WO2023115761A1 - 基于时序知识图谱的事件检测方法和装置 - Google Patents

基于时序知识图谱的事件检测方法和装置 Download PDF

Info

Publication number
WO2023115761A1
WO2023115761A1 PCT/CN2022/087213 CN2022087213W WO2023115761A1 WO 2023115761 A1 WO2023115761 A1 WO 2023115761A1 CN 2022087213 W CN2022087213 W CN 2022087213W WO 2023115761 A1 WO2023115761 A1 WO 2023115761A1
Authority
WO
WIPO (PCT)
Prior art keywords
event
knowledge graph
prediction model
training
detection
Prior art date
Application number
PCT/CN2022/087213
Other languages
English (en)
French (fr)
Inventor
鄂海红
宋美娜
许友日
Original Assignee
北京邮电大学
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 北京邮电大学 filed Critical 北京邮电大学
Publication of WO2023115761A1 publication Critical patent/WO2023115761A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36Creation of semantic tools, e.g. ontology or thesauri
    • G06F16/367Ontology

Definitions

  • the present disclosure relates to the fields of information technology and data services, and in particular to an event detection method and device based on a time series knowledge map.
  • a temporal knowledge graph is a collection of facts with temporal attributes.
  • the time series knowledge graph is characterized by incompleteness, that is, some facts may be missing under each timestamp.
  • it is characterized by constant updating, that is, new facts are constantly emerging. Therefore, the temporal knowledge map mainly has two tasks: completion and prediction.
  • the completion task is to complete the missing facts under each timestamp, and the prediction task is to predict what facts will happen in the future.
  • the current mainstream time series knowledge graph completion methods include Hyte, DE-SimplE, and TComplEx.
  • This type of method is based on model parameters and numerical vectors, and learns to judge true quadruples and false quadruples through scoring functions.
  • this type of model is a discriminative model, not a generative model. Due to the huge search space (the search space size of all possible quadruples is O(number of entities ⁇ number of relationships ⁇ number of entities ⁇ number of timestamps)), so Missing facts cannot be filled directly.
  • the current mainstream time series knowledge map prediction methods include RE-NET and CyGNet. This type of method is to predict future facts based on historical facts that have occurred. But none of them takes into account the impact of map incompleteness on predictions. Theoretically, due to the incompleteness of the map, the cues for prediction may be missing, and thus the performance of prediction may decrease. Therefore, there is a need to complete the map first, and then make predictions. However, since the existing completion models are all discriminative models, they cannot be completed directly.
  • the present disclosure aims to solve one of the technical problems in the related art at least to a certain extent.
  • the purpose of this disclosure is to improve the existing discriminative knowledge map completion model, solve the problem that the missing facts cannot be directly completed, thereby supplementing clues for event authenticity evaluation, and solving the problem of event authenticity evaluation performance problems.
  • an event detection method based on temporal knowledge graph is proposed.
  • Another object of the present disclosure is to propose an event detection device based on a temporal knowledge graph.
  • the present disclosure proposes an event detection method based on a time series knowledge graph, including the following steps:
  • the event to be detected includes a plurality of text data and a timestamp corresponding to each text data
  • the event detection method based on the time series knowledge map of the embodiment of the present disclosure, by acquiring the event to be detected, the event to be detected includes a plurality of text data and the timestamp corresponding to each text data; and training and predicting according to the completed time series knowledge map model to obtain a trained prediction model; input the event to be detected into the trained prediction model to obtain a detection result of the event to be detected.
  • the present disclosure can directly complete the sequence knowledge map, reduce the search space, and improve the performance of event detection.
  • the event detection method based on the temporal knowledge graph further includes: obtaining the completed temporal knowledge graph according to the temporal knowledge graph completion model training.
  • the training to obtain the completed time-series knowledge graph according to the completion model of the time-series knowledge graph includes:
  • time series knowledge map to perform complementary model training, learn all the facts of the training set, and obtain the first scoring function model to calculate the probability that each quadruple of the timestamp has occurred;
  • candidate triples are obtained from all the facts in the training set, and the candidate triples are combined with each of the occurred timestamps to obtain candidate quadruples ;
  • the candidate quadruples are scored by the completion model, and a preset number of candidate quadruples with the highest scores are selected to complete the current time stamp, so as to obtain the completed time series knowledge graph.
  • the training of the prediction model according to the completed time series knowledge graph to obtain the trained prediction model includes:
  • the probability that each quadruple of the future time stamp is established is calculated to obtain a trained prediction model.
  • the inputting the event to be detected into the trained prediction model to obtain the detection result of the event to be detected includes:
  • the event detection performance is evaluated by comparing with negative samples and ranking, so as to obtain the detection result of the event to be detected.
  • the training to obtain the completed sequence knowledge graph according to the sequence knowledge graph completion model further includes:
  • G static ⁇ (s,r,o)
  • the training of the prediction model according to the completed time series knowledge graph to obtain the trained prediction model includes:
  • G i:j represents the set of facts from timestamp i to j, and the probability of each head entity s at timestamp t is obtained from the graph feature Ht-1 at timestamp t-1 :
  • each tail entity o depends on the head entity s, the local features e s , e r of the relation r, and the historical features h t-1 (s, r) to obtain:
  • the probabilities for each entity are the index vector v q and the historical vocabulary Sum:
  • the generative module predicts new facts, without querying the relevant history:
  • the final probability is the sum of the copy probability and the generation probability:
  • the method also includes:
  • the second score function model is obtained to detect future events, including: head entity detection, tail entity detection and relationship detection; wherein,
  • the head entity detection includes: substituting the first entity into the first query to obtain the first score of the first entity, and selecting the highest entity as the answer from the first score to obtain the first detection event ;
  • the tail entity detection includes: substituting a second entity into a second query to obtain a second score of the second entity, and selecting the highest entity from the second score as an answer to obtain a second detection event ;
  • the relationship detection includes: substituting all relationships into a third query to obtain a third score of all relationships, and selecting the highest relationship from the third scores as an answer to obtain a third detection event.
  • Another aspect of the present disclosure proposes an event detection device based on a time series knowledge graph, including:
  • An acquisition module configured to acquire an event to be detected, where the event to be detected includes a plurality of text data and a timestamp corresponding to each text data;
  • the training module is used to train the prediction model according to the completed time-series knowledge map to obtain the trained prediction model
  • a detection module configured to input the event to be detected into the trained prediction model to obtain a detection result of the event to be detected.
  • Another aspect of the present disclosure proposes a computer device, including a memory, a processor, and a computer program stored on the memory and operable on the processor.
  • the processor executes the computer program, the above the method described.
  • Another aspect of the present disclosure provides a non-transitory computer-readable storage medium on which a computer program is stored, wherein the computer program implements the above method when executed by a processor.
  • Another aspect of the present disclosure proposes a computer program product, including computer instructions, which implement the method as described above when executed by at least one processor.
  • the event detection device based on the sequence knowledge graph of the embodiment of the present disclosure, by acquiring the event to be detected, the event to be detected includes a plurality of text data and the timestamp corresponding to each text data; and training and predicting according to the completed sequence knowledge graph model to obtain a trained prediction model; input the event to be detected into the trained prediction model to obtain a detection result of the event to be detected.
  • the present disclosure can directly complete the sequence knowledge map, reduce the search space, and improve the performance of event detection.
  • This disclosure improves the existing discriminative knowledge map completion model, realizes the direct completion of the time series knowledge map, supplements the clues for event authenticity evaluation, reduces the search space, and makes up for the existence of event authenticity evaluation performance. defects, improving the performance of event detection.
  • FIG. 1 is a schematic diagram of an event detection framework based on a time series knowledge graph according to an embodiment of the present disclosure
  • FIG. 2 is a flow chart of an event detection method based on a time series knowledge graph according to an embodiment of the present disclosure
  • FIG. 3 is a schematic diagram of prediction model training according to an embodiment of the present disclosure
  • Fig. 4 is a schematic diagram of an event detection method based on a time series knowledge graph according to an embodiment of the present disclosure.
  • Fig. 5 is a schematic structural diagram of an event detection device based on a time series knowledge graph according to an embodiment of the present disclosure.
  • the idea of the framework is to complete the time series knowledge graph first, and then make predictions.
  • the framework is mainly composed of three modules: (1) time series knowledge graph; (2) completion module; (3) detection module, as shown in Figure 1.
  • the framework uses a discriminative completion model for temporal knowledge graph completion, so that the completed graph can be used for event detection.
  • time series knowledge map completion model training, time stamp completion, event detection model training and event authenticity assessment There are mainly four interaction processes between the three modules: time series knowledge map completion model training, time stamp completion, event detection model training and event authenticity assessment.
  • This disclosure defines the problem, in Represents the knowledge map at time t i , which can be expressed as a set of quadruples, namely s, o ⁇ V, V is the entity set of G; r ⁇ R, R is the relationship set of G.
  • the goal is to enable the system to learn and train on this map, and obtain the ability to predict future facts, that is, to be able to answer queries (s,r,?,t) or (?,r,o,t) or (s,? ,o,t), where t>t n .
  • Fig. 2 is a flowchart of an event detection method based on a time series knowledge graph according to an embodiment of the present disclosure.
  • the event detection method based on time series knowledge graph includes the following steps:
  • step S1 the event to be detected is acquired, and the event to be detected includes a plurality of text data and a time stamp corresponding to each text data.
  • the event to be detected can be obtained by browsing websites such as microblogs and news webpages.
  • Step S2 train the prediction model according to the completed time-series knowledge graph, and obtain the trained prediction model.
  • the completed time-series knowledge graph is obtained by training the time-series knowledge graph completion model first.
  • the time series knowledge graph completion model is trained first, and the time series knowledge graph completion model is responsible for learning all events in the training set to obtain a discriminant scoring function model f(s, r, o, t), which can measure the time that has occurred The probability of any four-tuple (s,r,o,t i ) in ⁇ t 1 ,...,t n ⁇ being established (where 1 ⁇ i ⁇ n).
  • the present disclosure takes the completion model DE-SimplE as an example to describe the training of the completion model.
  • each entity e has 2 embeddings h e , t e , and each relation r has 2 embeddings v r .
  • DE-SimplE endows each entity v with a time encoder, so that the embedding of some dimensions of entities changes with time:
  • timestamp-by-timestamp completion because many events show periodic characteristics, that is, events that have occurred may occur again, so for an event that is missing at a certain timestamp t, it may be an event that appears at other timestamps, and these events The probability of being true is higher than that of a randomly generated event.
  • this disclosure takes all the events that have occurred as candidate triples for completion, and can obtain candidate triples from all events in the training set:
  • G static ⁇ (s,r,o)
  • the completed time-series knowledge graph is input into the prediction model for model training to obtain the second scoring function model, and the probability of each quaternion of future timestamps being established is calculated to obtain a trained prediction model.
  • the completed time series knowledge map is used as the input of the prediction model, and the score function model ⁇ (s,r,o,t) is obtained after training, which can measure any future time stamp ⁇ t n+1 ,...,t n+j ⁇
  • the probability that a quaternion (s,r,o,t i ) holds (where n+1 ⁇ i ⁇ n+j).
  • the present disclosure takes the recurring event network RE-NET and the copy generation network CyGNet as examples for illustration.
  • RE-NET defines the probability distribution of all facts in the time series knowledge graph G as
  • G i:j represents the set of facts from timestamp i to j, and the probability of each head entity s at timestamp t is obtained from the graph feature Ht-1 at timestamp t-1 :
  • each tail entity o depends on the head entity s, the local features e s , e r of the relation r, and the historical features h t-1 (s, r) to obtain:
  • the copy generation network CyGNet utilizes the copy and generation mechanism to identify and predict the fact with periodic repetition. For time t k , the historical vocabulary of each quadruple (s,p,?,t k ) to be queried is:
  • the copy module first generates the index vector
  • the probabilities for each entity are the index vector v q and the historical vocabulary Sum:
  • the generation module directly predicts new facts without querying the relevant history:
  • the final probability is the sum of the copy probability and the generation probability:
  • Step S3 inputting the event to be detected into the trained prediction model to obtain the detection result of the event to be detected.
  • each quadruple of the fact set G test ⁇ G tn-1 ,...,G tn+j ⁇ of future time stamps can be scored, by comparing and ranking with negative samples , to evaluate the fact prediction performance of the whole framework to obtain the detection results of the event to be detected.
  • the detection results obtained in the present disclosure are more authentic, and the prediction is more reliable.
  • the training, query and detection process of the present disclosure is shown in FIG. 3 .
  • Head entity detection For a certain query (?, r, o, t), substitute all entities s' into the query to get the scores of all entities Select the highest entity s as the answer, and get the event (s, r, o, t).
  • Tail entity detection For a certain query (s, r,?, t), substitute all entities o' into the query to get the scores of all entities Select the highest entity o as the answer, and get the event (s, r, o, t).
  • Relationship detection (s,?, o, t): For a certain query (s,?, o, t), substitute all relationships r' into the query to get the scores of all relationships Select the highest relationship r as the answer, and get the event (s, r, o, t).
  • the usage flow of the detection in the present disclosure is shown in FIG. 4 .
  • the event to be detected includes multiple text data and the timestamp corresponding to each text data; and training the prediction model according to the completed time series knowledge map to obtain the trained prediction model;
  • the event to be detected is input into the trained prediction model to obtain the detection result of the event to be detected.
  • the present disclosure can directly complete the sequence knowledge map, reduce the search space, and improve the performance of event detection.
  • this embodiment also provides an event detection device 10 based on a time series knowledge graph.
  • the device 10 includes: an acquisition module 100 , a training module 200 , and a detection module 300 .
  • An acquisition module 100 configured to acquire an event to be detected, where the event to be detected includes a plurality of text data and a timestamp corresponding to each text data;
  • the training module 200 is used to train the prediction model according to the completed time-series knowledge map to obtain the trained prediction model;
  • the detection module 300 is configured to input the event to be detected into the trained prediction model to obtain a detection result of the event to be detected.
  • the above-mentioned training module 200 is further configured to: obtain the completed sequence knowledge graph according to the sequence knowledge graph completion model training.
  • the event to be detected includes a plurality of text data and the timestamp corresponding to each text data; and training according to the completed timing knowledge graph
  • a prediction model is used to obtain a trained prediction model; the event to be detected is input into the trained prediction model to obtain a detection result of the event to be detected.
  • the present disclosure can directly complete the sequence knowledge map, reduce the search space, and improve the performance of event detection.
  • Another aspect of the present disclosure proposes a computer device, including a memory, a processor, and a computer program stored on the memory and operable on the processor.
  • the processor executes the computer program, the above the method described.
  • Another aspect of the present disclosure provides a non-transitory computer-readable storage medium on which a computer program is stored, wherein the computer program implements the above method when executed by a processor.
  • Another aspect of the present disclosure proposes a computer program product, including computer instructions, which implement the method as described above when executed by at least one processor.
  • a "computer-readable medium” may be any device that can contain, store, communicate, propagate or transmit a program for use in or in conjunction with an instruction execution system, device or device.
  • computer-readable media include the following: electrical connection with one or more wires (electronic device), portable computer disk case (magnetic device), random access memory (RAM), Read Only Memory (ROM), Erasable and Editable Read Only Memory (EPROM or Flash Memory), Fiber Optic Devices, and Portable Compact Disc Read Only Memory (CDROM).
  • the computer-readable medium may even be paper or other suitable medium on which the program can be printed, as it may be possible, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or other suitable processing if necessary.
  • the program is processed electronically and stored in computer memory.
  • each functional unit in each embodiment of the present invention may be integrated into one processing module, each unit may exist separately physically, or two or more units may be integrated into one module.
  • the above-mentioned integrated modules can be implemented in the form of hardware or in the form of software function modules. If the integrated modules are implemented in the form of software function modules and sold or used as independent products, they can also be stored in a computer-readable storage medium.
  • the storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, and the like.
  • first and second are used for descriptive purposes only, and cannot be interpreted as indicating or implying relative importance or implicitly specifying the quantity of indicated technical features.
  • the features defined as “first” and “second” may explicitly or implicitly include at least one of these features.
  • “plurality” means at least two, such as two, three, etc., unless otherwise specifically defined.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Animal Behavior & Ethology (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本公开公开了一种基于时序知识图谱的事件检测方法和装置,其中,该方法包括:获取待检测事件,待检测事件包括多个文本数据以及每个文本数据对应的时间戳;以及,根据补全后的时序知识图谱训练预测模型,得到训练好的预测模型;将待检测事件输入训练好的预测模型,以得到待检测事件的检测结果。本公开能够对时序知识图谱直接进行补全,减小了搜索空间,提升了事件检测的性能。

Description

基于时序知识图谱的事件检测方法和装置
相关申请的交叉引用
本申请基于申请号为202111566708.X、申请日为2021年12月20日的中国专利申请提出,并要求该中国专利申请的优先权,该中国专利申请的全部内容在此引入本申请作为参考。
技术领域
本公开涉及信息技术及数据业务领域,尤其涉及一种基于时序知识图谱的事件检测方法和装置。
背景技术
时序知识图谱是带有时间属性的事实的集合。时序知识图谱可表示为四元组的集合G={(s,r,o,t)},其中s代表头实体,o代表尾实体,r代表关系,t代表时间,如(梅西,获奖,世界足球先生,2009年)。通常,时序知识图谱具有不完整性的特点,即每个时间戳下可能缺失了部分事实。此外,它还具有不断更新的特点,即新的事实会不断出现。因此,时序知识图谱主要有2个任务:补全和预测。补全任务即补全每个时间戳下缺失的事实,而预测任务即预测未来会发生什么事实。
目前主流的时序知识图谱补全方法包括Hyte,DE-SimplE,TComplEx。这一类方法是基于模型参数和数值向量,通过得分函数学会判断真四元组和假四元组。但这一类模型都是判别式模型,不是生成式模型,由于搜索空间巨大(所有可能的四元组的搜索空间大小为O(实体数×关系数×实体数×时间戳数)),因此无法直接补全缺失的事实。
目前主流的时序知识图谱预测方法包括RE-NET,CyGNet。这一类方法是基于已发生的历史事实,预测未来的事实。但它们都没有考虑到图谱不完整性对预测的影响。理论上,由于图谱的不完整性,用于预测的线索可能缺失,因此可能导致预测的性能下降。因此,产生了先对图谱补全,再做预测的需求。但由于现有补全模型都是判别式模型,因此无法直接进行补全。
发明内容
本公开旨在至少在一定程度上解决相关技术中的技术问题之一。
为此,本公开的目的在于改进现有判别式的知识图谱补全模型,解决无法直接补全缺失事实的问题,从而补充了用于事件真实性评估的线索,解决了事件真实性评估性能存在缺陷的问题,提出了一种基于时序知识图谱的事件检测方法。
本公开的另一个目的在于提出一种基于时序知识图谱的事件检测装置。
本公开一方面提出了基于时序知识图谱的事件检测方法,包括以下步骤:
获取待检测事件,所述待检测事件包括多个文本数据以及每个文本数据对应的时间戳;以及,
根据补全后的时序知识图谱训练预测模型,得到训练好的预测模型;
将所述待检测事件输入所述训练好的预测模型,以得到所述待检测事件的检测结果。
本公开实施例的基于时序知识图谱的事件检测方法,通过获取待检测事件,待检测事件包括多个文本数据以及每个文本数据对应的时间戳;以及,根据补全后的时序知识图谱训练预测模型,得到训练好的预测模型;将待检测事件输入训练好的预测模型,以得到待检测事件的检测结果。本公开能够对时序知识图谱直接进行补全,减小了搜索空间,提升了事件检测的性能。
在一些实施方式中,所述基于时序知识图谱的事件检测方法还包括:根据时序知识图谱补全模型训练得到所述补全后的时序知识图谱。
在一些实施方式中,所述根据时序知识图谱补全模型训练得到所述补全后的时序知识图谱,包括:
利用所述时序知识图谱进行补全模型训练,学习训练集的所有事实,得到第一得分函数模型,以计算已发生时间戳每个四元组成立的概率;
通过对所述时序知识图谱的逐时间戳补全,从所述训练集的所有事实得到候选三元组,将所述候选三元组与每个所述已发生时间戳组合得到候选四元组;
通过所述补全模型对所述候选四元组进行评分,选出评分最高的预设个数的候选四元组补全当前时间戳,以得到所述补全后的时序知识图谱。
在一些实施方式中,所述根据补全后的时序知识图谱训练预测模型,以得到训练好的预测模型,包括:
将所述补全后的时序知识图谱输入预测模型进行模型训练;
基于所述模型训练,得到第二得分函数模型;
根据所述第二得分函数模型,计算未来时间戳每个四元组成立的概率,以得到训练好的预测模型。
在一些实施方式中,所述将所述待检测事件输入所述训练好的预测模型,以得到所述待检测事件的检测结果,包括:
在完成所述预测模型的训练后,对所述未来时间戳每个四元组进行评分,得到评分结果;
基于所述评分结果,通过与负样本进行比较和排名评估事件检测性能,以得到所述待检测事件的检测结果。
在一些实施方式中,所述根据时序知识图谱补全模型训练得到所述补全后的时序知识图谱,还包括:
通过忽略时间戳,将所有事实用于补全的候选三元组:
G static={(s,r,o)|(s,r,o,t)∈G t}
对每个时间戳t,与所述补全的候选三元组组合,得到所述候选四元组:
Figure PCTCN2022087213-appb-000001
使用在所述时序知识图谱上训练得到的DE-SimplE对所述候选四元组进行评分,选出得分最高的前k个四元组,作为所述每个时间戳t的补全事件:
Figure PCTCN2022087213-appb-000002
则所述每个时间戳t的图谱更新为:
Figure PCTCN2022087213-appb-000003
在一些实施方式中,所述根据补全后的时序知识图谱训练预测模型,以得到训练好的预测模型,包括:
通过循环事件网络RE-NET定义时序知识图谱G的所有事实的概率分布为:
Figure PCTCN2022087213-appb-000004
其中,G i:j代表从时间戳i到j的事实集合,时间戳t的每个头实体s的概率从时间戳t-1的图特征H t-1得到:
Figure PCTCN2022087213-appb-000005
每个关系r的概率由头实体s的局部特征e s和历史特征h t-1(s)得到:
Figure PCTCN2022087213-appb-000006
每个尾实体o的概率依赖于头实体s、关系r的局部特征e s,e r以及历史特征h t-1(s,r)得到:
Figure PCTCN2022087213-appb-000007
通过拷贝生成网络CyGNet,对于时间t k,每一个要查询的四元组(s,p,?,t k)的历史词汇为:
Figure PCTCN2022087213-appb-000008
其中,
Figure PCTCN2022087213-appb-000009
是N维的multi-hot向量,拷贝模块生成索引向量:
v q=tanh(W c[s,p,t k]+b c)
每个实体的概率为索引向量v q与历史词汇
Figure PCTCN2022087213-appb-000010
之和:
Figure PCTCN2022087213-appb-000011
生成模块预测新事实,不查询相关历史:
p(g)=softmax(W g[s,p,t k]+b g)
最终概率为拷贝概率与生成概率之和:
p(o|s,p,t)=α·p(c)+(1-α)·p(g)。
在一些实施方式中,所述方法,还包括:
在完成对所述预测模型的训练后,得到所述第二得分函数模型,以检测未来的事件,包括:头实体检测、尾实体检测和关系检测;其中,
所述头实体检测,包括:将第一实体代入第一查询,得到所述第一实体的第一得分,从所述第一得分中选出得到最高的实体作为答案,以得到第一检测事件;
所述尾实体检测,包括:将第二实体代入第二查询,得到所述第二实体的第二得分,从所述第二得分中选出得到最高的实体作为答案,以得到第二检测事件;
所述关系检测,包括:将所有关系代入第三查询,得到所述所有关系的第三得分,从所述第三得分中选出得到最高的关系作为答案,以得到第三检测事件。
本公开另一方面提出了一种基于时序知识图谱的事件检测装置,包括:
获取模块,用于获取待检测事件,所述待检测事件包括多个文本数据以及每个文本数据对应的时间戳;
训练模块,用于根据补全后的时序知识图谱训练预测模型,得到训练好的预测模型;
检测模块,用于将所述待检测事件输入所述训练好的预测模型,以得到所述待检测事件的检测结果。
本公开另一方面提出了一种计算机设备,包括存储器、处理器及存储在所述存储器上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,实现如上所述的方法。
本公开另一方面提出了一种非临时性计算机可读存储介质,其上存储有计算机程序,其中,所述计算机程序被处理器执行时实现如上所述的方法。
本公开另一方面提出了一种计算机程序产品,包括计算机指令,所述计算机指令被至少一个处理器执行时实现如上所述的方法。
本公开实施例的基于时序知识图谱的事件检测装置,通过获取待检测事件,待检测事件包括多个文本数据以及每个文本数据对应的时间戳;以及,根据补全后的时序知识图谱训练预测模型,得到训练好的预测模型;将待检测事件输入训练好的预测模型,以得到待检测事件的检测结果。本公开能够对时序知识图谱直接进行补全,减小了搜索空间,提升了事件检测的性能。
本公开改进现有判别式的知识图谱补全模型,实现对时序知识图谱直接进行补全,补充了用于事件真实性评估的线索,减小了搜索空间,弥补了事件真实性评估性能存在的缺陷,提升了事件检测的性能。
本公开附加的方面和优点将在下面的描述中部分给出,部分将从下面的描述中变得明显,或通过本公开的实践了解到。
附图说明
本公开上述的和/或附加的方面和优点从下面结合附图对实施例的描述中将变得明显和容易理解,其中:
图1为根据本公开实施例的基于时序知识图谱的事件检测框架示意图;
图2为根据本公开实施例的基于时序知识图谱的事件检测方法的流程图;
图3为根据本公开实施例的预测模型训练示意图;
图4为根据本公开实施例的基于时序知识图谱的事件检测方法的使用示意图。
图5为根据本公开实施例的基于时序知识图谱的事件检测装置的结构示意图。
具体实施方式
需要说明的是,在不冲突的情况下,本申请中的实施例及实施例中的特征可以相互组合。下面将参考附图并结合实施例来详细说明本公开。
为了使本技术领域的人员更好地理解本公开方案,下面将结合本公开实施例中的附图,对本公开实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例仅仅是本公开一部分的实施例,而不是全部的实施例。基于本公开中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都应当属于本公开保护的范围。
下面参照附图描述根据本公开实施例提出的基于时序知识图谱的事件检测方法及装置,首先将参照附图描述根据本公开实施例提出的基于时序知识图谱的事件检测方法。
在时序知识图谱中,每个时间戳有缺失或潜在的事件。框架的思想是先对时序知识图谱补全,再进行预测。该框架主要由3个模块组成:(1)时序知识图谱;(2)补全模块;(3)检测模块,如图1所示。同时,该框架以一种减小搜索空间的补全策略,将判别式补全模型用于时序知识图谱补全,以便将补全后的图谱用于事件检测。3个模块之间主要有4个交互流程:时序知识图谱补全模型训练,逐时间戳补全,事件检测模型训练及评估事件真实性。
本公开对问题进行定义,
Figure PCTCN2022087213-appb-000012
其中
Figure PCTCN2022087213-appb-000013
表示在t i时刻的知识图谱,可表示为四元组的集合,即
Figure PCTCN2022087213-appb-000014
s,o∈V,V为G的实体集合;r∈R,R为G的关系集合。
给定一个已发生的时序知识图谱
Figure PCTCN2022087213-appb-000015
用于训练,目标是使系统在该图谱上学习训练,获得预测未来事实的能力,即能够回答查询(s,r,?,t)或(?,r,o,t)或(s,?,o,t),其中t>t n
图2是本公开一个实施例的基于时序知识图谱的事件检测方法的流程图。
如图2所示,该基于时序知识图谱的事件检测方法包括以下步骤:
步骤S1,获取待检测事件,待检测事件包括多个文本数据以及每个文本数据对应的时间戳。
待检测事件可通过微博、新闻网页等浏览网站获得,例如,该待检测事件的文本数据包括:四元组集合,G={(s,r,o,t)},其中s代表头实体,o代表尾实体,r代表关系,t代表时 间。
步骤S2,根据补全后的时序知识图谱训练预测模型,得到训练好的预测模型。
可以理解的是,本公开先根据时序知识图谱补全模型训练得到补全后的时序知识图谱。
先进行时序知识图谱补全模型训练,时序知识图谱补全模型负责学习训练集的所有事件,以得到一个判别式的得分函数模型f(s,r,o,t),能够衡量已发生的时间戳内{t 1,…,t n}任意一个四元组(s,r,o,t i)成立的概率(其中1≤i≤n)。
作为一种示例,本公开以补全模型DE-SimplE为例说明补全模型的训练。首先假设每个实体e有2个嵌入h e、t e,每个关系r有2个嵌入v r
Figure PCTCN2022087213-appb-000016
假设实体的部分特征是静态的,部分特征是动态的,DE-SimplE赋予每个实体v一个时间编码器,使实体的部分维度的嵌入随时间变化:
Figure PCTCN2022087213-appb-000017
其中,
Figure PCTCN2022087213-appb-000018
代表实体v的静态嵌入,ω v、b v代表实体v编码时间的权值向量,σ为激活函数(如sin函数)。四元组(s,r,o,t)的得分定义为
Figure PCTCN2022087213-appb-000019
Figure PCTCN2022087213-appb-000020
再进行逐时间戳补全,由于许多事件呈现出周期性的特点,即已发生的事件可能再次发生,从而对于某个时间戳t缺失的事件,可能是其他时间戳出现的事件,并且这些事件成立的概率比随机生成的事件概率高。
因此,为减小补全的搜索空间,通过忽略时间戳,本公开将所有发生过的事件作为用于补全的候选三元组,可从训练集的所有事件得到候选三元组:
G static={(s,r,o)|(s,r,o,t)∈G t}
然后,对每个时间戳t,与候选三元组组合,得到候选四元组
Figure PCTCN2022087213-appb-000021
然后,这些候选四元组的真实性可以用判别式补全模型去衡量。一个四元组的分数越高代表着模型认为该四元组的真实性越高。使用在不完整图谱上训练得到的DE-SimplE对候选四元组评分,选出得分最高的前k个四元组,作为该时间戳t的补全事件:
Figure PCTCN2022087213-appb-000022
最终,每个时间戳t的图谱更新为:
Figure PCTCN2022087213-appb-000023
在一些实施方式中,将补全后的时序知识图谱输入预测模型进行模型训练,得到第二得分函数模型,计算未来时间戳每个四元组成立的概率,以得到训练好的预测模型。
将补全后的时序知识图谱作为预测模型的输入,训练后得到得分函数模型φ(s,r,o,t), 能够衡量未来时间戳{t n+1,…,t n+j}任意一个四元组(s,r,o,t i)成立的概率(其中n+1≤i≤n+j)。
作为一种示例,本公开以循环事件网络RE-NET和拷贝生成网络CyGNet为例进行阐述。
RE-NET定义时序知识图谱G的所有事实的概率分布为
Figure PCTCN2022087213-appb-000024
其中,G i:j代表从时间戳i到j的事实集合,时间戳t的每个头实体s的概率从时间戳t-1的图特征H t-1得到:
Figure PCTCN2022087213-appb-000025
每个关系r的概率由头实体s的局部特征e s和历史特征h t-1(s)得到:
Figure PCTCN2022087213-appb-000026
每个尾实体o的概率依赖于头实体s、关系r的局部特征e s,e r以及历史特征h t-1(s,r)得到:
Figure PCTCN2022087213-appb-000027
拷贝生成网络CyGNet利用拷贝和生成机制,识别和预测出具有周期重复性的事实。对于时间t k,每一个要查询的四元组(s,p,?,t k)的历史词汇为:
Figure PCTCN2022087213-appb-000028
其中,
Figure PCTCN2022087213-appb-000029
是N维的multi-hot向量。拷贝模块首先生成索引向量
v q=tanh(W c[s,p,t k]+b c)
每个实体的概率为索引向量v q与历史词汇
Figure PCTCN2022087213-appb-000030
之和:
Figure PCTCN2022087213-appb-000031
生成模块直接预测新事实,而不查询相关历史:
p(g)=softmax(W g[s,p,t k]+b g)
最终概率为拷贝概率与生成概率之和:
p(o|s,p,t)=α·p(c)+(1-α)·p(g)
步骤S3,将待检测事件输入训练好的预测模型,以得到待检测事件的检测结果。
在完成事实预测模型的训练后,可对未来时间戳的事实集合G test={G tn-1,…,G tn+j}的每个四元组进行评分,通过与负样本进行比较和排名,评估整个框架的事实预测性能,以得到待检测事件的检测结果。本公开得到的检测结果更加具有真实性,预测更加具有可靠性。
在完成预测模型的训练后,得到得分函数模型
Figure PCTCN2022087213-appb-000032
可对未来时间戳的事件集合G test={G tn+1,…,G tn+j}的每个四元组进行评分,进而得到待检测事件的检测结果,包括:头实体检测,尾实体检测和关系检测。本公开的训练、查询、检测过程如图3所示。
(1)头实体检测:对于某个查询(?,r,o,t),将所有的实体s’代入该查询,得到所有实体的得分
Figure PCTCN2022087213-appb-000033
从中选出得到最高的实体s作为答案,得到事件(s,r,o,t)。
(2)尾实体检测:对于某个查询(s,r,?,t),将所有的实体o’代入该查询,得到所有实体的得分
Figure PCTCN2022087213-appb-000034
从中选出得到最高的实体o作为答案,得到事件(s,r,o,t)。
(3)关系检测(s,?,o,t):对于某个查询(s,?,o,t),将所有的关系r’代入该查询,得到所有关系的得分
Figure PCTCN2022087213-appb-000035
从中选出得到最高的关系r作为答案,得到事件(s,r,o,t)。
在一些实施方式中,以查询尾实体检测为例,本公开检测的使用流程如图4所示。
通过上述步骤,通过获取待检测事件,待检测事件包括多个文本数据以及每个文本数据对应的时间戳;以及,根据补全后的时序知识图谱训练预测模型,得到训练好的预测模型;将待检测事件输入训练好的预测模型,以得到待检测事件的检测结果。本公开能够对时序知识图谱直接进行补全,减小了搜索空间,提升了事件检测的性能。
为了实现上述实施例,如图5所示,本实施例中还提供了一种基于时序知识图谱的事件检测装置10,该装置10包括:获取模块100,训练模块200,检测模块300。
获取模块100,用于获取待检测事件,待检测事件包括多个文本数据以及每个文本数据对应的时间戳;
训练模块200,用于根据补全后的时序知识图谱训练预测模型,得到训练好的预测模型;
检测模块300,用于将待检测事件输入训练好的预测模型,以得到待检测事件的检测结果。
在一些实施方式中,上述训练模块200,还用于:根据时序知识图谱补全模型训练得到所述补全后的时序知识图谱。
根据本公开实施例的基于时序知识图谱的事件检测装置,通过获取待检测事件,待检测事件包括多个文本数据以及每个文本数据对应的时间戳;以及,根据补全后的时序知识图谱训练预测模型,得到训练好的预测模型;将待检测事件输入训练好的预测模型,以得到待检测事件的检测结果。本公开能够对时序知识图谱直接进行补全,减小了搜索空间,提升了事件检测的性能。
需要说明的是,前述对基于时序知识图谱的事件检测方法实施例的解释说明也适用于该实施例的基于时序知识图谱的事件检测装置,此处不再赘述。
本公开另一方面提出了一种计算机设备,包括存储器、处理器及存储在所述存储器上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,实现如上所述的 方法。
本公开另一方面提出了一种非临时性计算机可读存储介质,其上存储有计算机程序,其中,所述计算机程序被处理器执行时实现如上所述的方法。
本公开另一方面提出了一种计算机程序产品,包括计算机指令,所述计算机指令被至少一个处理器执行时实现如上所述的方法。
流程图中或在此以其他方式描述的任何过程或方法描述可以被理解为,表示包括一个或更多个用于实现定制逻辑功能或过程的步骤的可执行指令的代码的模块、片段或部分,并且本发明的优选实施方式的范围包括另外的实现,其中可以不按所示出或讨论的顺序,包括根据所涉及的功能按基本同时的方式或按相反的顺序,来执行功能,这应被本发明的实施例所属技术领域的技术人员所理解。
在流程图中表示或在此以其他方式描述的逻辑和/或步骤,例如,可以被认为是用于实现逻辑功能的可执行指令的定序列表,可以具体实现在任何计算机可读介质中,以供指令执行系统、装置或设备(如基于计算机的系统、包括处理器的系统或其他可以从指令执行系统、装置或设备取指令并执行指令的系统)使用,或结合这些指令执行系统、装置或设备而使用。就本说明书而言,"计算机可读介质"可以是任何可以包含、存储、通信、传播或传输程序以供指令执行系统、装置或设备或结合这些指令执行系统、装置或设备而使用的装置。计算机可读介质的更具体的示例(非穷尽性列表)包括以下:具有一个或多个布线的电连接部(电子装置),便携式计算机盘盒(磁装置),随机存取存储器(RAM),只读存储器(ROM),可擦除可编辑只读存储器(EPROM或闪速存储器),光纤装置,以及便携式光盘只读存储器(CDROM)。另外,计算机可读介质甚至可以是可在其上打印所述程序的纸或其他合适的介质,因为可以例如通过对纸或其他介质进行光学扫描,接着进行编辑、解译或必要时以其他合适方式进行处理来以电子方式获得所述程序,然后将其存储在计算机存储器中。
应当理解,本发明的各部分可以用硬件、软件、固件或它们的组合来实现。在上述实施方式中,多个步骤或方法可以用存储在存储器中且由合适的指令执行系统执行的软件或固件来实现。如,如果用硬件来实现和在另一实施方式中一样,可用本领域公知的下列技术中的任一项或他们的组合来实现:具有用于对数据信号实现逻辑功能的逻辑门电路的离散逻辑电路,具有合适的组合逻辑门电路的专用集成电路,可编程门阵列(PGA),现场可编程门阵列(FPGA)等。
本技术领域的普通技术人员可以理解实现上述实施例方法携带的全部或部分步骤是可以通过程序来指令相关的硬件完成,所述的程序可以存储于一种计算机可读存储介质中,该程序在执行时,包括方法实施例的步骤之一或其组合。
此外,在本发明各个实施例中的各功能单元可以集成在一个处理模块中,也可以是各个单元单独物理存在,也可以两个或两个以上单元集成在一个模块中。上述集成的模块既可以采用硬件的形式实现,也可以采用软件功能模块的形式实现。所述集成的模块如果以软件功能模块的形式实现并作为独立的产品销售或使用时,也可以存储在一个计算机可读取存储介质中。
上述提到的存储介质可以是只读存储器,磁盘或光盘等。
此外,术语“第一”、“第二”仅用于描述目的,而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此,限定有“第一”、“第二”的特征可以明示或者隐含地包括至少一个该特征。在本公开的描述中,“多个”的含义是至少两个,例如两个,三个等,除非另有明确具体的限定。
在本说明书的描述中,参考术语“一个实施例”、“一些实施例”、“示例”、“具体示例”、或“一些示例”等的描述意指结合该实施例或示例描述的具体特征、结构、材料或者特点包含于本公开的至少一个实施例或示例中。在本说明书中,对上述术语的示意性表述不必须针对的是相同的实施例或示例。而且,描述的具体特征、结构、材料或者特点可以在任一个或多个实施例或示例中以合适的方式结合。此外,在不相互矛盾的情况下,本领域的技术人员可以将本说明书中描述的不同实施例或示例以及不同实施例或示例的特征进行结合和组合。
尽管上面已经示出和描述了本公开的实施例,可以理解的是,上述实施例是示例性的,不能理解为对本公开的限制,本领域的普通技术人员在本公开的范围内可以对上述实施例进行变化、修改、替换和变型。

Claims (13)

  1. 一种基于时序知识图谱的事件检测方法,包括:
    获取待检测事件,所述待检测事件包括多个文本数据以及每个文本数据对应的时间戳;
    根据补全后的时序知识图谱训练预测模型,得到训练好的预测模型;和
    将所述待检测事件输入所述训练好的预测模型,以得到所述待检测事件的检测结果。
  2. 根据权利要求1所述的基于时序知识图谱的事件检测方法,还包括:
    根据时序知识图谱补全模型训练得到所述补全后的时序知识图谱。
  3. 根据权利要求2所述的基于时序知识图谱的事件检测方法,其中,所述根据时序知识图谱补全模型训练得到所述补全后的时序知识图谱,包括:
    利用所述时序知识图谱进行补全模型训练,学习训练集的所有事实,得到第一得分函数模型,以计算已发生时间戳每个四元组成立的概率;
    通过对所述时序知识图谱的逐时间戳补全,从所述训练集的所有事实得到候选三元组,将所述候选三元组与每个所述已发生时间戳组合得到候选四元组;
    通过所述补全模型对所述候选四元组进行评分,选出评分最高的预设个数的候选四元组补全当前时间戳,以得到所述补全后的时序知识图谱。
  4. 根据权利要求2或3所述的基于时序知识图谱的事件检测方法,其中,所述根据补全后的时序知识图谱训练预测模型,以得到训练好的预测模型,包括:
    将所述补全后的时序知识图谱输入预测模型进行模型训练;
    基于所述模型训练,得到第二得分函数模型;
    根据所述第二得分函数模型,计算未来时间戳每个四元组成立的概率,以得到训练好的预测模型。
  5. 根据权利要求4所述的基于时序知识图谱的事件检测方法,其中,所述将所述待检测事件输入所述训练好的预测模型,以得到所述待检测事件的检测结果,包括:
    在完成所述预测模型的训练后,对所述未来时间戳每个四元组进行评分,得到评分结果;
    基于所述评分结果,通过与负样本进行比较和排名评估事件检测性能,以得到所述待检测事件的检测结果。
  6. 根据权利要求3所述的基于时序知识图谱的事件检测方法,其中,所述根据时序知识图谱补全模型训练得到所述补全后的时序知识图谱,还包括:
    通过忽略时间戳,将所有事实用于补全的候选三元组:
    G static={(s,r,o)|(s,r,o,t)∈G t}
    对每个时间戳t,与所述补全的候选三元组组合,得到所述候选四元组:
    Figure PCTCN2022087213-appb-100001
    使用在所述时序知识图谱上训练得到的DE-SimplE对所述候选四元组进行评分,选出得分最高的前k个四元组,作为所述每个时间戳t的补全事件:
    Figure PCTCN2022087213-appb-100002
    则所述每个时间戳t的图谱更新为:
    Figure PCTCN2022087213-appb-100003
  7. 根据权利要求4所述的基于时序知识图谱的事件检测方法,其中,所述根据补全后的时序知识图谱训练预测模型,以得到训练好的预测模型,包括:
    通过循环事件网络RE-NET定义时序知识图谱G的所有事实的概率分布为:
    Figure PCTCN2022087213-appb-100004
    其中,G i:j代表从时间戳i到j的事实集合,时间戳t的每个头实体s的概率从时间戳t-1的图特征H t-1得到:
    Figure PCTCN2022087213-appb-100005
    每个关系r的概率由头实体s的局部特征e s和历史特征h t-1(s)得到:
    Figure PCTCN2022087213-appb-100006
    每个尾实体o的概率依赖于头实体s、关系r的局部特征e s,e r以及历史特征h t-1(s,r)得到:
    Figure PCTCN2022087213-appb-100007
    通过拷贝生成网络CyGNet,对于时间t k,每一个要查询的四元组(s,p,?,t k)的历史词汇为:
    Figure PCTCN2022087213-appb-100008
    其中,
    Figure PCTCN2022087213-appb-100009
    是N维的multi-hot向量,拷贝模块生成索引向量:
    v q=tanh(W c[s,p,t k]+b c)
    每个实体的概率为索引向量v q与历史词汇
    Figure PCTCN2022087213-appb-100010
    之和:
    Figure PCTCN2022087213-appb-100011
    生成模块预测新事实,不查询相关历史:
    p(g)=softmax(W g[s,p,t k]+b g)
    最终概率为拷贝概率与生成概率之和:
    p(o|s,p,t)=α·p(c)+(1-α)·p(g)。
  8. 根据权利要求1至7中任一项所述的基于时序知识图谱的事件检测方法,还包括:
    在完成对所述预测模型的训练后,得到所述第二得分函数模型,以检测未来的事件,包 括:头实体检测、尾实体检测和关系检测;其中,
    所述头实体检测,包括:将第一实体代入第一查询,得到所述第一实体的第一得分,从所述第一得分中选出得到最高的实体作为答案,以得到第一检测事件;
    所述尾实体检测,包括:将第二实体代入第二查询,得到所述第二实体的第二得分,从所述第二得分中选出得到最高的实体作为答案,以得到第二检测事件;
    所述关系检测,包括:将所有关系代入第三查询,得到所述所有关系的第三得分,从所述第三得分中选出得到最高的关系作为答案,以得到第三检测事件。
  9. 一种基于时序知识图谱的事件检测装置,包括:
    获取模块,用于获取待检测事件,所述待检测事件包括多个文本数据以及每个文本数据对应的时间戳;
    训练模块,用于根据补全后的时序知识图谱训练预测模型,得到训练好的预测模型;
    检测模块,用于将所述待检测事件输入所述训练好的预测模型,以得到所述待检测事件的检测结果。
  10. 根据权利要求9所述的基于时序知识图谱的事件检测装置,其中,所述训练模块,还用于:
    根据时序知识图谱补全模型训练得到所述补全后的时序知识图谱。
  11. 一种计算机设备,包括存储器、处理器及存储在所述存储器上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,实现如权利要求1-8中任一所述的方法。
  12. 一种非临时性计算机可读存储介质,其上存储有计算机程序,其中,所述计算机程序被处理器执行时实现如权利要求1-8中任一所述的方法。
  13. 一种计算机程序产品,包括计算机指令,所述计算机指令被至少一个处理器执行时实现如权利要求1-8中的任一项所述的方法。
PCT/CN2022/087213 2021-12-20 2022-04-15 基于时序知识图谱的事件检测方法和装置 WO2023115761A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111566708.X 2021-12-20
CN202111566708.XA CN115577114A (zh) 2021-12-20 2021-12-20 一种基于时序知识图谱的事件检测方法和装置

Publications (1)

Publication Number Publication Date
WO2023115761A1 true WO2023115761A1 (zh) 2023-06-29

Family

ID=84579241

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/087213 WO2023115761A1 (zh) 2021-12-20 2022-04-15 基于时序知识图谱的事件检测方法和装置

Country Status (2)

Country Link
CN (1) CN115577114A (zh)
WO (1) WO2023115761A1 (zh)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116611523A (zh) * 2023-07-18 2023-08-18 山东师范大学 涡轮风扇发动机的可解释故障预测方法及系统
CN117094395A (zh) * 2023-10-19 2023-11-21 腾讯科技(深圳)有限公司 对知识图谱进行补全的方法、装置和计算机存储介质
CN117093727A (zh) * 2023-10-16 2023-11-21 湖南董因信息技术有限公司 基于时间关系感知的时序知识图谱补全方法
CN117493583A (zh) * 2024-01-03 2024-02-02 安徽思高智能科技有限公司 结合事件日志和知识图谱的流程操作序列生成方法及系统
CN117610513A (zh) * 2024-01-22 2024-02-27 南开大学 一种基于知识保护及选择的主题文本生成方法
CN117671979A (zh) * 2023-12-25 2024-03-08 中傲智能科技(苏州)有限公司 一种基于知识图谱的智慧城市数据管理系统与方法

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117151227B (zh) * 2023-08-28 2024-05-17 深圳智现未来工业软件有限公司 一种半导体检测结果的推理方法及装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200065668A1 (en) * 2018-08-27 2020-02-27 NEC Laboratories Europe GmbH Method and system for learning sequence encoders for temporal knowledge graph completion
CN111881219A (zh) * 2020-05-19 2020-11-03 杭州中奥科技有限公司 动态知识图谱补全方法、装置、电子设备和存储介质
CN112395423A (zh) * 2020-09-09 2021-02-23 北京邮电大学 递归的时序知识图谱补全方法和装置
CN112905738A (zh) * 2021-02-05 2021-06-04 中山大学 一种基于时态知识图谱推理的社交关系演变预测方法
CN113190654A (zh) * 2021-05-08 2021-07-30 北京工业大学 一种基于实体联合嵌入和概率模型的知识图谱补全方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200065668A1 (en) * 2018-08-27 2020-02-27 NEC Laboratories Europe GmbH Method and system for learning sequence encoders for temporal knowledge graph completion
CN111881219A (zh) * 2020-05-19 2020-11-03 杭州中奥科技有限公司 动态知识图谱补全方法、装置、电子设备和存储介质
CN112395423A (zh) * 2020-09-09 2021-02-23 北京邮电大学 递归的时序知识图谱补全方法和装置
CN112905738A (zh) * 2021-02-05 2021-06-04 中山大学 一种基于时态知识图谱推理的社交关系演变预测方法
CN113190654A (zh) * 2021-05-08 2021-07-30 北京工业大学 一种基于实体联合嵌入和概率模型的知识图谱补全方法

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116611523A (zh) * 2023-07-18 2023-08-18 山东师范大学 涡轮风扇发动机的可解释故障预测方法及系统
CN116611523B (zh) * 2023-07-18 2023-09-26 山东师范大学 涡轮风扇发动机的可解释故障预测方法及系统
CN117093727A (zh) * 2023-10-16 2023-11-21 湖南董因信息技术有限公司 基于时间关系感知的时序知识图谱补全方法
CN117093727B (zh) * 2023-10-16 2024-01-05 湖南董因信息技术有限公司 基于时间关系感知的时序知识图谱补全方法
CN117094395A (zh) * 2023-10-19 2023-11-21 腾讯科技(深圳)有限公司 对知识图谱进行补全的方法、装置和计算机存储介质
CN117094395B (zh) * 2023-10-19 2024-02-09 腾讯科技(深圳)有限公司 对知识图谱进行补全的方法、装置和计算机存储介质
CN117671979A (zh) * 2023-12-25 2024-03-08 中傲智能科技(苏州)有限公司 一种基于知识图谱的智慧城市数据管理系统与方法
CN117493583A (zh) * 2024-01-03 2024-02-02 安徽思高智能科技有限公司 结合事件日志和知识图谱的流程操作序列生成方法及系统
CN117493583B (zh) * 2024-01-03 2024-05-14 安徽思高智能科技有限公司 结合事件日志和知识图谱的流程操作序列生成方法及系统
CN117610513A (zh) * 2024-01-22 2024-02-27 南开大学 一种基于知识保护及选择的主题文本生成方法
CN117610513B (zh) * 2024-01-22 2024-04-02 南开大学 一种基于知识保护及选择的主题文本生成方法

Also Published As

Publication number Publication date
CN115577114A (zh) 2023-01-06

Similar Documents

Publication Publication Date Title
WO2023115761A1 (zh) 基于时序知识图谱的事件检测方法和装置
US11288444B2 (en) Optimization techniques for artificial intelligence
Wang et al. Images don’t lie: Duplicate crowdtesting reports detection with screenshot information
Stein et al. Intrinsic plagiarism analysis
US9146987B2 (en) Clustering based question set generation for training and testing of a question and answer system
US9230009B2 (en) Routing of questions to appropriately trained question and answer system pipelines using clustering
CN111506714A (zh) 基于知识图嵌入的问题回答
CN110163376B (zh) 样本检测方法、媒体对象的识别方法、装置、终端及介质
KR101452082B1 (ko) 리서치 미션 식별
JPH07295989A (ja) データを解析するためのインタプリタを形成する装置
CN111666766B (zh) 数据处理方法、装置和设备
US11669740B2 (en) Graph-based labeling rule augmentation for weakly supervised training of machine-learning-based named entity recognition
CN112184089B (zh) 试题难度预测模型的训练方法、装置、设备及存储介质
CN112420125A (zh) 分子属性预测方法、装置、智能设备和终端
CN114511085A (zh) 实体属性值的识别方法、装置、设备、介质及程序产品
Zhang et al. Detecting SMART Model Cognitive Operations in Mathematical Problem-Solving Process.
CN110826616A (zh) 信息处理方法及装置、电子设备、存储介质
WO2023000725A1 (zh) 电力计量的命名实体识别方法、装置和计算机设备
Li et al. HAPZSL: A hybrid attention prototype network for knowledge graph zero-shot relational learning
CN114639044A (zh) 标签确定方法、装置、电子设备及存储介质
US11989626B2 (en) Generating performance predictions with uncertainty intervals
CN116778210A (zh) 教学影像评价系统以及教学影像评价方法
CN109213830B (zh) 专业性技术文档的文档检索系统
CN112149743A (zh) 一种访问控制方法、装置、设备及介质
CN110633363A (zh) 一种基于nlp和模糊多准则决策的文本实体推荐方法