CN115221292A - Generating knowledge question-answering method and device - Google Patents
Generating knowledge question-answering method and device Download PDFInfo
- Publication number
- CN115221292A CN115221292A CN202110412948.8A CN202110412948A CN115221292A CN 115221292 A CN115221292 A CN 115221292A CN 202110412948 A CN202110412948 A CN 202110412948A CN 115221292 A CN115221292 A CN 115221292A
- Authority
- CN
- China
- Prior art keywords
- event
- question
- data set
- answer
- knowledge base
- Prior art date
- Legal status (The legal status 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 status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 42
- 238000012549 training Methods 0.000 claims abstract description 50
- 238000010276 construction Methods 0.000 claims abstract description 14
- 238000000605 extraction Methods 0.000 claims description 31
- 238000002372 labelling Methods 0.000 claims description 10
- 238000013528 artificial neural network Methods 0.000 claims description 9
- 238000013145 classification model Methods 0.000 claims description 7
- JEIPFZHSYJVQDO-UHFFFAOYSA-N iron(III) oxide Inorganic materials O=[Fe]O[Fe]=O JEIPFZHSYJVQDO-UHFFFAOYSA-N 0.000 claims description 5
- 238000004590 computer program Methods 0.000 claims description 4
- 238000012544 monitoring process Methods 0.000 claims description 3
- 230000001364 causal effect Effects 0.000 claims description 2
- 238000012545 processing Methods 0.000 claims description 2
- 230000001502 supplementing effect Effects 0.000 claims description 2
- 238000013518 transcription Methods 0.000 claims description 2
- 230000035897 transcription Effects 0.000 claims description 2
- 238000013136 deep learning model Methods 0.000 abstract description 3
- 239000002585 base Substances 0.000 description 44
- 238000005516 engineering process Methods 0.000 description 4
- 238000013473 artificial intelligence Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000003058 natural language processing Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000012458 free base Substances 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 230000009469 supplementation Effects 0.000 description 1
- 238000013519 translation Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation or dialogue systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/31—Indexing; Data structures therefor; Storage structures
- G06F16/316—Indexing structures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
- G06F40/35—Discourse or dialogue representation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Databases & Information Systems (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Software Systems (AREA)
- Human Computer Interaction (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
A method and a device for generating knowledge question-answering are characterized in that questions and answers in an appointed question-answering application scene are used as a question-answer pair, a question-answer pair data set is divided into a database construction data set and a training data set according to a proportion and used for automatic construction of a knowledge base and training of a question model respectively, and the question model and the knowledge base are used for generating the knowledge question-answering for new questions in the appointed question-answering application scene. The invention uses the question-answering system of the generating type, the answer got is closer to the natural language, more fluent, more rational; according to the method, the problem set can be automatically constructed by the event map without the help of a manually constructed knowledge base, and the database construction data set and the training data set are easily expanded; the method exceeds the similar prior art in the application of data in the actual field, and the whole set of method has more interpretability and modularized decoupling compared with an end-to-end deep learning model, and is more suitable for actual application scenes.
Description
Technical Field
The invention belongs to the technical field of artificial intelligence and natural language processing, relates to a knowledge base automatic construction technology in a question-answering system and a knowledge base-based generating type question-answering technology, and discloses a generating type knowledge question-answering method and device constructed based on an automatic knowledge base.
Background
With the rapid development of the information society, the corresponding artificial intelligence technology is also continuously improved. The question-answering system in artificial intelligence is a very wide application scenario, and simultaneously has very challenging and difficult-to-solve problems.
The question-answering system can be divided into: a selection-type question-answer, a search-type question-answer, and a generation-type question-answer. The first two forms can be modeled as classification models as well as ranking models, however, generating questions and answers requires a higher level of natural language understanding. Compared with the former two, the generative question-answering system can construct answer descriptions closer to human languages, and answers look more natural and intelligent.
The generative question-answering system generally adopts a Seq2Seq framework in deep learning, generates answers in a natural language form through a language model mechanism, and is suitable for being used in many scenes. The generative framework has good effect on the tasks of machine translation and text summarization in the field of natural language processing, however, the knowledge question answering needs additional knowledge as a supplement for deep learning, and has very precise requirements on the dependent knowledge, which is difficult to be precisely stored by the neural network. Knowledge base is often required as a support for the knowledge question answering to generate a proper answer.
The generation-type knowledge question-answering requires that a knowledge base is searched and partially selected, and then the selected knowledge is fused with a deep learning model, so that the model generates a meaningful and accurate answer under the support of the knowledge. Recent generative knowledge question-answering systems such as GenQA and CoreQA select knowledge points supporting the Seq2Seq model by making selections on the knowledge base. The method can only be used for answering complex questions depending on knowledge in a single-hop range on the knowledge base, and has high requirements on the structure and the integrity of the knowledge base.
The traditional knowledge question-answering system has rich knowledge base support for applicable question-answering range, such as knowledge base DBpedia, freeBase, wikiData and the like constructed based on Wiki Pedia. The knowledge bases can well support knowledge question answering of common sense types, and play a vital role in a traditional question answering system in the form of examination knowledge, such as Google and hundred-degree search engines. However, such a knowledge question-answering system is difficult to be adequate for the knowledge question-answering in a specific scene or some question-answering beyond the knowledge point form, such as the question-answering form which is more complicated and requires more deep understanding of the question and knowledge, such as cause and effect, time sequence and comparison. On the other hand, the existing common method usually needs to deal with the demand by manually constructing a knowledge base, and manpower and material resources are consumed. For application scenarios in specific fields, manually constructing the knowledge base is a very tedious, complex and difficult-to-amplify scheme.
Disclosure of Invention
The problem to be solved by the invention is to extract knowledge from data and automatically construct a knowledge base according to the problem answers of the knowledge question answers in the corresponding field under the condition of lacking the knowledge base, and then to generate the formula knowledge question answers according to the constructed knowledge base.
The technical scheme of the invention is as follows: a generating knowledge question-answering method comprises the steps of taking a question and an answer in a specified question-answering application scene as a question-answer pair, dividing a question-answer pair data set into a database building data set and a training data set according to a proportion, respectively using the database building data set and the training data set for automatic construction of a knowledge base and training of a question answering model, marking the training data set to obtain an event extraction data set, training the event extraction data set to obtain an extraction model, and extracting events and event relations of the database building data set to obtain an event map knowledge base; the method comprises the steps of conducting answer model training after a knowledge base is built, extracting events of the problems through a training data set, corresponding to event nodes of an event map knowledge base according to similarity, selecting sub-images on a map as problem-related sub-images through the event nodes, inputting a problem text and the selected sub-images into an answer model based on a Seq2Seq framework, outputting an answer text, monitoring output through standard answers of the training data set, and conducting training to obtain the answer model;
and finally, generating a knowledge question and answer for the new question under the specified question and answer application scene by the answer model and the knowledge base.
Further, the automatic construction of the knowledge base and the training of the answer model are specifically as follows:
the automatic construction part of the knowledge base comprises the following steps:
the method comprises the following steps: for the training data set, obtaining an event and an event relation corresponding to the question-answer pair through manual marking, and using the event and the event relation as knowledge points for constructing a knowledge base to obtain an event extraction data set; combining the events therein to carry out semantic relation labeling to obtain an event-to-semantic relation data set;
step two: training a BERT-based joint extraction model by using the labeled event extraction data set, wherein the joint extraction model performs event extraction and relationship extraction by a sequence labeling method;
step three: applying the combined extraction model obtained in the step two to a database building data set, and extracting an event set in the database building data set and a relation between events to obtain a basic event map;
step four: training a sentence pair classification model based on BERT (binary inverted transcription) by using events to a semantic relation data set, predicting each event pair by using the sentence pair classification model after forming the event pair by two events extracted in the step three to obtain a semantic relation between the event pairs, supplementing the obtained relation into a basic event map, adding a co-occurrence relation between every two events extracted by the same question-answer pair, enriching the relation between the events, and forming a knowledge base event map by using all the obtained events and event relations to construct a knowledge base;
the steps of the answering model training part are as follows:
step five: extracting on a training data set by using a joint extraction model to obtain an event corresponding to each problem in the training data set, using the event corresponding to the problem as a summary description of the problem, then predicting the problem event and all nodes in a knowledge base event map pairwise by a sentence pair classification model, selecting event nodes marked as a co-reference relationship, and enabling the summary description of the problem to be linked to the nodes of the knowledge base event map; the same processing is carried out on the answers in the training data set, so that the answer events are linked to the nodes of the knowledge base event map;
step six: according to the selected event nodes, running a Personalized PageRank algorithm on a knowledge base event map, and selecting sub-maps on the map as problem-related sub-maps;
step seven: inputting the question text and the selected subgraph into an answer model combining a graph neural network GNN and a copying mechanism based on a Seq2Seq framework, and outputting an answer text.
The invention also proposes a generative knowledge question answering device in which a computer program is configured, which computer program, when executed, implements the method described above.
The invention has the following beneficial effects: (1) The method uses a generating question-answering system, and the obtained answer is closer to the natural language and looks smoother and more reasonable; (2) According to the method, a knowledge base constructed manually is not needed, the event map can be automatically constructed for the question set, the database construction data set and the training data set can be easily expanded through expanding the question and answer to the data set, and compared with the method for manually constructing the whole knowledge base, the method can realize the automatic construction of the whole knowledge base only by marking partial data, and can realize the algorithm iteration of expanding the knowledge base by expanding the data set; (3) The method of the invention exceeds a series of latest retrieval-based and generation-based question-answering technologies in the application of data in the actual field, and the whole set of method has more interpretability and modular decoupling compared with an end-to-end deep learning model, and is more suitable for the actual application scene. Compared with the results of ROUGE values of the query and answer method BM25 and the SRMRS based on retrieval and the ROUGE values of the mBART25 based on the generated query and answer method on a test set, the results of the ROUGE values are as follows:
| Method | ROUGE-1 | ROUGE-2 | ROUGE-L |
| BM25 | 30.3 | 15.0 | 18.6 |
| SRMRS | 32.2 | 15.2 | 18.9 |
| mBART25 | 34.1 | 18.7 | 24.3 |
| the invention | 37.1 | 31.2 | 26.8 |
As can be seen from the test results, the test results of the invention are obviously superior to those of the existing method.
Drawings
FIG. 1 is an overall flow chart of the present invention.
Fig. 2 is a diagram of a map-based generated question-answering model of the answering module.
Detailed Description
The present invention will be further described with reference to the following examples and flow charts.
The application scenario of the embodiment is a fruit-type question and answer data set for geographic college entrance examination, and for the scenario, the manual construction of the knowledge base is very complicated and low in efficiency, and the method can well solve the problem. Taking the question and the answer thereof as a question-answer pair, splitting the question-answer pair in a question-answer pair data set into a knowledge base automatic construction part and an answer model training part according to the ratio of 3; the method comprises the steps of conducting answer model training after a knowledge base is built, extracting events of the problems through a training data set, corresponding to event nodes of an event map knowledge base according to similarity, selecting sub-images on the map as problem-related sub-images through the event nodes, inputting a problem text and the selected sub-images into an answer model based on a Seq2Seq framework, outputting the answer text by combining a graph neural network GNN and a copying mechanism, monitoring output by using standard answers of the training data set, and conducting training to obtain the answer model.
The following description will be made. First, the automatic construction part of the knowledge base is described:
the method comprises the following steps: and performing event labeling on 50% of the training data set, labeling all events in the questions and all events in the answers of each question-answer pair, and labeling context and causal relationship between the events to obtain an event extraction data set. And then, for the whole marked event set, marking the semantic relation between every two events by a rule and manual marking method, wherein the semantic relation comprises common reference, correlation, opposite and irrelevant, and obtaining an event-to-semantic relation data set.
Step two: a BERT-based joint extraction model is trained by event extraction data sets, represented as an 'event extraction and relation extraction model' module in FIG. 1, and descriptions of events and relations among events are obtained through output of sequence labels. Wherein, the BERT uses model parameters obtained by pre-training on Chinese corpus, the sequence labeling part indicates the beginning and the end of an event by a double-row BOE, and then the classification result of event relation is obtained by inputting the embedded representation of two events into a single classification layer.
Step three: and predicting each question-answer pair on the database building data set by using the trained combined extraction model to obtain an event and an event relation in each question-answer pair. And converting the obtained data into a graph structure, and combining all event nodes with the same text to obtain a primary event graph knowledge base, namely a basic event graph. Denoted in fig. 1 as "events and event relationships" module.
Step four: furthermore, a BERT sentence pair multi-classifier is trained by using the labeled event pair semantic relation data set, and the configuration adopted by the BERT part is the same as that of the joint extraction model. The sentence-pair multi-classifier is used for splicing two events and then directly inputting the two events into BERT to obtain embedded representation of the two events, and then inputting the embedded representation into a linear classification layer to obtain a final classification result. And using the classifier obtained by training for the basic event map, and predicting every two events once, thereby obtaining the semantic relation between any two events in the basic event map. And then, the semantic relations are added into the basic event graph, so that the connectivity between each node in the graph is enriched. Wherein all event pairs labeled as coreferences are merged such that an event of one meaning appears as a node in the graph. In addition, a co-occurrence relation is added between the events extracted from the same question and answer pair, the co-occurrence relation is also put into the basic event map, the relation supplementation is completed, and the automatically constructed event map knowledge base is obtained.
The following is the training portion of the answer model:
step five: for each question on the whole training data set, the previously obtained joint extraction model is applied, all events mentioned in the question are extracted, the events are used as a summary description of the question, and then the summary description is corresponding to nodes on the knowledge base graph. And classifying each node in the summary description event and the map once by using a BerT sentence trained before for the multiple classifiers, and then obtaining the map node most relevant to the summary description event, namely the map node linked with the problem event, by a heuristic similarity sorting algorithm obtained by fusing the fastText similarity and the ROUGE value in all the obtained display common-finger results. In some cases, it is the event node where the text is completely consistent. For the answers on the training data set, the graph nodes obtained by linking the answer events are obtained through the process.
Step six: starting from a graph node most relevant to the summary description event, according to the selected event node, a Personalized page rank Personalized PageRank algorithm is operated on the whole graph to obtain the weight of each point relative to the event nodes, then the first 500 nodes are selected from the graph nodes in a descending order to be used as a partial knowledge base sub-graph representation relevant to the problem. Only the subgraph needs to be used as the relevant knowledge range of the problem in the later process. And step five and step six complete the work of the sub-graph selection module in the figure 1.
Step seven: inputting the question text and the selected subgraph into an answer model based on a Seq2Seq framework, and outputting an answer text by combining a graph neural network GNN and a copy mechanism. The method comprises the following specific steps: and inputting the subgraph obtained in the step six and the natural language text of the question into a map-based generating question-answering model shown in figure 2. The model structure is as follows: firstly, coding a problem text and each event node on a subgraph through a BERT coder; then taking the result obtained by coding as initialization, and operating a 3-layer graph neural network GNN on the subgraph to obtain a further coding result; the output of each event node is judged in advance whether the node is an answer or not through a linear answer classifier, and the node is supervised by the answer event link map node obtained in the fifth step; sorting the sub-graph nodes according to the output of the linear answer classifier, and selecting the first 20 nodes as event nodes which are finally input into a Seq2Seq decoder; stitching the results, including the embedded representation of the neural network of the graph of the first 20 nodes, the text representation of the nodes and the question text; and finally, inputting the splicing result into a 6-layer decoder based on a Transformer decoder architecture to obtain a final output text. The standard answer text corresponding to the question in the training data set supervises the output of the decoder, so that the decoder can directly output the answer obtained by the model.
By the method, the knowledge base is automatically constructed based on the known questions and answers of the specific scene, and the generative question-answering is realized for the new questions of the scene.
Claims (8)
1. A generating knowledge question-answering method is characterized in that questions and answers in a specified question-answering application scene are used as a question-answer pair, a question-answer pair data set is divided into a database building data set and a training data set according to a proportion and used for automatic building of a knowledge base and training of a question answering model respectively, the training data set is marked to obtain an event extraction data set, an extraction model is obtained through training of the event extraction data set, and then events and event relations of the database building data set are extracted to obtain an event map knowledge base; the method comprises the steps of conducting answer model training after a knowledge base is built, extracting events of the problems through a training data set, corresponding to event nodes of an event map knowledge base according to similarity, selecting sub-images on a map as problem-related sub-images through the event nodes, inputting a problem text and the selected sub-images into an answer model based on a Seq2Seq framework, outputting an answer text, monitoring output through standard answers of the training data set, and conducting training to obtain the answer model;
and finally, generating a knowledge question and answer for the new question under the specified question and answer application scene by the answer model and the knowledge base.
2. The generative knowledge question-answering method according to claim 1, wherein the knowledge base automatic construction and answer model training specifically comprises:
the automatic construction part of the knowledge base comprises the following steps:
the method comprises the following steps: for the data set used for training, obtaining an event and an event relation corresponding to the question-answer pair through manual marking, and taking the event and the event relation as knowledge points for constructing a knowledge base to obtain an event extraction data set; combining the events therein to carry out semantic relation labeling to obtain an event-to-semantic relation data set;
step two: training a BERT-based combined extraction model by using an event extraction data set, wherein the combined extraction model performs event extraction and relation extraction by a sequence labeling method;
step three: applying the combined extraction model obtained in the step two to a database building data set, and extracting an event set in the database building data set and a relation between events to obtain a basic event map;
step four: training a sentence pair classification model based on BERT (binary inverted transcription) by using events to a semantic relation data set, predicting each event pair by using the sentence pair classification model after forming the event pair by two events extracted in the step three to obtain a semantic relation between the event pairs, supplementing the obtained relation into a basic event map, adding a co-occurrence relation between every two events extracted by the same question-answer pair, enriching the relation between the events, and forming a knowledge base event map by using all the obtained events and event relations to construct a knowledge base;
the steps of the answering model training part are as follows:
step five: extracting on a training data set by using a joint extraction model to obtain an event corresponding to each problem in the training data set, using the event corresponding to the problem as a summary description of the problem, then predicting the problem event and all nodes in a knowledge base event map pairwise by a sentence pair classification model, selecting event nodes marked as a co-reference relationship, and enabling the summary description of the problem to be linked to the nodes of the knowledge base event map; the same processing is carried out on the answers in the training data set, so that the answer events are linked to the nodes of the knowledge base event map;
step six: according to the selected event nodes, operating a Personalized PageRank algorithm on a knowledge base event map, and selecting sub-maps on the map as problem-related sub-maps;
step seven: inputting the question text and the selected subgraph into an answer model combining a graph neural network GNN and a copying mechanism based on a Seq2Seq framework, and outputting an answer text.
3. The method as claimed in claim 2, wherein 50% of the training data sets are subjected to event labeling, all events in questions and all events in answers are labeled for each question-answer pair, context and causal relationship between the events are labeled to obtain an event extraction data set, and then semantic relationship between every two events, including common index, correlation, reversal and irrelevance, is labeled for the whole labeled event set by a rule and manual labeling method to obtain an event-to-semantic relationship data set.
4. The method according to claim 2, wherein in the fourth step, all event pairs marked as co-reference are merged according to the semantic relationship between the event pairs.
5. The method according to claim 2, wherein in the fifth step, for event nodes of the coreference relationship, a heuristic similarity sorting algorithm obtained by fusing fastText similarity and ROUGE value is used to obtain graph nodes most relevant to the summary description event, namely graph nodes linked with the problem events.
6. The method for generating knowledge question-answering according to claim 2, wherein the sixth step is specifically as follows: for a problem event, starting from a graph node which is linked with the summary description, a Personalized page rank Personalized PageRank algorithm is operated on the whole graph to obtain the weight of each node relative to the linked graph node, then the nodes are sorted from large to small, the first N nodes are selected to be used as a partial knowledge base sub-graph representation related to the problem, and only the sub-graph is used as a related knowledge range of the problem in the subsequent process.
7. The generative knowledge question-answering method according to claim 2, wherein in step seven, the answer model is a map-based generative question-answering model having a structure of: firstly, coding a problem text and each event node on a subgraph through a BERT coder; then, taking the result obtained by coding as initialization, and operating a 3-layer graph neural network GNN on the subgraph to obtain a further coding result; the output of each event node is judged in advance whether the node is an answer or not through a linear answer classifier, and the node is supervised by the answer event link map node obtained in the fifth step; sorting the sub-graph nodes according to the output of the linear answer classifier, and selecting the first M nodes as event nodes which are finally input into a Seq2Seq decoder; splicing the obtained results, wherein the obtained results comprise the graph neural network embedded representation of the first M nodes, the text representation of the nodes and the problem text; and finally, inputting the splicing result into a 6-layer decoder based on a Transformer decoder architecture to obtain a final output text.
8. A generative knowledge question answering apparatus, wherein a computer program is provided in the apparatus, which computer program when executed implements the method of any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110412948.8A CN115221292A (en) | 2021-04-16 | 2021-04-16 | Generating knowledge question-answering method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110412948.8A CN115221292A (en) | 2021-04-16 | 2021-04-16 | Generating knowledge question-answering method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115221292A true CN115221292A (en) | 2022-10-21 |
Family
ID=83605660
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110412948.8A Pending CN115221292A (en) | 2021-04-16 | 2021-04-16 | Generating knowledge question-answering method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115221292A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116884393A (en) * | 2023-08-03 | 2023-10-13 | 北京中科深智科技有限公司 | Pressure-spring type multistage buffering generation type AI communication method |
-
2021
- 2021-04-16 CN CN202110412948.8A patent/CN115221292A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116884393A (en) * | 2023-08-03 | 2023-10-13 | 北京中科深智科技有限公司 | Pressure-spring type multistage buffering generation type AI communication method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP7064262B2 (en) | Knowledge graph understanding support system based on natural language generation technology | |
| CN107004000A (en) | A kind of language material generating means and method | |
| KR101835345B1 (en) | Concept graph expansion system based on knowledge base | |
| CN110188147B (en) | Knowledge graph-based document entity relationship discovery method and system | |
| DE102019001267A1 (en) | Dialog-like system for answering inquiries | |
| WO2020010834A1 (en) | Faq question and answer library generalization method, apparatus, and device | |
| Paulheim | Exploiting Linked Open Data as Background Knowledge in Data Mining. | |
| CN102955848A (en) | Semantic-based three-dimensional model retrieval system and method | |
| CN106156023A (en) | The methods, devices and systems of semantic matches | |
| CN113886567A (en) | Teaching method and system based on knowledge graph | |
| CN114238653B (en) | Method for constructing programming education knowledge graph, completing and intelligently asking and answering | |
| CN112597316A (en) | Interpretable reasoning question-answering method and device | |
| CN113779220A (en) | Mongolian multi-hop question-answering method based on three-channel cognitive map and graph attention network | |
| Nguyen et al. | Seagull: A bird’s-eye view of the evolution of technical games research | |
| Pavlić et al. | Graph-based formalisms for knowledge representation | |
| CN115982379A (en) | User portrait construction method and system based on knowledge graph | |
| Karkar et al. | Illustrate it! An Arabic multimedia text-to-picture m-learning system | |
| CN114911893A (en) | Method and system for automatically constructing knowledge base based on knowledge graph | |
| CN115221292A (en) | Generating knowledge question-answering method and device | |
| Zhou et al. | Assisting in requirements goal modeling: a hybrid approach based on machine learning and logical reasoning | |
| Feng et al. | ModelsKG: A Design and Research on Knowledge Graph of Multimodal Curriculum Based on PaddleOCR and DeepKE | |
| CN115269806A (en) | Question-answering method, electronic device and storage medium applied to mineral domain knowledge graph | |
| Algosaibi et al. | Using the semantics inherent in sitemaps to learn ontologies | |
| Miksatko et al. | What’s in a cluster? automatically detecting interesting interactions in student e-discussions | |
| Deng et al. | From the Construction of Knowledge System to the Generation of Learner Portrait Based on Knowledge Graph |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |