WO2022041294A1 - Procédé destiné à générer des questions en combinant un triplet et un type d'entité dans une base de connaissances - Google Patents
Procédé destiné à générer des questions en combinant un triplet et un type d'entité dans une base de connaissances Download PDFInfo
- Publication number
- WO2022041294A1 WO2022041294A1 PCT/CN2020/112924 CN2020112924W WO2022041294A1 WO 2022041294 A1 WO2022041294 A1 WO 2022041294A1 CN 2020112924 W CN2020112924 W CN 2020112924W WO 2022041294 A1 WO2022041294 A1 WO 2022041294A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- entity
- word
- triplet
- vector
- question
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/10—Text processing
- G06F40/166—Editing, e.g. inserting or deleting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
- G06F40/284—Lexical analysis, e.g. tokenisation or collocates
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
Definitions
- the invention relates to the field of natural language text generation in natural language processing technology, in particular to a method for generating problems combining triples and entity types in a knowledge base.
- Question generation is an extremely important task in the field of natural language processing. In recent years, there have been more and more studies on question generation in text generation. According to different data sources, existing methods can be divided into knowledge base-based question generation. , text-based question generation, image and text-based question generation.
- Serban et al. first proposed the use of recurrent neural networks to generate factual questions (Serban IV, Garcia-Duran A, Gulcehre C, et al. Generating Factoid Questions With Recurrent Neural Networks: The 30M Factoid Question-Answer Corpus[C]/ /Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics(Volume 1:Long Papers).2016:588-598.), based on this, Indurthi et al.
- Liu et al. proposed the existing knowledge base-based question generation field (Liu C, Liu K, He S, et al.
- the purpose of the present invention is to aim at the deficiencies of the prior art, consider part-of-speech tagging for each context by the method of part-of-speech tagging, so as to obtain the entity words contained in each context, and then obtain the word of each word through the ConceptNet network for the words of the context. Three-tuple information, and then pre-sequence the context word, context entity word and knowledge through the pre-trained Glove word embedding method to obtain the corresponding word vector.
- first encode the first sentence and its corresponding knowledge output the hidden state information, and add this output as an input to the knowledge encoding corresponding to the second round of harmony to obtain the hidden state information at this moment.
- the present invention is realized by at least one of the following technical solutions.
- a method for generating problems combining triples and entity types in a knowledge base comprises the following steps:
- the input of the reconstructed triplet model is the triplet and the corresponding head entity and tail entity in the triplet entity type, the output is a new set of triples based on the entity type;
- the training set can be generated by using public questions. , such as the SQuAD dataset;
- decode through a decoder composed of a gate control recurrent neural unit based on an attention mechanism, and obtain a new word vector sequence representing the new triple, thereby obtaining a set of word vector sequences representing the generated problem;
- the word vector sequence obtained in step 4) uses the word vector sequence obtained in step 4) to obtain the words represented by the vector, the word vector sequence is a matrix, each column in this matrix is a vector, each vector represents a word, and the value of each vector is The length is equal to the number of words in the entire vocabulary.
- the word corresponding to the largest dimension in the vector is the word represented by the vector.
- step 1) the step of reconstructing triple model comprises:
- E 1 and E 3 represent the head entity and the tail entity in a triple, respectively
- E 2 represents the defined relationship between E 1 and E 3
- E 4 and E 5 represent the entity types corresponding to E 1 and E 3 respectively;
- step 1.3) perform step 1.1) and step 1.2) iteratively in the original input data to obtain a new data set composed of new triples after triple reconstruction, and finally divide them into new training sets and test sets in proportion , the validation set.
- step 2) includes:
- step 3) includes:
- the question is related to the head entity and relationship in the triplet
- the answer to the question is the tail entity in the triplet
- three weights are calculated by the attention mechanism network to represent the relationship between the head entity, the head entity and the tail entity, and the importance of the tail entity in the triplet. The greater the weight obtained by the element in the triplet, the generation of the When using words, you should pay more attention to this element;
- v e1 , v e2 , v e3 respectively represent the head entity, the relationship between the head entity and the tail entity, and the tail entity in each triplet
- ⁇ s,t , ⁇ p,t , ⁇ o,t respectively represent At time t when the problem is generated, the head entity, the relationship between the head entity and the tail entity, and the weight of the tail entity are calculated by the attention mechanism network.
- the attention mechanism network is used to obtain the relationship between the entity, the head entity and the tail entity in the input triplet, and the weight of the tail entity, which includes: :
- s t-1 represents the representation of the words generated at time t-1
- v a , W a , U a respectively represent the weight matrices that can be trained in generating the attention mechanism network
- v e1 , v e2 , v e3 respectively represent the head entity, the relationship between the head entity and the tail entity, and the tail entity in each triplet
- h p, t represent the new vector representing the relationship
- h s, t represent the new vector representing the head entity
- h o, t represents the new vector representing the tail entity
- the attention mechanism network calculates three scalar weights to represent the head entity, tail entity, and tail entity respectively. Importance of entities and relationships:
- ⁇ s,t , ⁇ p,t , ⁇ o,t respectively represent the relationship between the head entity, the head entity and the tail entity, and the weight of the tail entity at the time t of generating the problem.
- step 4) includes:
- in and z t respectively represent the word embedding representation of words obtained by combining the representation of triples at time t-1 at time t, word embedding representation and one-hot encoding one-hot vector representation of the vector obtained through the fully connected network, s t Represents the generation at time t the words in the question;
- r t ⁇ (w r E w y t-1 +U r s t-1 +A r v s,t )
- step 5 includes:
- the present invention has the following advantages and beneficial effects:
- the present invention not only considers one-sided information, but also considers three important information: contextual content; contextual entity words and triplet knowledge information corresponding to each word. And combine the three information through a reasonable cumulative encoding method. Its beneficial effects: compared with the results obtained by the prior art, the present invention can generate ending sentences that are more in line with the trend of contextual plots.
- FIG. 1 is a flowchart of a method for generating a story ending combining contextual entity words and knowledge according to an embodiment of the present invention
- FIG. 2 is a design diagram of an overall model adopted in an embodiment of the present invention.
- this embodiment provides a method for generating problems combining triples and entity types in a knowledge base.
- the model diagram is shown in FIG. 2, including the following steps:
- the steps to reconstruct the triple model include:
- E 1 and E 3 represent the head entity and tail entity in a triple, respectively
- E 2 represents the defined relationship between E 1 and E 3
- E 4 and E 5 represent the entity types corresponding to E 1 and E 3 respectively;
- step 11) and step 12) iteratively in the original input data to obtain a new data set consisting of new triples after triple reconstruction, and finally divided into new training set and test set according to the proportion , the validation set.
- the input of the deep learning joint model is a word vector spliced together by vectors representing context, entity words and common sense knowledge respectively, and the output is a set of sequences related to the context.
- the specific steps of constructing the deep learning joint model include:
- step 1.3 input the word vector obtained in step 1.2) into the long-term memory network model in a step-by-step iterative manner;
- the model compares the output parameters of the attention mechanism model with the contextual entity word vector, such as when the input of the model is a triple (Obama, wife, Michelle), and entities Obama and Michelle, the model's
- the outgoing question is who is Obama's wife? , assuming that the words in the corpus are Ao, Ba, Ma, Wife, Zi, Mi, Xie, Er, De, Is, Who, then the word vector corresponding to Ao should be [1,0,0,0,0,0 ,0,0,0,0,0], the word vector corresponding to the Ba word should be [0,1,0,0,0,0,0,0,0,0,0,0], and suppose the result of the model output , the word vector corresponding to the Austrian word is [0.2,0.5,0.9,0,0,0,0,0,0,0,0], etc., the model will be based on the word vector [1,0, 0,0,0,0,0,0,0,0] adjustment, according
- step 1.5 iteratively execute step 1.5), when the difference between the accuracy of the long-short-term memory network model and the attention mechanism model parameters is stable, that is, when the fluctuation range is less than a certain range (usually a small value, such as 10e -5), get the final attention sequence-to-sequence deep learning joint model.
- a certain range usually a small value, such as 10e -5
- part-of-speech tagging tool uses the part-of-speech tagging tool to classify the words in the context by part-of-speech, and obtain the nouns and plural nouns contained therein;
- the obtained knowledge graph vector is combined with the context entity word to select the information of the triplet of more important words through the attention mechanism model.
- the specific process of selection is as follows:
- g(x) represents the knowledge graph vector
- hi , ri , t i represent the head entity, relationship, and tail entity of each word , respectively
- W r , W h , W t represent the learnable parameters for training relational entities, head entities, and tail entities
- tanh is As the hyperbolic tangent function of the activation function, ⁇ Ri refers to the representation of head entity, tail entity, relation entity, Etc. is a new representation obtained by normalization.
- the gate mechanism based on the attention mechanism is used to learn the ability of the context when encoding each sentence, and the final model outputs the state vector of the context hidden layer, which is obtained by splicing the state vector of the context hidden layer and the word vector of the context noun entity word.
- the final input vector as follows:
- the question is related to the head entity and relation in the triplet
- the answer to the question is the tail entity in the triplet
- the vector of the triplet is input into the attention mechanism network.
- the attention mechanism network three weights are calculated to represent the relationship between the head entity, the head entity and the tail entity, and the importance of the tail entity in the triplet. More attention should be paid to this element;
- a new representation of the triplet is obtained at each instant in which the problem is generated by weighted summation of the weight of each element in the resulting triplet and the vector of each element:
- v e1 , v e2 , v e3 respectively represent the head entity, the relationship between the head entity and the tail entity, and the tail entity in each triplet
- ⁇ s,t , ⁇ p,t , ⁇ o,t respectively represent At time t when the problem is generated, the head entity, the relationship between the head entity and the tail entity, and the weight of the tail entity are calculated by the attention mechanism network.
- a new word vector sequence representing the new triple is obtained, thereby obtaining a set of word vector sequences representing the generated problem, including:
- the representation of the word at time t-1 is combined with the representation of the triplet at time t to obtain the representation of the output word:
- in and z t respectively represent the word embedding representation of words obtained by combining the representation of triples at time t-1 at time t, word embedding representation and one-hot encoding one-hot vector representation of the vector obtained through the fully connected network, s t Represents the generation at time t the words in the question;
- the representation of the word at time t-1 is combined with the representation of the triplet at time t, and the representation of the output word is obtained through the gate mechanism recurrent neural network:
- r t ⁇ (w r E w y t-1 +U r s t-1 +A r v s,t )
- step 5 Input the hidden layer state vector in step 4) into the gate mechanism recurrent neural network based on the attention mechanism, and use the negative log-likelihood as the loss function to track the encoding and decoding stage, so that the final output is A set of context-dependent sequences, step 5) includes the following steps:
- the current attention sequence-to-sequence deep learning joint model is used as the best attention sequence-to-sequence deep learning joint model, and its specific formula is:
- Y t represents each word in the question
- E 1 and E 3 respectively represent a certain triple
- E 2 represents the relationship between E 1 and E 3
- E 4 and E 5 represent the entity types corresponding to E 1 and E 3 respectively;
- the entity type words that appear in the question Y that is, E 4 or E 5 , are replaced with E 1 and E 3 , respectively, to obtain a new question Y' as the final output.
Abstract
L'invention concerne un procédé destiné à générer des questions en combinant un triplet et un type d'entité dans une base de connaissances. Une entrée d'un modèle de réseau neuronal est une séquence de vecteurs de mots représentant un triplet reconstruit, et une sortie, obtenue en traitant la séquence de vecteurs de mots, est une séquence de vecteurs de mots servant à représenter une question. Le présent procédé consiste : à construire un nouveau triplet sur la base de types d'entités correspondant à une entité de tête et à une entité de queue dans un triplet; à utiliser le procédé d'intégration de mots GloVe pour obtenir une représentation du nouveau triplet et une question correspondant au nouveau triplet; à utiliser un réseau neuronal récurrent à mécanisme de portillonnage sur la base d'un mécanisme d'attention pour coder la représentation du nouveau triplet, et à sortir une représentation d'un triplet incorporant des informations contextuelles; à utiliser le réseau neuronal récurrent à mécanisme de portillonnage sur la base d'un mécanisme d'attention pour décoder la représentation du triplet, et à sortir une représentation de la question correspondant au triplet, sortant ainsi la question; et à remplacer les types d'entités dans la question sortie par le modèle par des entités spécifiques pour obtenir une nouvelle question. Selon le présent procédé, le triplet est combiné à des informations de types d'entités correspondant à l'entité de tête et à l'entité de queue dans le triplet dans une base de connaissances et, au moyen du modèle de réseau neuronal sur la base d'un mécanisme d'attention, une question plus fluide grammaticalement et mieux associée au triplet d'entrée est obtenue.
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010872496.7 | 2020-08-26 | ||
CN202010872496.7A CN112115687B (zh) | 2020-08-26 | 2020-08-26 | 一种结合知识库中的三元组和实体类型的生成问题方法 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022041294A1 true WO2022041294A1 (fr) | 2022-03-03 |
Family
ID=73804357
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2020/112924 WO2022041294A1 (fr) | 2020-08-26 | 2020-09-02 | Procédé destiné à générer des questions en combinant un triplet et un type d'entité dans une base de connaissances |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN112115687B (fr) |
WO (1) | WO2022041294A1 (fr) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114764566A (zh) * | 2022-04-11 | 2022-07-19 | 中国航空综合技术研究所 | 用于航空领域的知识元抽取方法 |
CN114942998A (zh) * | 2022-04-25 | 2022-08-26 | 西北工业大学 | 融合多源数据的知识图谱邻域结构稀疏的实体对齐方法 |
CN115936737A (zh) * | 2023-03-10 | 2023-04-07 | 云筑信息科技(成都)有限公司 | 一种确定建材真伪的方法和系统 |
CN116432750A (zh) * | 2023-04-13 | 2023-07-14 | 华中师范大学 | 一种基于盒嵌入的少样本知识图谱补全方法 |
CN117540035A (zh) * | 2024-01-09 | 2024-02-09 | 安徽思高智能科技有限公司 | 一种基于实体类型信息融合的rpa知识图谱构建方法 |
CN117540035B (zh) * | 2024-01-09 | 2024-05-14 | 安徽思高智能科技有限公司 | 一种基于实体类型信息融合的rpa知识图谱构建方法 |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113051929A (zh) * | 2021-03-23 | 2021-06-29 | 电子科技大学 | 一种基于细粒度语义信息增强的实体关系抽取的方法 |
CN113221571B (zh) * | 2021-05-31 | 2022-07-01 | 重庆交通大学 | 基于实体相关注意力机制的实体关系联合抽取方法 |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110647620A (zh) * | 2019-09-23 | 2020-01-03 | 中国农业大学 | 一种基于置信超平面和词典信息的知识图谱表示学习方法 |
US20200183963A1 (en) * | 2016-09-29 | 2020-06-11 | Koninklijke Philips N.V. | Question generation |
CN111368528A (zh) * | 2020-03-09 | 2020-07-03 | 西南交通大学 | 一种面向医学文本的实体关系联合抽取方法 |
CN111488440A (zh) * | 2020-03-30 | 2020-08-04 | 华南理工大学 | 一种基于多任务联合的问题生成方法 |
CN111563146A (zh) * | 2020-04-02 | 2020-08-21 | 华南理工大学 | 一种基于推理的难度可控问题生成方法 |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109308316B (zh) * | 2018-07-25 | 2021-05-14 | 华南理工大学 | 一种基于主题聚类的自适应对话生成系统 |
CN110502749B (zh) * | 2019-08-02 | 2023-10-03 | 中国电子科技集团公司第二十八研究所 | 一种基于双层注意力机制与双向gru的文本关系抽取方法 |
CN111159368B (zh) * | 2019-12-12 | 2023-04-28 | 华南理工大学 | 一种个性化对话的回复生成方法 |
CN111339774B (zh) * | 2020-02-07 | 2022-11-29 | 腾讯科技(深圳)有限公司 | 文本的实体关系抽取方法和模型训练方法 |
-
2020
- 2020-08-26 CN CN202010872496.7A patent/CN112115687B/zh active Active
- 2020-09-02 WO PCT/CN2020/112924 patent/WO2022041294A1/fr active Application Filing
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20200183963A1 (en) * | 2016-09-29 | 2020-06-11 | Koninklijke Philips N.V. | Question generation |
CN110647620A (zh) * | 2019-09-23 | 2020-01-03 | 中国农业大学 | 一种基于置信超平面和词典信息的知识图谱表示学习方法 |
CN111368528A (zh) * | 2020-03-09 | 2020-07-03 | 西南交通大学 | 一种面向医学文本的实体关系联合抽取方法 |
CN111488440A (zh) * | 2020-03-30 | 2020-08-04 | 华南理工大学 | 一种基于多任务联合的问题生成方法 |
CN111563146A (zh) * | 2020-04-02 | 2020-08-21 | 华南理工大学 | 一种基于推理的难度可控问题生成方法 |
Non-Patent Citations (1)
Title |
---|
SONG ZEHAN: "Multi-Source Co-Attention Networks for Composite Question Generation", MASTER THESIS, TIANJIN POLYTECHNIC UNIVERSITY, CN, 1 August 2020 (2020-08-01), CN , pages 1 - 75, XP055905924, ISSN: 1674-0246, DOI: 10.27461/d.cnki.gzjdx.2020.000454 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114764566A (zh) * | 2022-04-11 | 2022-07-19 | 中国航空综合技术研究所 | 用于航空领域的知识元抽取方法 |
CN114764566B (zh) * | 2022-04-11 | 2024-01-23 | 中国航空综合技术研究所 | 用于航空领域的知识元抽取方法 |
CN114942998A (zh) * | 2022-04-25 | 2022-08-26 | 西北工业大学 | 融合多源数据的知识图谱邻域结构稀疏的实体对齐方法 |
CN114942998B (zh) * | 2022-04-25 | 2024-02-13 | 西北工业大学 | 融合多源数据的知识图谱邻域结构稀疏的实体对齐方法 |
CN115936737A (zh) * | 2023-03-10 | 2023-04-07 | 云筑信息科技(成都)有限公司 | 一种确定建材真伪的方法和系统 |
CN115936737B (zh) * | 2023-03-10 | 2023-06-23 | 云筑信息科技(成都)有限公司 | 一种确定建材真伪的方法和系统 |
CN116432750A (zh) * | 2023-04-13 | 2023-07-14 | 华中师范大学 | 一种基于盒嵌入的少样本知识图谱补全方法 |
CN116432750B (zh) * | 2023-04-13 | 2023-10-27 | 华中师范大学 | 一种基于盒嵌入的少样本知识图谱补全方法 |
CN117540035A (zh) * | 2024-01-09 | 2024-02-09 | 安徽思高智能科技有限公司 | 一种基于实体类型信息融合的rpa知识图谱构建方法 |
CN117540035B (zh) * | 2024-01-09 | 2024-05-14 | 安徽思高智能科技有限公司 | 一种基于实体类型信息融合的rpa知识图谱构建方法 |
Also Published As
Publication number | Publication date |
---|---|
CN112115687A (zh) | 2020-12-22 |
CN112115687B (zh) | 2024-04-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2022041294A1 (fr) | Procédé destiné à générer des questions en combinant un triplet et un type d'entité dans une base de connaissances | |
CN113255755B (zh) | 一种基于异质融合网络的多模态情感分类方法 | |
CN111312245B (zh) | 一种语音应答方法、装置和存储介质 | |
Firdaus et al. | EmoSen: Generating sentiment and emotion controlled responses in a multimodal dialogue system | |
CN114694076A (zh) | 基于多任务学习与层叠跨模态融合的多模态情感分析方法 | |
CN115329779B (zh) | 一种多人对话情感识别方法 | |
Liu et al. | Reinforcement learning for emotional text-to-speech synthesis with improved emotion discriminability | |
CN111125333B (zh) | 一种基于表示学习与多层覆盖机制的生成式知识问答方法 | |
US11475225B2 (en) | Method, system, electronic device and storage medium for clarification question generation | |
Chen et al. | Delving deeper into the decoder for video captioning | |
Tang et al. | Modelling student behavior using granular large scale action data from a MOOC | |
CN113536804B (zh) | 一种基于关键词强化的GRU和Kronecker的自然语言特征提取方法 | |
CN114242071A (zh) | 一种低资源语音识别方法、系统、语音模型训练方法 | |
Lin et al. | Sentiment analysis of low-carbon travel APP user comments based on deep learning | |
EP4235485A1 (fr) | Procédé de conversion de données textuelles en caractéristique acoustique, dispositif électronique et support de stockage | |
Wang et al. | Empathetic response generation through graph-based multi-hop reasoning on emotional causality | |
CN111353040A (zh) | 基于gru的属性级别情感分析方法 | |
Sun et al. | Emotional conversation generation orientated syntactically constrained bidirectional-asynchronous framework | |
Xu et al. | Generating emotional controllable response based on multi-task and dual attention framework | |
CN114743056A (zh) | 一种基于动态早退的图像描述生成模型及模型训练方法 | |
CN116150334A (zh) | 基于UniLM模型和Copy机制的中文共情语句训练方法及系统 | |
Sakti et al. | Incremental sentence compression using LSTM recurrent networks | |
CN111813907A (zh) | 一种自然语言问答技术中的问句意图识别方法 | |
Miao et al. | Multi-turn dialogue model based on the improved hierarchical recurrent attention network | |
Gao et al. | Dialogue Summarization with Static-Dynamic Structure Fusion Graph |
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: 20950963 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: 20950963 Country of ref document: EP Kind code of ref document: A1 |