WO2021139278A1 - 一种智能面试方法、装置及终端设备 - Google Patents

一种智能面试方法、装置及终端设备 Download PDF

Info

Publication number
WO2021139278A1
WO2021139278A1 PCT/CN2020/119298 CN2020119298W WO2021139278A1 WO 2021139278 A1 WO2021139278 A1 WO 2021139278A1 CN 2020119298 W CN2020119298 W CN 2020119298W WO 2021139278 A1 WO2021139278 A1 WO 2021139278A1
Authority
WO
WIPO (PCT)
Prior art keywords
sentence
reply
information
vector information
interview
Prior art date
Application number
PCT/CN2020/119298
Other languages
English (en)
French (fr)
Inventor
邓悦
郑立颖
徐亮
Original Assignee
平安科技(深圳)有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技(深圳)有限公司 filed Critical 平安科技(深圳)有限公司
Publication of WO2021139278A1 publication Critical patent/WO2021139278A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00Handling natural language data
    • G06F40/20Natural language analysis
    • G06F40/205Parsing
    • G06F40/211Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/27Replication, distribution or synchronisation of data between databases or within a distributed database system; Distributed database system architectures therefor
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00Handling natural language data
    • G06F40/20Natural language analysis
    • G06F40/279Recognition of textual entities
    • G06F40/289Phrasal analysis, e.g. finite state techniques or chunking
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00Administration; Management
    • G06Q10/06Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063Operations research, analysis or management
    • G06Q10/0639Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00Administration; Management
    • G06Q10/10Office automation; Time management
    • G06Q10/105Human resources
    • G06Q10/1053Employment or hiring

Definitions

  • This application relates to the field of artificial intelligence technology, in particular to an intelligent interview method, device and terminal equipment.
  • One of the purposes of the embodiments of this application is to provide a smart interview method, device, and terminal equipment, which aims to solve the problem of low reliability of interview results due to inaccurate relationship extraction when conducting smart interviews in the prior art .
  • the first aspect of the embodiments of this application provides a smart interview method, including:
  • response information of the candidate during the interview where the response information includes multiple response sentences
  • an interview question for the candidate is generated.
  • the second aspect of the embodiments of the present application provides a smart interview device, including:
  • the obtaining module is used to obtain the reply information of the candidate in the interview process, the reply information includes a plurality of reply sentences;
  • the conversion module is configured to use a preset language model to convert the multiple reply sentences into corresponding sentence vector information
  • the determining module is configured to determine the sentence set vector information corresponding to the reply information according to the sentence vector information of the multiple reply sentences;
  • a calculation module configured to use the sentence set vector information to calculate the relationship probability between multiple entities included in the reply message
  • An extraction module configured to extract target relationship information from the multiple entities according to the relationship probability
  • a generating module is used to generate interview questions for the candidate based on the target relationship information.
  • the third aspect of the embodiments of the present application provides a terminal device, including: a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor executes the computer program When realized:
  • response information of the candidate during the interview where the response information includes multiple response sentences
  • an interview question for the candidate is generated.
  • the fourth aspect of the embodiments of the present application provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following is achieved:
  • response information of the candidate during the interview where the response information includes multiple response sentences
  • an interview question for the candidate is generated.
  • the fifth aspect of the embodiments of the present application also provides a computer program product, which when the computer program product runs on a terminal device, enables the terminal device to execute:
  • response information of the candidate during the interview where the response information includes multiple response sentences
  • an interview question for the candidate is generated.
  • the embodiments of the present application include the following advantages:
  • FIG. 1 is a schematic diagram of a step flow diagram of an intelligent interview method according to an embodiment of the present application
  • Fig. 2 is a schematic diagram of the flow of steps of another smart interview method according to an embodiment of the present application
  • FIG. 3 is a schematic diagram of a smart interview device according to an embodiment of the present application.
  • Fig. 4 is a schematic diagram of a terminal device according to an embodiment of the present application.
  • FIG. 1 there is shown a schematic flow chart of the steps of a smart interview method according to an embodiment of the present application, which may specifically include the following steps:
  • S101 Obtain response information of a candidate during the interview process, where the response information includes multiple response sentences;
  • this method can be applied to intelligent interviews, that is, interacting with interview candidates through machines such as terminal devices to complete the interview evaluation of the candidates.
  • the terminal device in this embodiment may be an electronic device such as a mobile phone, a tablet computer, or a personal computer, and this embodiment does not limit the specific type of the terminal device.
  • the candidate's reply information may be inputted to the terminal device in the form of text by the candidate through the input device provided by the terminal device during the interview; or the terminal device may also interact through voice Candidates are interviewed in the form of interviews, by collecting the voice information of the candidates during the interview and converting the voice information into text content, so as to obtain the candidate's response information, which is not limited in this embodiment.
  • the candidate's reply information may be the reply information for a certain question raised by the terminal device.
  • the terminal device can require candidates to introduce themselves first.
  • the content of the candidate's self-introduction is the response information that needs to be obtained, and the terminal device can process the response information, so that the entire interview process can proceed smoothly.
  • the candidate's reply message may include multiple reply sentences. After converting each reply sentence, the corresponding sentence vector information can be obtained, and the sentence vector information is the sentence representation of the corresponding reply sentence.
  • a pre-trained language model may be used to perform vector conversion on each reply sentence to obtain a corresponding sentence representation.
  • the byte pair encoder (BPE) function can be preset in the language model, and the word in each reply sentence can be encoded using the byte pair encoding method to obtain the word vector of each word ; Then the word vector of each word in each sentence is added to the position vector corresponding to the position of the word, so as to obtain the sentence vector information that characterizes the sentence.
  • a language model based on a natural language processing (Natural Language Processing, NLP) transformer can also be used to convert each reply sentence.
  • NLP transformer includes an encoder and a decoder.
  • the encoder and decoder can be used to characterize the sentence vector of the sentence. This embodiment does not limit how to convert the reply sentence.
  • S103 Determine the sentence set vector information corresponding to the reply information according to the sentence vector information of the multiple reply sentences;
  • each reply sentence can only be used to characterize the content of the current sentence, and each reply of a candidate is often a detailed elaboration on a certain topic, and the focus of each sentence may be has a difference.
  • the sentence set vector information of the entire reply message can also be determined based on all the sentence vector information, and the content of the entire reply message can be represented by the sentence set vector information.
  • multiple entities included in the reply message may be obtained by segmenting each reply sentence and determining the part of speech of each word after the segmentation. For example, each word with a noun part of speech can be marked as an entity. These entities will be used as the basis for the follow-up of the candidates, that is, the corresponding interview questions will be generated through these entities.
  • the candidate needs to be asked for follow-up, or the candidate needs to conduct an in-depth analysis of a topic, which is often the content that the candidate mentions many times or focuses on during the interview. Therefore, in this embodiment, it is possible to determine which content needs to be focused on by determining the relationship probability between each entity.
  • the relationship probability between each entity can be output.
  • a group of entity pairs with the largest probability value can be extracted as the target relationship, or multiple groups of entity pairs with a probability value exceeding a certain threshold can be extracted as target relationship information as the input sentence of the terminal device.
  • Extracting some entities according to the relationship probability as the basis for questioning in subsequent interviews can reduce the amount of data processed by the machine and improve the questioning efficiency of the terminal device during the interview.
  • S106 Based on the target relationship information, generate an interview question for the candidate.
  • the terminal device After receiving the above-mentioned target relationship information, the terminal device can determine the next interview question based on the relationship and continue to interview the candidate.
  • FIG. 2 there is shown a schematic diagram of a step flow diagram of another smart interview method according to an embodiment of the present application, which may specifically include the following steps:
  • This method can be applied to an intelligent interview, by using a terminal device as an artificial intelligence (AI) interviewer, interacting with the interview candidate, completing the interview evaluation of the candidate, and improving the efficiency of the interview.
  • AI artificial intelligence
  • the candidate can directly use voice to communicate with the AI interviewer.
  • candidates use mobile phones, tablets or personal computers to conduct interviews.
  • AI interviewers can convert the voice information into text content, and then understand the specific meaning of the text content based on natural language processing technology , And ask the candidates questions on this basis.
  • each reply of the candidate may include multiple reply sentences.
  • the AI interviewer can process each response sentence separately to identify the key content of the candidate's entire response message.
  • the AI interviewer can first identify the entity in each reply sentence and generate a sequence of entities to be processed.
  • entity in each reply sentence may refer to each word with noun part of speech in the reply sentence.
  • each entity can correspond to a word with word order number in the preset corpus.
  • the entity sequence of the reply sentence can be obtained. .
  • a language model based on an NLP transformer decoder can be used to convert a candidate's reply sentence into a responsive sentence vector.
  • the above-mentioned language model based on the NLP transformer decoder may be a transformer decoder model with a shielded multi-head self-attention mechanism based on position feedforward operation.
  • the NLP transformer decoder in this embodiment is different from the original NLP transformer that only decodes, but includes a shielded multi-head self-attention mechanism based on position feedforward operation, which can be based on a given input on multiple NLP transformer blocks Characterize repeated coding. And because there is no encoder block, the NLP transformer decoder does not contain any unshielded self-attention mechanism.
  • the NLP transformer decoder can be generated by adopting the following encoding methods:
  • T is a matrix sentence corresponding hot code one-hot vector composition
  • W e is a marker embedded in a matrix
  • W p is the position of the embedded matrix
  • L is the number of transformer blocks
  • h l is a state transformer block l-th layer.
  • the NLP transformer Since the NLP transformer does not have the implicit concept of marking position, the first layer of NLP transformer will add the position embedding to be learned e p ⁇ Rd to each marked embedding at position p of the input sequence on.
  • the self-attention architecture allows the output state block to be expressed through all input states h l-1 This is important for effectively modeling remote dependencies.
  • the NLP transformer decoder in this embodiment also needs to limit self-attention, so that the model only needs to pay attention to the following text of the current mark, and does not need to pay attention to the above of the current mark.
  • the above-mentioned mark is the entity in each reply sentence.
  • S204 Generate sentence vector information of the target reply sentence according to the probability distribution of each entity in the target reply sentence;
  • the objective function can be set to maximize the log-likelihood function:
  • k is the context window considered for predicting the next marker c i through the conditional probability P.
  • the probability distribution of each entity can be calculated as follows:
  • h L is a state sequence after the last layer of the transformer L, W e is embedded in a matrix, ⁇ is a model parameter optimization lowered by stochastic gradient obtained.
  • the sentence vector information corresponding to each reply sentence can be composed, that is, the sentence representation.
  • S205 Determine the weight value of the sentence vector information of each reply sentence, and perform a weighted summation on the sentence vector information of each reply sentence according to the weight value to obtain the sentence set vector information corresponding to the reply information;
  • the sentence collection representation of the entire piece of reply information can be obtained, that is, the sentence collection vector information corresponding to the candidate's reply information.
  • the ratio of the sentence vector information of each reply sentence to the sum of the sentence vector information of all reply sentences can be calculated separately, and the The ratio is used as the weight value of the sentence vector information of the corresponding reply sentence.
  • weight value of the sentence vector information of each reply sentence can be expressed by the following formula:
  • ⁇ i is the weight value of the sentence vector information of the i-th reply sentence
  • exp(s i r) is the sentence vector information of the i-th reply sentence.
  • the structure of the NLP transformer can also be extended. That is, the model can be pre-trained with equation (1) as the goal first, and then the language model can be fine-tuned for the task of relation extraction.
  • tags x i [x 1 ,...,x m ], head i and tail i are the positions of the two entities in the relationship in the marked entity sequence, and r i is the relationship corresponding to remote supervision label. Due to noisy annotations, the response variable trained with label r i is unreliable. Instead, apply relational classification to bag-level text, representing each entity pair as a collection among them It is composed of the entity pairs of all sentences.
  • the final state is used to represent the last state of h L Represents the sentence vector information s i of the reply sentence
  • S206 Perform linear transformation and logistic regression softmax transformation on the sentence set vector information in sequence to obtain the mutual relationship probabilities of multiple entities included in the reply information;
  • the sentence set vector information used to characterize the entire piece of reply information can be linearly transformed and then subjected to the softmax transformation, and the output distribution P(l) on the relationship label can be obtained:
  • W r is the representation matrix of relation r, Is the deviation vector.
  • the above output distribution is the probability of the relationship between each entity in the candidate's answer content.
  • each parameter in the output distribution can also be fine-tuned.
  • the goal of fine-tuning is to maximize the likelihood function as follows:
  • the language model can be used as an auxiliary target to improve the versatility and convergence speed of the model. Therefore, combining the above formulas (1) and (2), the final objective function can be obtained:
  • the scalar value ⁇ is the weight of the language model objective function in the fine-tuning process.
  • S207 Extract one or more entity pairs whose relationship probability exceeds a preset threshold as target relationship information
  • a group of entity pairs with the highest probability can be extracted as the target relationship, or multiple groups of entity pairs with a probability exceeding a certain threshold can be extracted as the target relationship, as the input sentence of the AI interviewer.
  • the AI interviewer can determine the next interview question based on the relationship and continue to interview the candidate.
  • an interview evaluation report for the candidate can be generated, and the above report can be uploaded to the Blockchain to ensure its safety The transparency and fairness of the evaluation results.
  • the evaluation report can be downloaded from the blockchain through the user device to verify whether the report has been tampered with.
  • the blockchain referred to in this embodiment is a new application mode of computer technology such as distributed data storage, point-to-point transmission, consensus mechanism, and encryption algorithm.
  • the blockchain is essentially a decentralized database, which is a series of data blocks associated with cryptographic methods. Each data block contains a batch of network transaction information to verify the validity of the information. (Anti-counterfeiting) and generate the next block.
  • the blockchain can include the underlying platform of the blockchain, the platform product service layer, and the application service layer.
  • the AI interviewer can quickly extract the important content of the candidate's answer. Since the amount of data used for machine understanding after extracting the content is less, the AI interviewer can quickly give necessary and reasonable follow-up questions. In practical applications, because the judgment is more accurate, the response speed of the hardware is also improved, which not only saves the hardware space, but also improves the running speed and the interview experience of the candidates.
  • FIG. 3 a schematic diagram of an intelligent interview device according to an embodiment of the present application is shown, which may specifically include the following modules:
  • the obtaining module 301 is used to obtain the reply information of the candidate during the interview, and the reply information includes a plurality of reply sentences;
  • the conversion module 302 is configured to use a preset language model to respectively convert the multiple reply sentences into corresponding sentence vector information
  • the determining module 303 is configured to determine the sentence set vector information corresponding to the reply information according to the sentence vector information of the multiple reply sentences;
  • the calculation module 304 is configured to use the sentence set vector information to calculate the relationship probability between multiple entities included in the reply message;
  • the extraction module 305 is configured to extract target relationship information from the multiple entities according to the relationship probability
  • the generating module 306 is configured to generate interview questions for the candidate based on the target relationship information.
  • the conversion module 302 may specifically include the following sub-modules:
  • the entity sequence generation sub-module is used to identify multiple entities in the target reply sentence, and generate the entity sequence to be processed according to the multiple entities, and the target reply sentence is any one of the multiple reply sentences;
  • the probability distribution calculation sub-module is used to input the sequence of entities to be processed into a preset language model to obtain the probability distribution of each entity in the target reply sentence.
  • the language model is based on a position feedforward operation.
  • the sentence vector information generating sub-module is used to generate sentence vector information of the target reply sentence according to the probability distribution of each entity in the target reply sentence.
  • the transformer decoder with shielded multi-head self-attention mechanism based on position feedforward operation is generated by adopting the following encoding method:
  • T is a matrix sentence corresponding hot code one-hot vector composition
  • W e is a marker embedded in a matrix
  • W p is the position of the embedded matrix
  • L is the number of transformer blocks
  • h l is a state transformer block l-th layer.
  • the determining module 303 may specifically include the following sub-modules:
  • the weight value determination sub-module is used to determine the weight value of the sentence vector information of each reply sentence
  • the sentence set vector information generating submodule is used to perform a weighted summation of the sentence vector information of each reply sentence according to the weight value to obtain the sentence set vector information corresponding to the reply information.
  • the weight value determining submodule may specifically include the following units:
  • the weight value calculation unit is used to calculate the ratio of the sentence vector information of each reply sentence to the sum of the sentence vector information of all reply sentences, and use the ratio as the weight value of the sentence vector information of the corresponding reply sentence.
  • the calculation module 304 may specifically include the following sub-modules:
  • the relationship probability calculation sub-module is used to sequentially perform linear transformation and logistic regression softmax transformation on the sentence set vector information to obtain the relationship probabilities between multiple entities contained in the reply information.
  • the extraction module 305 may specifically include the following sub-modules:
  • the target relationship information extraction submodule is used to extract one or more entity pairs whose relationship probability exceeds a preset threshold as target relationship information.
  • the device may further include the following modules:
  • the interview evaluation report generation module is used to generate an interview evaluation report for the candidate according to the candidate's reply information to multiple interview questions;
  • the interview evaluation report upload module is used to upload the interview evaluation report to the blockchain.
  • the description is relatively simple, and for related parts, please refer to the description of the method embodiment part.
  • the terminal device 400 of this embodiment includes: a processor 410, a memory 420, and a computer program 421 stored in the memory 420 and running on the processor 410.
  • the processor 410 executes the computer program 421
  • the steps in each embodiment of the smart interview method described above are implemented, for example, steps S101 to S106 shown in FIG. 1.
  • the processor 410 executes the computer program 421
  • the functions of the modules/units in the foregoing device embodiments, such as the functions of the modules 301 to 306 shown in FIG. 3, are realized.
  • the computer program 421 may be divided into one or more modules/units, and the one or more modules/units are stored in the memory 420 and executed by the processor 410 to complete This application.
  • the one or more modules/units may be a series of computer program instruction segments capable of completing specific functions, and the instruction segments may be used to describe the execution process of the computer program 421 in the terminal device 400.
  • the computer program 421 can be divided into an acquisition module, a conversion module, a determination module, a calculation module, an extraction module, and a generation module.
  • the specific functions of each module are as follows:
  • the obtaining module is used to obtain the reply information of the candidate in the interview process, the reply information includes a plurality of reply sentences;
  • the conversion module is configured to use a preset language model to convert the multiple reply sentences into corresponding sentence vector information
  • the determining module is configured to determine the sentence set vector information corresponding to the reply information according to the sentence vector information of the multiple reply sentences;
  • a calculation module configured to use the sentence set vector information to calculate the relationship probability between multiple entities included in the reply message
  • An extraction module configured to extract target relationship information from the multiple entities according to the relationship probability
  • a generating module is used to generate interview questions for the candidate based on the target relationship information.
  • Said using a preset language model to respectively convert the multiple reply sentences into corresponding sentence vector information includes:
  • the language model is based on a position feedforward operation with a masked multi-head self-attention mechanism Transformer decoder
  • the sentence vector information of the target reply sentence is generated.
  • the processor when the processor executes the computer program, it also realizes the generation of a transformer decoder, and the transformer decoder with a shielded multi-head self-attention mechanism based on position feedforward operation is generated by adopting the following encoding method :
  • T is a matrix sentence corresponding hot code one-hot vector composition
  • W e is a marker embedded in a matrix
  • W p is the position of the embedded matrix
  • L is the number of transformer blocks
  • h l is a state transformer block l-th layer.
  • the sentence vector information of each reply sentence is weighted and summed to obtain the sentence set vector information corresponding to the reply information.
  • the processor when the processor executes the computer program, it also realizes: separately calculating the ratio of the sentence vector information of each reply sentence to the sum of the sentence vector information of all reply sentences, and using the ratio as the corresponding The weight value of the sentence vector information of the reply sentence.
  • the linear transformation and the logistic regression softmax transformation are sequentially performed on the sentence set vector information to obtain the mutual relationship probabilities of the multiple entities included in the reply information.
  • One or more entity pairs whose relationship probability exceeds a preset threshold are extracted as target relationship information.
  • the embodiment of the present application also provides a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it realizes:
  • response information of the candidate during the interview where the response information includes multiple response sentences
  • an interview question for the candidate is generated.
  • the language model is based on a position feedforward operation with a masked multi-head self-attention mechanism Transformer decoder
  • the sentence vector information of the target reply sentence is generated.
  • a transformer decoder when the computer program is executed by the processor, a transformer decoder is also realized, and the transformer decoder with a shielded multi-head self-attention mechanism based on position feedforward operation is generated by adopting the following encoding method:
  • T is a matrix sentence corresponding hot code one-hot vector composition
  • W e is a marker embedded in a matrix
  • W p is the position of the embedded matrix
  • L is the number of transformer blocks
  • h l is a state transformer block l-th layer.
  • the sentence vector information of each reply sentence is weighted and summed to obtain the sentence set vector information corresponding to the reply information.
  • the ratio of the sentence vector information of each reply sentence to the sum of the sentence vector information of all reply sentences is calculated separately, and the ratio is used as the weight value of the sentence vector information of the corresponding reply sentence.
  • the linear transformation and the logistic regression softmax transformation are sequentially performed on the sentence set vector information to obtain the mutual relationship probabilities of the multiple entities included in the reply information.
  • One or more entity pairs whose relationship probability exceeds a preset threshold are extracted as target relationship information.
  • the terminal device 400 may be a computing device such as a desktop computer, a notebook, or a palmtop computer.
  • the terminal device 400 may include, but is not limited to, a processor 410 and a memory 420.
  • FIG. 4 is only an example of the terminal device 400, and does not constitute a limitation on the terminal device 400. It may include more or less components than shown in the figure, or combine certain components, or different components.
  • the terminal device 400 may also include input and output devices, network access devices, buses, and so on.
  • the processor 410 may be a central processing unit (Central Processing Unit, CPU), or other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (ASIC), Ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.
  • the memory 420 may be an internal storage unit of the terminal device 400, such as a hard disk or a memory of the terminal device 400.
  • the memory 420 may also be an external storage device of the terminal device 400, such as a plug-in hard disk equipped on the terminal device 400, a smart memory card (Smart Media Card, SMC), or a Secure Digital (SD). Card, Flash Card, etc.
  • the memory 420 may also include both an internal storage unit of the terminal device 400 and an external storage device.
  • the memory 420 is used to store the computer program 421 and other programs and data required by the terminal device 400.
  • the memory 420 can also be used to temporarily store data that has been output or will be output.
  • Non-volatile memory may include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • ROM read only memory
  • PROM programmable ROM
  • EPROM electrically programmable ROM
  • EEPROM electrically erasable programmable ROM
  • Volatile memory may include random access memory (RAM) or external cache memory.
  • RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain Channel (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

Abstract

一种智能面试方法、装置及终端设备,所述方法包括:获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句(S101);采用预设的语言模型,分别将多个回复语句转换为对应的句向量信息(S102);根据多个回复语句的句向量信息,确定回复信息对应的语句集合向量信息(S103);采用语句集合向量信息,计算回复信息中包含的多个实体相互间的关系概率(S104);根据关系概率,从多个实体中提取出目标关系信息(S105);基于目标关系信息,生成针对候选人的面试题目(S106)。可以快速地抽取出候选人回答内容中的重要部分,方便人工智能面试官给出必要和合理的追问,进而生成面试评价报告。此外,面试评价报告可以上传至区块链中,以保证其安全性和公正透明性。

Description

一种智能面试方法、装置及终端设备
本申请要求于2020年05月28日在中国专利局提交的、申请号为202010466693.9、发明名称为“一种智能面试方法、装置及终端设备”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请涉及人工智能技术领域,具体涉及一种智能面试方法、装置及终端设备。
背景技术
招聘面试是一项费时费力的工作。尤其是在招聘量大的时候,由于等待面试的候选人众多但面试官却有限,面试官通常需要连续进行多个场次的面试,严重影响面试效率。为了节省面试时间,提高招聘效率,智能面试应运而生。智能面试可以通过机器与候选人进行交互,自动完成对候选人的评价。
在进行智能面试时,为了使机器能够准确地对候选人做出提问,需要对候选人针对上一个问题的回答进行处理,抽取候选人前面回答的重要部分作为参考依据。关系抽取在这个环节中起着重要作用。
相关技术中可以通过两种方式来实现关系抽取。一种是使用文本中的概念与知识库中对应的关系实例,启发式地生成标记数据,然后再采用标记数据训练关系抽取模型。但是,按照这种方式生成的标记数据会产生噪声标记,得到错误的判断和不完整的知识库信息,使机器做出不准确的提问。另一种方式可以基于预先提供的语义和句法知识进行多实例学习,并根据学习结果来指导模型的训练。但是,发明人意识到,根据预先提供的语义和句法知识训练出的模型只能识别与提供的语义和句法知识对应的某一部分或某一类的关系,在其他类的关系识别上表现较差,适用范围较窄,也无法广泛地应用于各种不同场景下的智能面试。
技术问题
本申请实施例的目的之一在于:提供一种智能面试方法、装置及终端设备,旨在解决现有技术中在进行智能面试时,由于关系抽取不准确导致面试结果可信度较低的问题。
技术解决方案
为解决上述技术问题,本申请实施例采用的技术方案是:
本申请实施例的第一方面提供了一种智能面试方法,包括:
获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
根据所述关系概率,从所述多个实体中提取出目标关系信息;
基于所述目标关系信息,生成针对所述候选人的面试题目。
本申请实施例的第二方面提供了一种智能面试装置,包括:
获取模块,用于获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
转换模块,用于采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
确定模块,用于根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
计算模块,用于采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
提取模块,用于根据所述关系概率,从所述多个实体中提取出目标关系信息;
生成模块,用于基于所述目标关系信息,生成针对所述候选人的面试题目。
本申请实施例的第三方面提供了一种终端设备,包括:存储器、处理器以及存储在所 述存储器中并可在所述处理器上运行的计算机程序,所述处理器执行所述计算机程序时实现:
获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
根据所述关系概率,从所述多个实体中提取出目标关系信息;
基于所述目标关系信息,生成针对所述候选人的面试题目。
本申请实施例的第四方面提供了一种计算机可读存储介质,所述计算机可读存储介质存储有计算机程序,所述计算机程序被处理器执行时实现:
获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
根据所述关系概率,从所述多个实体中提取出目标关系信息;
基于所述目标关系信息,生成针对所述候选人的面试题目。
本申请实施例的第五方面还提供了一种计算机程序产品,当所述计算机程序产品在终端设备上运行时,使得所述终端设备执行时实现:
获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
根据所述关系概率,从所述多个实体中提取出目标关系信息;
基于所述目标关系信息,生成针对所述候选人的面试题目。
有益效果
与现有技术相比,本申请实施例包括以下优点:
本申请实施例,通过获取候选人在面试过程中的回复信息并采用预设的语言模型,将回复信息中的多个回复语句转换为对应的句向量信息,从而可以基于各个回复语句的句向量信息,计算出整段回复信息的语句集合向量信息,便于终端设备根据语句集合向量信息快速地抽取出候选人回复信息中的重要部分,减少后续机器理解的数据量,提高终端设备给出必要和合理追问的速度。在实际的面试应用中,根据候选人回复信息中的重要部分内容进行针对性的处理,使得终端设备对候选人的评价更准确。
附图说明
为了更清楚地说明本申请实施例中的技术方案,下面将对实施例或示范性技术描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其它的附图。
图1是本申请一个实施例的一种智能面试方法的步骤流程示意图;
图2是本申请一个实施例的另一种智能面试方法的步骤流程示意图;
图3是本申请一个实施例的一种智能面试装置的示意图;
图4是本申请一个实施例的一种终端设备的示意图。
本发明的实施方式
以下描述中,为了说明而不是为了限定,提出了诸如特定系统结构、技术之类的具体细节,以便透彻理解本申请实施例。然而,本领域技术人员应当清楚,在没有这些具体细节的其他实施例中也可以实现本申请。在其他情况中,省略对众所周知的系统、装置、电路以及方法的详细说明,以免不必要的细节妨碍本申请的描述。
下面通过具体实施例来说明本申请的技术方案。
参照图1,示出了本申请一个实施例的一种智能面试方法的步骤流程示意图,具体可以包括如下步骤:
S101、获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
需要说明的是,本方法可以应用于智能面试中,即通过终端设备等机器与面试候选人进行交互,完成对候选人的面试评价。
本实施例中的终端设备可以是手机、平板电脑、个人计算机等电子设备,本实施例对终端设备的具体类型不作限定。
在本申请实施例中,候选人的回复信息可以是在面试过程中,由候选人通过终端设备提供的输入装置,以文字的形式输入至终端设备的;或者,终端设备也可以通过语音交互的形式对候选人进行面试,通过采集候选人在面试过程中的语音信息,并将语音信息转换为文本内容,从而得到候选人的回复信息,本实施例对此不作限定。
在本申请实施例中,候选人的回复信息可以是针对终端设备提出的某一个问题所进行回复的信息。在面试开始时,终端设备可以要求候选人首先进行自我介绍。候选人自我介绍的内容,即是需要获取的回复信息,终端设备可以针对该回复信息进行处理,使整个面试过程顺畅地进行下去。
S102、采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
通常,候选人的回复信息可能包括多个回复语句,在对每个回复语句进行转换后,可以得到相应的句向量信息,该句向量信息也就是对应回复语句的句子表征。
在本申请实施例中,可以采用预先训练得到的语言模型对每个回复语句进行向量转换,得到相应的句子表征。
在具体实现中,可以在语言模型中预置字节对编码(byte pair encoder,BPE)功能,使用字节对编码的方式对每个回复语句中的词语进行编码,得到每个词语的词向量;然后将各个句子中每个词语的词向量与该词语所在位置对应的位置向量相加,从而得到表征该句子的句向量信息。
或者,也可以采用基于自然语言处理(Natural Language Processing,NLP)变压器的语言模型来对每个回复语句进行转换。通常,NLP变压器包括编码器和解码器两部分,对于给定的语句序列,通过编码器和解码器的处理,可以用于表征该语句的句向量。本实施例对如何回复语句进行转换不作限定。
S103、根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
通常,根据每个回复语句得到的句向量信息仅可以用于表征当前句子的内容,而候选人的每一次回复往往是针对某一个话题而进行的详细阐述,各个句子所体现出的关注点可能存在差异。
因此,在分别得到每个回复语句的句向量信息后,还可以根据全部的句向量信息,确定出整个回复信息的语句集合向量信息,通过语句集合向量信息表征整个回复信息的内容。
S104、采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
在本申请实施例中,回复信息中包含的多个实体可以是对每个回复语句进行分词并通过确定分词后的各个词语的词性来得到的。例如,可以将具有名词词性的各个词语标记为实体。这些实体将被作为后续对候选人进行追问的基础,即通过这些实体生成相应面试题目。
通常,需要对候选人进行追问,或者需要请候选人对某一话题进行深入分析的,往往是候选人在面试过程中多次或重点提及的内容。因此,本实施例可以通过对各个实体之间的关系概率来确定哪些是需要重点关注的内容。
在具体实现中,可以基于前述语言模型,输出各个实体之间的关系概率。
S105、根据所述关系概率,从所述多个实体中提取出目标关系信息;
在本申请实施例中,可以提取概率值最大的一组实体对作为目标关系,或者提取概率值超过一定阈值的多组实体对作为目标关系信息,作为终端设备的输入语句。
根据关系概率的大小提取部分实体作为后续面试的提问基础,可以减少机器处理的数据量,提高终端设备在面试过程中的提问效率。
S106、基于所述目标关系信息,生成针对所述候选人的面试题目。
终端设备在接收到上述目标关系信息后,可以根据该关系,确定下一个面试问题,继续对候选人进行面试。
在本申请实施例中,通过获取候选人在面试过程中的回复信息并采用预设的语言模型,将回复信息中的多个回复语句转换为对应的句向量信息,从而可以基于各个回复语句的句向量信息,计算出整段回复信息的语句集合向量信息,便于终端设备根据语句集合向量信息快速地抽取出候选人回复信息中的重要部分,减少后续机器理解的数据量,提高终端设备给出必要和合理追问的速度。在实际的面试应用中,根据候选人回复信息中的重要部分内容进行针对性的处理,使得终端设备对候选人的评价更准确。
参照图2,示出了本申请一个实施例的另一种智能面试方法的步骤流程示意图,具体可以包括如下步骤:
S201、获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
本方法可以应用于智能面试中,通过使用终端设备作为人工智能(Artificial Intelligence,AI)面试官,与面试候选人进行交互,完成对候选人的面试评价,提高面试效率。
在本申请实施例中,候选人可以直接使用语音与AI面试官进行交流。例如,候选人使用手机、平板电脑或个人计算机进行面试,通过采集面试过程中的语音信息,AI面试官可以将语音信息转换为文本内容,然后基于自然语言处理技术理解文本内容所包含的具体意思,并在此基础上对候选人进行提问。
S202、识别目标回复语句中的多个实体,根据所述多个实体生成待处理的实体序列,所述目标回复语句为所述多个回复语句中的任意一个;
在本申请实施例中,候选人的每一次回复可能包含多个回复语句。AI面试官可以针对每个回复语句分别进行处理,从而识别出候选人整段回复信息中的重点内容。
在本申请实施例中,AI面试官可以首先识别出每个回复语句中的实体,生成待处理的实体序列。
需要说明的是,每个回复语句中的实体可以是指该回复语句中具有名词词性的各个词语。
在具体实现中,每个实体可以对应预设的语料库中的一个具有词序编号的词语,通过按照各个实体在回复语句中的先后顺序对对应的词序编号进行排列,可以得到该回复语句的实体序列。
S203、将所述待处理的实体序列输入预设的语言模型中,获得所述目标回复语句中每个实体的概率分布;
在本申请实施例中,可以使用基于NLP变压器解码器的语言模型来将候选人的回复语句转换为响应的句向量。上述基于NLP变压器解码器的语言模型可以是基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器模型。
本实施例中的NLP变压器解码器,不同于仅解码的原始NLP变压器,而是包括基于位置前馈操作的有屏蔽的多头自注意力机制,可以在多个NLP变压器块上基于给定的输入表征重复编码。而且因为没有编码器块,NLP变压器解码器不包含任何非屏蔽的自注意力机制。
在具体实现中,可以通过采用如下编码方式生成NLP变压器解码器:
h 0=TW e+W p
Figure PCTCN2020119298-appb-000001
其中,T是句子对应的独热码one-hot向量组成的矩阵,W e是标记嵌入矩阵,W p是位置嵌入矩阵,L是变压器块的数量,h l是第l层变压器块的状态。
由于NLP变压器没有标记位置这一隐含概念,因此第一层NLP变压器会将待学习的位置嵌入e p∈R d添加到输入序列位置p的每个标记嵌入
Figure PCTCN2020119298-appb-000002
上。自注意力的体系结构允许通过所有输入状态h l-1表示输出状态块
Figure PCTCN2020119298-appb-000003
这对有效地对远程依赖关系建模很重要。但是,本实施例中的NLP变压器解码器同时需要限制自注意力,使模型只需关注当前标记的下文,而无需关注当前标记的上文。上述标记即是各个回复语句中的实体。
S204、根据所述目标回复语句中每个实体的概率分布,生成所述目标回复语句的句向量信息;
在本申请实施例中,基于上述NLP变压器解码器模型,对于给定的实体序列
Figure PCTCN2020119298-appb-000004
可以设定目标函数为最大化对数似然函数:
Figure PCTCN2020119298-appb-000005
其中,k是考虑用于通过条件概率P预测下一个标记c i的上下文窗口。
使用上述NLP变压器解码器模型,可以计算出各个实体的概率分布如下:
Figure PCTCN2020119298-appb-000006
其中,h L是变压器最后一层L之后的状态序列,W e是嵌入矩阵,θ是通过随机梯度下降优化得到的模型参数。
通过输出每个实体的概率分布,可以组成每个回复语句对应的句向量信息,即句子表征。
S205、确定每个回复语句的句向量信息的权重值,根据所述权重值,对所述每个回复语句的句向量信息进行加权求和,得到所述回复信息对应的语句集合向量信息;
在本申请实施例中,可以通过汇总每个回复语句的句子表征,得到整段回复信息的语句集合表征,即与候选人回复信息相对应的语句集合向量信息。
在具体实现中,为了更清晰地表达每个回复语句与整段回复信息的关系,可以将分别计算每个回复语句的句向量信息与全部回复语句的句向量信息之和的比值,并将该比值作为对应的回复语句的句向量信息的权重值。
例如,每个回复语句的句向量信息的权重值可以采用如下公式表示:
Figure PCTCN2020119298-appb-000007
其中,α i为第i个回复语句的句向量信息的权重值,exp(s ir)为第i个回复语句的句向量信息。
在本申请实施例中,为了在使用基于上述NLP变压器生成的语言模型实现对远程监管数据集的多实例学习时更有效,还可以对NLP变压器进行结构扩展。即,可以首先以等式(1)为目标对模型进行预训练,然后针对关系提取任务对语言模型进行微调。
例如,令一个标记的数据集
Figure PCTCN2020119298-appb-000008
其中每个示例都由标记x i=[x 1,…,x m]组成,head i和tail i是关系中两个实体在标记的实体序列中的位置,r i是远距离监督对应的关系标签。由于有噪声注释,以标签r i为训练的响应变量是不可靠的。相反,将关系分类应用于袋级别的文本,将每个实体对表示为集合
Figure PCTCN2020119298-appb-000009
其中
Figure PCTCN2020119298-appb-000010
由所有句子的实体对组成。
通过将句子的实体序列馈入模型中,使用最终状态表征h L的上一个状态
Figure PCTCN2020119298-appb-000011
表示回复语句的句向量信息s i
然后,将每个回复语句的句向量信息进行加权求和,可以得到回复信息对应的语句集合向量信息:
Figure PCTCN2020119298-appb-000012
S206、依次对所述语句集合向量信息进行线性变换和逻辑回归softmax变换,获得所述回复信息中包含的多个实体相互间的关系概率;
在本申请实施例中,可以将用于表征整段回复信息的语句集合向量信息进行线性变换后再做softmax变换,可以得到关系标签上的输出分布P(l):
P(l|S,θ)=softmax(W rs+b)
其中,W r是关系r的表示矩阵,
Figure PCTCN2020119298-appb-000013
是偏差矢量。
上述输出分布即是候选人回答内容中各个实体之间的关系概率。
在本申请实施例中,为了提高模型输出的分布概率的准确性,还可以对输出分布中各个参数进行微调。
具体地,微调的目标是最大化如下似然函数:
Figure PCTCN2020119298-appb-000014
在微调时,可以将语言模型作为辅助目标,提高模型的通用性和收敛速度。因此,结合上述公式(1)和(2),可以得到最终的目标函数:
Figure PCTCN2020119298-appb-000015
其中,标量值λ是微调过程中语言模型目标函数的权重。
S207、提取所述关系概率超过预设阈值的一个或多个实体对,作为目标关系信息;
在本申请实施例中,可以提取概率最大的一组实体对作为目标关系,或者提取概率超过一定阈值的多组实体对作为目标关系,作为AI面试官的输入语句。AI面试官在接收到上述目标关系信息后,可以根据该关系,确定下一个面试问题,继续对候选人进行面试。
作为一个具体的示例,若面试官提问:“我们有一个新产品,你们可以通过讨论决定是否要发售这个新产品。”
候选人回答:“我认为新产品的临床试验时间还不够长,如果直接投入使用会为用户带来一定的风险,并且我们可能没有办法解决这些问题。所以我认为应该在临床试验的结果足够充分以后再投入销售”
通过对该候选人的回答进行编码,并得到相应的隐藏状态h L,然后对每句话加权得到整段文本的表征,然后输出每种实体的概率。例如<新产品,风险>(0.7),<新产品,问题>(0.2),……。那么,可以选择概率最大的关系,也就是<新产品,风险>作为目标关系信息,作为进一步面试的提问基础。
S208、基于所述目标关系信息,生成针对所述候选人的面试题目;
S209、根据所述候选人对多个面试题目的回复信息,生成针对所述候选人的面试评价报告;将所述面试评价报告上传至区块链中。
在本申请实施例中,根据候选人对AI面试官每个问题的回复,可以生成针对该候选人的面试评价报告,上述报告可以被上传至区块链(Block chain)中,以保证其安全性以及评价结果的透明公正性。
在后续过程中,例如需要对某个候选人的信息进行回溯时,可以通过用户设备从区块链中下载该评价报告,以便查证报告是否被篡改。
需要说明的是,本实施例中所指区块链是分布式数据存储、点对点传输、共识机制、加密算法等计算机技术的新型应用模式。区块链本质上是一个去中心化的数据库,是一串使用密码学方法相关联产生的数据块,每一个数据块中包含了一批次网络交易的信息,用于验证其信息的有效性(防伪)和生成下一个区块。区块链可以包括区块链底层平台、平台产品服务层以及应用服务层等。
在本申请实施例中,在智能面试过程中,通过采用基于远程监督NLP变压器的语言模型进行关系抽取,AI面试官可以迅速抽取候选人回答的重要内容。由于抽取内容后用于机器理解的数据量少了,所以AI面试官可以迅速给出必要和合理的追问。在实际应用中,因为判断更精准,硬件的应答速度也得到了提高,不仅节省了硬件空间,而且提高了运行速度和候选人的面试体验。
其次,通过抽取候选人的回答内容,可以对候选人回答中体现的关系与实际信息不一致的地方进行追问,进而可以更好地理解候选人的回答,有助于提供更有效的选拔候选人的依据,避免了面试官因为候选人的外貌,以及长时间面试造成的精力不足等因素而导致对候选人的表现产生错误判断的可能性,保证面试结果的准确性和可靠性。
需要说明的是,上述实施例中各步骤的序号的大小并不意味着执行顺序的先后,各过程的执行顺序应以其功能和内在逻辑确定,而不应对本申请实施例的实施过程构成任何限定。
参照图3,示出了本申请一个实施例的一种智能面试装置的示意图,具体可以包括如下模块:
获取模块301,用于获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
转换模块302,用于采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
确定模块303,用于根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
计算模块304,用于采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
提取模块305,用于根据所述关系概率,从所述多个实体中提取出目标关系信息;
生成模块306,用于基于所述目标关系信息,生成针对所述候选人的面试题目。
在本申请实施例中,所述转换模块302具体可以包括如下子模块:
实体序列生成子模块,用于识别目标回复语句中的多个实体,根据所述多个实体生成待处理的实体序列,所述目标回复语句为所述多个回复语句中的任意一个;
概率分布计算子模块,用于将所述待处理的实体序列输入预设的语言模型中,获得所述目标回复语句中每个实体的概率分布,所述语言模型为基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器;
句向量信息生成子模块,用于根据所述目标回复语句中每个实体的概率分布,生成所述目标回复语句的句向量信息。
在本申请实施例中,所述基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器通过采用如下编码方式生成:
h 0=TW e+W p
Figure PCTCN2020119298-appb-000016
其中,T是句子对应的独热码one-hot向量组成的矩阵,W e是标记嵌入矩阵,W p是位置嵌入矩阵,L是变压器块的数量,h l是第l层变压器块的状态。
在本申请实施例中,所述确定模块303具体可以包括如下子模块:
权重值确定子模块,用于确定每个回复语句的句向量信息的权重值;
语句集合向量信息生成子模块,用于根据所述权重值,对所述每个回复语句的句向量信息进行加权求和,得到所述回复信息对应的语句集合向量信息。
在本申请实施例中,所述权重值确定子模块具体可以包括如下单元:
权重值计算单元,用于分别计算每个回复语句的句向量信息与全部回复语句的句向量信息之和的比值,将所述比值作为对应的回复语句的句向量信息的权重值。
在本申请实施例中,所述计算模块304具体可以包括如下子模块:
关系概率计算子模块,用于依次对所述语句集合向量信息进行线性变换和逻辑回归softmax变换,获得所述回复信息中包含的多个实体相互间的关系概率。
在本申请实施例中,所述提取模块305具体可以包括如下子模块:
目标关系信息提取子模块,用于提取所述关系概率超过预设阈值的一个或多个实体对,作为目标关系信息。
在本申请实施例中,所述装置还可以包括如下模块:
面试评价报告生成模块,用于根据所述候选人对多个面试题目的回复信息,生成针对所述候选人的面试评价报告;
面试评价报告上传模块,用于将所述面试评价报告上传至区块链中。
对于装置实施例而言,由于其与方法实施例基本相似,所以描述得比较简单,相关之处参见方法实施例部分的说明即可。
参照图4,示出了本申请一个实施例的一种终端设备的示意图。如图4所示,本实施例的终端设备400包括:处理器410、存储器420以及存储在所述存储器420中并可在所述处理器410上运行的计算机程序421。所述处理器410执行所述计算机程序421时实现上述智能面试方法各个实施例中的步骤,例如图1所示的步骤S101至S106。或者,所述处理器410执行所述计算机程序421时实现上述各装置实施例中各模块/单元的功能,例如图3所示模块301至306的功能。
示例性的,所述计算机程序421可以被分割成一个或多个模块/单元,所述一个或者多个模块/单元被存储在所述存储器420中,并由所述处理器410执行,以完成本申请。所述一个或多个模块/单元可以是能够完成特定功能的一系列计算机程序指令段,该指令段可以用于描述所述计算机程序421在所述终端设备400中的执行过程。例如,所述计算机程序421可以被分割成获取模块、转换模块、确定模块、计算模块、提取模块和生成模块,各模块具体功能如下:
获取模块,用于获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
转换模块,用于采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
确定模块,用于根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
计算模块,用于采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
提取模块,用于根据所述关系概率,从所述多个实体中提取出目标关系信息;
生成模块,用于基于所述目标关系信息,生成针对所述候选人的面试题目。
在本申请实施例中,所述处理器执行所述计算机程序时还实现:
所述采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息,包括:
识别目标回复语句中的多个实体,根据所述多个实体生成待处理的实体序列,所述目标回复语句为所述多个回复语句中的任意一个;
将所述待处理的实体序列输入预设的语言模型中,获得所述目标回复语句中每个实体的概率分布,所述语言模型为基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器;
根据所述目标回复语句中每个实体的概率分布,生成所述目标回复语句的句向量信息。
在本申请实施例中,所述处理器执行所述计算机程序时还实现生成变压器解码器,所述基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器通过采用如下编码方式生成:
h 0=TW e+W p
Figure PCTCN2020119298-appb-000017
其中,T是句子对应的独热码one-hot向量组成的矩阵,W e是标记嵌入矩阵,W p是位置嵌入矩阵,L是变压器块的数量,h l是第l层变压器块的状态。
在本申请实施例中,所述处理器执行所述计算机程序时还实现:
确定每个回复语句的句向量信息的权重值;
根据所述权重值,对所述每个回复语句的句向量信息进行加权求和,得到所述回复信 息对应的语句集合向量信息。
在本申请实施例中,所述处理器执行所述计算机程序时还实现:分别计算每个回复语句的句向量信息与全部回复语句的句向量信息之和的比值,将所述比值作为对应的回复语句的句向量信息的权重值。
在本申请实施例中,所述处理器执行所述计算机程序时还实现:
依次对所述语句集合向量信息进行线性变换和逻辑回归softmax变换,获得所述回复信息中包含的多个实体相互间的关系概率。
在本申请实施例中,所述处理器执行所述计算机程序时还实现:
提取所述关系概率超过预设阈值的一个或多个实体对,作为目标关系信息。
在本申请实施例中,所述处理器执行所述计算机程序时还实现:
根据所述候选人对多个面试题目的回复信息,生成针对所述候选人的面试评价报告;
将所述面试评价报告上传至区块链中。
本申请实施例还提供了一种计算机可读存储介质,所述计算机可读存储介质存储有计算机程序,所述计算机程序被处理器执行时实现:
获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
根据所述关系概率,从所述多个实体中提取出目标关系信息;
基于所述目标关系信息,生成针对所述候选人的面试题目。
在本申请实施例中,所述计算机程序被处理器执行时还实现:
识别目标回复语句中的多个实体,根据所述多个实体生成待处理的实体序列,所述目标回复语句为所述多个回复语句中的任意一个;
将所述待处理的实体序列输入预设的语言模型中,获得所述目标回复语句中每个实体的概率分布,所述语言模型为基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器;
根据所述目标回复语句中每个实体的概率分布,生成所述目标回复语句的句向量信息。
在本申请实施例中,所述计算机程序被处理器执行时还实现生成变压器解码器,所述基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器通过采用如下编码方式生成:
h 0=TW e+W p
Figure PCTCN2020119298-appb-000018
其中,T是句子对应的独热码one-hot向量组成的矩阵,W e是标记嵌入矩阵,W p是位置嵌入矩阵,L是变压器块的数量,h l是第l层变压器块的状态。
在本申请实施例中,所述计算机程序被处理器执行时还实现:
确定每个回复语句的句向量信息的权重值;
根据所述权重值,对所述每个回复语句的句向量信息进行加权求和,得到所述回复信息对应的语句集合向量信息。
在本申请实施例中,所述计算机程序被处理器执行时还实现:
分别计算每个回复语句的句向量信息与全部回复语句的句向量信息之和的比值,将所述比值作为对应的回复语句的句向量信息的权重值。
在本申请实施例中,所述计算机程序被处理器执行时还实现:
依次对所述语句集合向量信息进行线性变换和逻辑回归softmax变换,获得所述回复信息中包含的多个实体相互间的关系概率。
在本申请实施例中,所述计算机程序被处理器执行时还实现:
提取所述关系概率超过预设阈值的一个或多个实体对,作为目标关系信息。
在本申请实施例中,所述计算机程序被处理器执行时还实现:
根据所述候选人对多个面试题目的回复信息,生成针对所述候选人的面试评价报告;
将所述面试评价报告上传至区块链中。
所述终端设备400可以是桌上型计算机、笔记本、掌上电脑等计算设备。所述终端设备400可包括,但不仅限于,处理器410、存储器420。本领域技术人员可以理解,图4仅仅是终端设备400的一种示例,并不构成对终端设备400的限定,可以包括比图示更多或更少的部件,或者组合某些部件,或者不同的部件,例如所述终端设备400还可以包括输入输出设备、网络接入设备、总线等。
所述处理器410可以是中央处理单元(Central Processing Unit,CPU),还可以是其他通用处理器、数字信号处理器(Digital Signal Processor,DSP)、专用集成电路(Application Specific Integrated Circuit,ASIC)、现成可编程门阵列(Field-Programmable Gate Array,FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
所述存储器420可以是所述终端设备400的内部存储单元,例如终端设备400的硬盘或内存。所述存储器420也可以是所述终端设备400的外部存储设备,例如所述终端设备400上配备的插接式硬盘,智能存储卡(Smart Media Card,SMC),安全数字(Secure Digital,SD)卡,闪存卡(Flash Card)等等。进一步地,所述存储器420还可以既包括所述终端设备400的内部存储单元也包括外部存储设备。所述存储器420用于存储所述计算机程序421以及所述终端设备400所需的其他程序和数据。所述存储器420还可以用于暂时地存储已经输出或者将要输出的数据。
本领域普通技术人员可以理解,实现上述实施例方法中的全部或部分流程,是可以通过计算机可读指令来指令相关的硬件来完成,所述的计算机可读指令可存储于一非易失性计算机可读取存储介质中,该计算机可读指令在执行时,可包括如上述各方法的实施例的流程。其中,本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用,均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限,RAM以多种形式可得,诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
以上所述实施例仅用以说明本申请的技术方案,而非对其限制。尽管参照前述实施例对本申请进行了详细的说明,本领域的普通技术人员应当理解:其依然可以对前述各实施例所记载的技术方案进行修改,或者对其中部分技术特征进行等同替换;而这些修改或者替换,并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围,均应包含在本申请的保护范围之内。

Claims (20)

  1. 一种智能面试方法,其中,包括:
    获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
    采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
    根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
    采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
    根据所述关系概率,从所述多个实体中提取出目标关系信息;
    基于所述目标关系信息,生成针对所述候选人的面试题目。
  2. 根据权利要求1所述的方法,其中,所述采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息,包括:
    识别目标回复语句中的多个实体,根据所述多个实体生成待处理的实体序列,所述目标回复语句为所述多个回复语句中的任意一个;
    将所述待处理的实体序列输入预设的语言模型中,获得所述目标回复语句中每个实体的概率分布,所述语言模型为基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器;
    根据所述目标回复语句中每个实体的概率分布,生成所述目标回复语句的句向量信息。
  3. 根据权利要求2所述的方法,其中,所述基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器通过采用如下编码方式生成:
    h 0=TW e+W p
    Figure PCTCN2020119298-appb-100001
    其中,T是句子对应的独热码one-hot向量组成的矩阵,W e是标记嵌入矩阵,W p是位置嵌入矩阵,L是变压器块的数量,h l是第l层变压器块的状态。
  4. 根据权利要求1-3任一项所述的方法,其中,所述根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息,包括:
    确定每个回复语句的句向量信息的权重值;
    根据所述权重值,对所述每个回复语句的句向量信息进行加权求和,得到所述回复信息对应的语句集合向量信息。
  5. 根据权利要求4所述的方法,其中,所述确定每个回复语句的句向量信息的权重值,包括:
    分别计算每个回复语句的句向量信息与全部回复语句的句向量信息之和的比值,将所述比值作为对应的回复语句的句向量信息的权重值。
  6. 根据权利要求1或2或3或5所述的方法,其中,所述采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率,包括:
    依次对所述语句集合向量信息进行线性变换和逻辑回归softmax变换,获得所述回复信息中包含的多个实体相互间的关系概率。
  7. 根据权利要求6所述的方法,其中,所述根据所述关系概率,从所述多个实体中提取出目标关系信息,包括:
    提取所述关系概率超过预设阈值的一个或多个实体对,作为目标关系信息。
  8. 根据权利要求1或2或3或5或7所述的方法,其中,还包括:
    根据所述候选人对多个面试题目的回复信息,生成针对所述候选人的面试评价报告;
    将所述面试评价报告上传至区块链中。
  9. 一种智能面试装置,其中,包括:
    获取模块,用于获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
    转换模块,用于采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
    确定模块,用于根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
    计算模块,用于采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
    提取模块,用于根据所述关系概率,从所述多个实体中提取出目标关系信息;
    生成模块,用于基于所述目标关系信息,生成针对所述候选人的面试题目。
  10. 如权利要求9所述的智能面试装置,其中,所述转换模块具体可以包括如下子模块:
    实体序列生成子模块,用于识别目标回复语句中的多个实体,根据所述多个实体生成待处理的实体序列,所述目标回复语句为所述多个回复语句中的任意一个;
    概率分布计算子模块,用于将所述待处理的实体序列输入预设的语言模型中,获得所述目标回复语句中每个实体的概率分布,所述语言模型为基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器;
    句向量信息生成子模块,用于根据所述目标回复语句中每个实体的概率分布,生成所述目标回复语句的句向量信息。
  11. 如权利要求10所述的智能面试装置,其中,所述基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器通过采用如下编码方式生成:
    h 0=TW e+W p
    Figure PCTCN2020119298-appb-100002
    其中,T是句子对应的独热码one-hot向量组成的矩阵,W e是标记嵌入矩阵,W p是位置嵌入矩阵,L是变压器块的数量,h l是第l层变压器块的状态。
  12. 如权利要求9-11任一项所述的智能面试装置,其中,所述确定模块具体可以包括如下子模块:
    权重值确定子模块,用于确定每个回复语句的句向量信息的权重值;
    语句集合向量信息生成子模块,用于根据所述权重值,对所述每个回复语句的句向量信息进行加权求和,得到所述回复信息对应的语句集合向量信息。
  13. 一种终端设备,其中,包括:存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序,所述处理器执行所述计算机程序时实现:
    获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
    采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
    根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
    采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
    根据所述关系概率,从所述多个实体中提取出目标关系信息;
    基于所述目标关系信息,生成针对所述候选人的面试题目。
  14. 根据权利要求13所述的终端设备,其中,所述处理器执行所述计算机程序时还实现:
    识别目标回复语句中的多个实体,根据所述多个实体生成待处理的实体序列,所述目标回复语句为所述多个回复语句中的任意一个;
    将所述待处理的实体序列输入预设的语言模型中,获得所述目标回复语句中每个实体的概率分布,所述语言模型为基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器;
    根据所述目标回复语句中每个实体的概率分布,生成所述目标回复语句的句向量信息。
  15. 根据权利要求14所述的终端设备,其中,所述处理器执行所述计算机程序时还实现生成变压器解码器,所述基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码 器通过采用如下编码方式生成:
    h 0=TW e+W p
    Figure PCTCN2020119298-appb-100003
    其中,T是句子对应的独热码one-hot向量组成的矩阵,W e是标记嵌入矩阵,W p是位置嵌入矩阵,L是变压器块的数量,h l是第l层变压器块的状态。
  16. 根据权利要求13-15任一项所述的终端设备,其中,所述处理器执行所述计算机程序时还实现:
    确定每个回复语句的句向量信息的权重值;
    根据所述权重值,对所述每个回复语句的句向量信息进行加权求和,得到所述回复信息对应的语句集合向量信息。
  17. 一种计算机可读存储介质,所述计算机可读存储介质存储有计算机程序,其中,所述计算机程序被处理器执行时实现:
    获取候选人在面试过程中的回复信息,所述回复信息包括多个回复语句;
    采用预设的语言模型,分别将所述多个回复语句转换为对应的句向量信息;
    根据所述多个回复语句的句向量信息,确定所述回复信息对应的语句集合向量信息;
    采用所述语句集合向量信息,计算所述回复信息中包含的多个实体相互间的关系概率;
    根据所述关系概率,从所述多个实体中提取出目标关系信息;
    基于所述目标关系信息,生成针对所述候选人的面试题目。
  18. 根据权利要求17所述的计算机可读存储介质,其中,所述计算机程序被处理器执行时还实现:
    识别目标回复语句中的多个实体,根据所述多个实体生成待处理的实体序列,所述目标回复语句为所述多个回复语句中的任意一个;
    将所述待处理的实体序列输入预设的语言模型中,获得所述目标回复语句中每个实体的概率分布,所述语言模型为基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器;
    根据所述目标回复语句中每个实体的概率分布,生成所述目标回复语句的句向量信息。
  19. 根据权利要求18所述的计算机可读存储介质,其中,所述计算机程序被处理器执行时还实现生成变压器解码器,所述基于位置前馈操作的有屏蔽的多头自注意力机制的变压器解码器通过采用如下编码方式生成:
    h 0=TW e+W p
    Figure PCTCN2020119298-appb-100004
    其中,T是句子对应的独热码one-hot向量组成的矩阵,W e是标记嵌入矩阵,W p是位置嵌入矩阵,L是变压器块的数量,h l是第l层变压器块的状态。
  20. 根据权利要求17-19任一项所述的计算机可读存储介质,其中,所述计算机程序被处理器执行时还实现:
    确定每个回复语句的句向量信息的权重值;
    根据所述权重值,对所述每个回复语句的句向量信息进行加权求和,得到所述回复信息对应的语句集合向量信息。
PCT/CN2020/119298 2020-05-28 2020-09-30 一种智能面试方法、装置及终端设备 WO2021139278A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010466693.9 2020-05-28
CN202010466693.9A CN111695335A (zh) 2020-05-28 2020-05-28 一种智能面试方法、装置及终端设备

Publications (1)

Publication Number Publication Date
WO2021139278A1 true WO2021139278A1 (zh) 2021-07-15

Family

ID=72478712

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/119298 WO2021139278A1 (zh) 2020-05-28 2020-09-30 一种智能面试方法、装置及终端设备

Country Status (2)

Country Link
CN (1) CN111695335A (zh)
WO (1) WO2021139278A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113672698A (zh) * 2021-08-01 2021-11-19 北京网聘咨询有限公司 一种基于表达分析的智能面试方法、系统、设备和存储介质

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111695335A (zh) * 2020-05-28 2020-09-22 平安科技(深圳)有限公司 一种智能面试方法、装置及终端设备
CN112613306A (zh) * 2020-12-31 2021-04-06 恒安嘉新(北京)科技股份公司 抽取实体关系的方法、装置、电子设备、及存储介质
CN114648315A (zh) * 2022-05-09 2022-06-21 北京全景智招科技有限公司 虚拟面试的方法、装置、设备及存储介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108733792A (zh) * 2018-05-14 2018-11-02 北京大学深圳研究生院 一种实体关系抽取方法
CN109670023A (zh) * 2018-12-14 2019-04-23 平安城市建设科技(深圳)有限公司 人机自动面试方法、装置、设备和存储介质
US10372763B2 (en) * 2015-07-13 2019-08-06 International Business Machines Corporation Generating probabilistic annotations for entities and relations using reasoning and corpus-level evidence
CN110889583A (zh) * 2019-10-11 2020-03-17 平安科技(深圳)有限公司 Ai面试方法、装置、计算机设备及计算机可读存储介质
CN111104517A (zh) * 2019-10-01 2020-05-05 浙江工商大学 一种基于两个三元组的中文问题生成方法
CN111695335A (zh) * 2020-05-28 2020-09-22 平安科技(深圳)有限公司 一种智能面试方法、装置及终端设备

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10372763B2 (en) * 2015-07-13 2019-08-06 International Business Machines Corporation Generating probabilistic annotations for entities and relations using reasoning and corpus-level evidence
CN108733792A (zh) * 2018-05-14 2018-11-02 北京大学深圳研究生院 一种实体关系抽取方法
CN109670023A (zh) * 2018-12-14 2019-04-23 平安城市建设科技(深圳)有限公司 人机自动面试方法、装置、设备和存储介质
CN111104517A (zh) * 2019-10-01 2020-05-05 浙江工商大学 一种基于两个三元组的中文问题生成方法
CN110889583A (zh) * 2019-10-11 2020-03-17 平安科技(深圳)有限公司 Ai面试方法、装置、计算机设备及计算机可读存储介质
CN111695335A (zh) * 2020-05-28 2020-09-22 平安科技(深圳)有限公司 一种智能面试方法、装置及终端设备

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113672698A (zh) * 2021-08-01 2021-11-19 北京网聘咨询有限公司 一种基于表达分析的智能面试方法、系统、设备和存储介质

Also Published As

Publication number Publication date
CN111695335A (zh) 2020-09-22

Similar Documents

Publication Publication Date Title
WO2021139278A1 (zh) 一种智能面试方法、装置及终端设备
WO2021114840A1 (zh) 基于语义分析的评分方法、装置、终端设备及存储介质
US11348352B2 (en) Contract lifecycle management
CN109992664B (zh) 争议焦点的标注分类方法、装置、计算机设备和存储介质
CN110163478B (zh) 一种合同条款的风险审查方法及装置
WO2020224219A1 (zh) 中文分词方法、装置、电子设备及可读存储介质
JP2021089705A (ja) 翻訳品質を評価するための方法と装置
CN110765763A (zh) 语音识别文本的纠错方法、装置、计算机设备和存储介质
WO2020233131A1 (zh) 问答处理方法、装置、计算机设备和存储介质
CN112328761B (zh) 一种意图标签设置方法、装置、计算机设备及存储介质
CN113743099B (zh) 基于自注意力机制方面术语提取系统、方法、介质、终端
WO2021027125A1 (zh) 序列标注方法、装置、计算机设备和存储介质
CN112686022A (zh) 违规语料的检测方法、装置、计算机设备及存储介质
CN110362798B (zh) 裁决信息检索分析方法、装置、计算机设备和存储介质
CN112883193A (zh) 一种文本分类模型的训练方法、装置、设备以及可读介质
CN113707299A (zh) 基于问诊会话的辅助诊断方法、装置及计算机设备
CN113505601A (zh) 一种正负样本对构造方法、装置、计算机设备及存储介质
Cornia et al. A unified cycle-consistent neural model for text and image retrieval
CN116070632A (zh) 一种非正式文本实体标签识别方法和装置
Afuecheta et al. Flexible models for stock returns based on Student's t distribution
CN112052329A (zh) 文本摘要生成方法、装置、计算机设备及可读存储介质
Sheng et al. Semantic-preserving abstractive text summarization with Siamese generative adversarial net
Zhang et al. Coattention based bilstm for answer selection
WO2022227196A1 (zh) 一种数据分析方法、装置、计算机设备及存储介质
CN113792120B (zh) 图网络的构建方法及装置、阅读理解方法及装置

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20911634

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20911634

Country of ref document: EP

Kind code of ref document: A1