured to predict a target slot value of a brand word slot of the input information if the target slot value of the product word slot in the input information is extracted.

In an exemplary embodiment, the semantic recognition apparatus 1700 may further include: a negative intent recognition module may be configured to recognize negative semantics in the input information.

In an exemplary embodiment, the semantic recognition apparatus 1700 may further include: the wrongly-written character word stock establishing module can be configured to establish a wrongly-written character word stock; the wrongly written character correcting module may be configured to automatically correct wrongly written characters of the input information based on the wrongly written character lexicon.

Since each functional module of the semantic recognition apparatus 1700 according to the exemplary embodiment of the present invention corresponds to the steps of the exemplary embodiment of the semantic recognition method described above, it is not described herein again.

In an exemplary embodiment of the present invention, there is also provided an electronic device capable of implementing the above method.

Referring now to FIG. 18, shown is a block diagram of a computer system 1800 suitable for use with the electronic device implementing an embodiment of the present invention. The computer system 1800 of the electronic device shown in fig. 18 is only an example, and should not bring any limitations to the function and scope of the embodiments of the present invention.

As shown in fig. 18, the computer system 1800 includes a Central Processing Unit (CPU)1801, which can perform various appropriate actions and processes in accordance with a program stored in a Read Only Memory (ROM)1802 or a program loaded from a storage portion 1808 into a Random Access Memory (RAM) 1803. In the RAM 1803, various programs and data necessary for system operation are also stored. The CPU 1801, ROM 1802, and RAM 1803 are connected to each other via a bus 1804. An input/output (I/O) interface 1805 is also connected to bus 1804.

The following components are connected to the I/O interface 1805: an input portion 1806 including a keyboard, a mouse, and the like; an output portion 1807 including a display such as a Cathode Ray Tube (CRT), a Liquid Crystal Display (LCD), and the like, and a speaker; a storage portion 1808 including a hard disk and the like; and a communication section 1809 including a network interface card such as a LAN card, a modem, or the like. The communication section 1809 performs communication processing via a network such as the internet. A driver 1810 is also connected to the I/O interface 1805 as needed. A removable medium 1811 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, or the like is mounted on the drive 1810 as necessary, so that a computer program read out therefrom is mounted in the storage portion 1808 as necessary.

In particular, according to an embodiment of the present invention, the processes described above with reference to the flowcharts may be implemented as computer software programs. For example, embodiments of the invention include a computer program product comprising a computer program embodied on a computer-readable medium, the computer program comprising program code for performing the method illustrated in the flow chart. In such embodiments, the computer program may be downloaded and installed from a network via the communication portion 1809, and/or installed from the removable media 1811. The computer program executes the above-described functions defined in the system of the present application when executed by the Central Processing Unit (CPU) 1801.

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

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

The units described in the embodiments of the present invention may be implemented by software, or may be implemented by hardware, and the described units may also be disposed in a processor. Wherein the names of the elements do not in some way constitute a limitation on the elements themselves.

As another aspect, the present application also provides a computer-readable medium, which may be contained in the electronic device described in the above embodiments; or may exist separately without being assembled into the electronic device. The computer readable medium carries one or more programs which, when executed by an electronic device, cause the electronic device to implement the semantic recognition method as described in the above embodiments.

For example, the electronic device may implement the following as shown in fig. 1: step S110, acquiring input information of a target object; step S120, extracting a target slot value of a target slot in the input information through a slot extraction model; the slot position extraction model comprises an input layer, a sharing layer and an output layer which are sequentially connected, wherein the output layer comprises n parallel sub-output layers, and each sub-output layer is used for outputting a slot value prediction result of a slot position; n is a positive integer of 2 or more.

It should be noted that although in the above detailed description several modules of a device or apparatus for action execution are mentioned, such division is not mandatory. Indeed, the features and functionality of two or more of the modules described above may be embodied in one module according to embodiments of the invention. Conversely, the features and functions of one module described above may be further divided into embodiments by a plurality of modules.

Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein may be implemented by software, or by software in combination with necessary hardware. Therefore, the technical solution according to the embodiment of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a usb disk, a removable hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the embodiment of the present invention.

Other embodiments of the invention will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the invention and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the invention being indicated by the following claims.

It will be understood that the invention is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the invention is limited only by the appended claims.

Claims

1. A method of semantic identification, comprising:

acquiring input information of a target object;

extracting a target slot value of a target slot in the input information through a slot extraction model;

the slot position extraction model comprises an input layer, a sharing layer and an output layer which are sequentially connected, wherein the output layer comprises n parallel sub-output layers, and each sub-output layer is used for outputting a slot value prediction result of a slot position; n is a positive integer of 2 or more.

2. The semantic recognition method according to claim 1, wherein the input information comprises input speech information and/or input text information.

3. The semantic recognition method of claim 1, wherein the target slot comprises any one or more of a product word slot, a brand word slot, a modifier slot, a gender slot, a price slot, a query range slot, and a channel number slot.

4. The semantic recognition method according to claim 1, further comprising:

n kinds of slots of the target product class are predefined.

5. The semantic recognition method according to claim 1, further comprising:

generating a product word bank, a brand word bank and a brand product word pairing word bank;

and establishing a knowledge graph based on the product word bank, the brand word bank and the brand product word pairing word bank.

6. The semantic recognition method according to claim 5, further comprising:

retrieving recommendation information from the knowledge graph according to a target slot value of the target slot position;

and sending the recommendation information to the target object.

7. The semantic recognition method according to claim 1, further comprising:

and determining a target business scene corresponding to the input information through a business scene classification model.

8. The semantic identification method according to claim 7, wherein the target business scenario comprises any one or more of a commodity query scenario, an after-sales service scenario, an ambiguous offer query scenario, a specific commodity offer query scenario, an order query scenario, a total station express scenario, and an unknown scenario.

9. The semantic recognition method according to claim 1, further comprising:

acquiring portrait information of the target object;

the portrait information includes any one or more of purchasing power level, member level, category preference information, and sex information.

10. The semantic identification method of claim 1, wherein the target slot comprises a product word slot and a brand word slot, the method further comprising:

and if the target slot value of the product word slot position in the input information is extracted, predicting the target slot value of the brand word slot position of the input information.

11. The semantic recognition method according to claim 1, further comprising:

negative semantics in the input information are identified.

12. The semantic recognition method according to claim 1, further comprising: establishing a wrongly written word library;

and automatically correcting the wrongly written characters of the input information based on the wrongly written character lexicon.

13. A semantic recognition apparatus, comprising:

an input information acquisition module configured to acquire input information of a target object;

the slot position extraction model is configured to extract a target slot value of a target slot position in the input information through the slot position extraction model;

14. An electronic device, comprising: a processor; and a memory having computer readable instructions stored thereon which, when executed by the processor, implement the semantic identification method of any one of claims 1 to 12.

15. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the semantic recognition method according to any one of claims 1 to 12.