CN113505209A - Intelligent question-answering system for automobile field - Google Patents
Intelligent question-answering system for automobile field Download PDFInfo
- Publication number
- CN113505209A CN113505209A CN202110778221.1A CN202110778221A CN113505209A CN 113505209 A CN113505209 A CN 113505209A CN 202110778221 A CN202110778221 A CN 202110778221A CN 113505209 A CN113505209 A CN 113505209A
- Authority
- CN
- China
- Prior art keywords
- module
- query
- question
- dictionary
- matching
- 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
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
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3325—Reformulation based on results of preceding query
- G06F16/3326—Reformulation based on results of preceding query using relevance feedback from the user, e.g. relevance feedback on documents, documents sets, document terms or passages
- G06F16/3328—Reformulation based on results of preceding query using relevance feedback from the user, e.g. relevance feedback on documents, documents sets, document terms or passages using graphical result space presentation or visualisation
-
- 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
- G06F16/353—Clustering; Classification into predefined classes
-
- 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
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- 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
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Animal Behavior & Ethology (AREA)
- Human Computer Interaction (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- General Health & Medical Sciences (AREA)
- Machine Translation (AREA)
Abstract
The invention discloses an intelligent question-answering system for the automobile field, which comprises: the system comprises a knowledge base module, a visual interaction module, an intention identification module, a graph matching module, a template matching module, a retrieval module and an end-to-end module; the knowledge base module stores a knowledge map and a corpus of the automobile field; after the user inputs the question, the input content of the user is judged, and the corresponding module of the system is called to process according to different judged user purposes to obtain the answer of the question. Dividing the user purpose into two categories of automobile field questioning and chatting, and aiming at the automobile field questioning, obtaining question answers by using a question answering method based on an automobile field knowledge map; an answer is generated for the chat using an end-to-end module based on deep learning. The invention can improve the classification precision and accurately identify the intention of the user.
Description
Technical Field
The invention relates to the technical field of intelligence, in particular to an intelligent question answering system for the field of automobiles.
Background
At present, artificial intelligence is undoubtedly one of the most popular research contents, and as natural language processing of one main direction of artificial intelligence, one of the application modes is intelligent question answering. The intelligent question-answering means that the computer automatically answers the questions asked by the user by analyzing the questions asked by the user, and the intelligent question-answering system is an advanced information service.
In recent years, various internet companies have gradually introduced and developed their own intelligent question-answering systems, such as "Siri" of apple, "asistat" of Google, "Alexa" of amazon, "alei honey" of acriba, and "lover classmates" of millet. Although the existing question-answering method at the present stage can well complete the human-computer interaction problem, the problems such as unclear semantic recognition, wrong or meaningless answer generation, overlong system response time and the like still exist.
A knowledge graph is essentially a structured semantic knowledge base, a graph data structure consisting of nodes and edges. Many important links related to language understanding, information query, language organization and the like in the intelligent question-answering technology all need to be guided by language knowledge, common knowledge and domain knowledge. The knowledge graph is very suitable for being injected into the intelligent question-answering technology as an external knowledge source, for example, the knowledge graph can be used for assisting in understanding of the question, and by means of the attributes and the relations of the nodes in the knowledge graph, the entities in the question are found through the corresponding technology, so that the user problems are better understood. In addition, at the decoding stage of the end-to-end model, relevant entities can be retrieved from the knowledge graph as answers, and the entity answers are spliced with the text answers to form replies.
At present, an intelligent question-answering system in the field of automobiles has single constraint on part-of-speech recognition conditions of user input words, recognition accuracy is low, semantic structure information of input questions and dependency relationships among words are not considered, shallow semantic analysis is only realized, and complicated questions are difficult to answer.
Therefore, how to provide an intelligent question-answering system oriented to the automobile field, which introduces a knowledge graph and can significantly improve the quality and efficiency of intelligent question-answering, is a technical problem that needs to be solved urgently by technical personnel in the field.
Disclosure of Invention
In view of the above, the invention provides an intelligent question-answering system for the automobile field, which divides the user input questions into two types of questions of the automobile field and chatting, and obtains question answers by using a question-answering method based on a knowledge graph of the automobile field aiming at the questions of the automobile field; for chatting, an end-to-end module based on deep learning is used to generate answers, so that the classification precision can be improved, and the user intention can be accurately identified.
In order to achieve the purpose, the invention adopts the following technical scheme:
an intelligent question-answering system for the automobile field comprises: the system comprises a knowledge base module, a visual interaction module, an intention identification module, a graph matching module, a template matching module, a retrieval module and an end-to-end module;
the knowledge base module is used for storing a knowledge map and a corpus of the automobile field;
the visual interaction module is used for a user to input questions and feed back answers to the questions;
the intention identification module is used for judging whether the type of the user input question is an automobile field question or a chatting question, inputting the automobile field question to the graph matching module and inputting the chatting question to the end-to-end module;
the map matching module is used for matching the automobile field questions with the knowledge map in the knowledge base module by using a map matching method, and if the matching is successful, the answers to the questions are fed back to the visual interaction module; if the matching fails, inputting the automobile field problem into the template matching module;
the template matching module is used for matching automobile field problems with problem templates predefined in the knowledge base module by using a template matching method, and if the matching is successful, the problem answers are fed back to the visual interaction module; if the matching fails, inputting the automobile field problem into the retrieval module;
the retrieval module is used for retrieving answers of the automobile field questions in the corpus, and if the retrieval is successful, the answers of the questions are fed back to the visual interaction module; if the retrieval fails, inputting the automobile field problem into the end-to-end module;
the end-to-end module is used for identifying automobile field questions or chatting questions according to a pre-established deep learning model, generating question answers and feeding the question answers back to the visual interaction module.
Preferably, in the above intelligent question-answering system for the automobile field, the knowledge map includes: triple knowledge generated after processing of structured data related to the automobile field; the corpus at least comprises chatting corpus, data sets and unstructured text data related to the automobile field.
Preferably, in the above intelligent question-answering system for the automobile field, the visual interaction module is a web interaction interface based on a Django framework or a WeChat interaction interface based on itChat.
Preferably, in the above intelligent question-answering system for the automobile field, the intention identification module is configured to perform tagging and word segmentation operations on a current input question of a user by using a pre-trained FastText classification model, and determine that the type of the input question is an automobile field question or a chatting question.
Preferably, in the above intelligent question answering system for the automobile field, the graph matching module includes:
the dictionary module is used for extracting and constructing dictionaries of different types from the knowledge graph in advance; the dictionary includes: the system comprises an entity dictionary, a relation dictionary, an attribute value dictionary, a type dictionary generated manually in advance, a query word dictionary and a stop word dictionary;
the dictionary tree module is used for correspondingly generating a dictionary tree according to the content of each dictionary;
the preprocessing module is used for removing punctuation marks and carrying out capital and lower English case conversion processing on the problems input by the user;
the matching module is used for matching each node of the dictionary tree downwards from a root node according to the character sequence until the preprocessed user input problem cannot be matched;
the word segmentation and part-of-speech tagging module is used for segmenting words of the matching result output by the matching module and tagging the word segmentation result in terms of speech;
the dependency tree construction module is used for analyzing semantic association among all parts in the word segmentation result and establishing a dependency tree according to the semantic association;
the node type judging module is used for traversing each node of the dependency tree in the dictionary of each type and determining the type of each node; if the word corresponding to the node appears in the stop word dictionary, setting the node as a stop word node; if the editing distance similarity between the words corresponding to the nodes and the words in the entity dictionary or the relation dictionary is larger than a set threshold value, setting the nodes as entity query nodes or relation query nodes;
the pronoun resolution module is used for finding out a node of which the node type is a pronoun in the dependency tree according to the part of speech tagging result, finding out an entity node closest to the node, and taking a word corresponding to the entity node as a specific object referred by the pronoun;
the intention identification submodule is used for finding out the query word nodes in the dependency tree and finding out the attribute nodes or the relationship nodes which are closest to the nodes as the intention;
the query graph building module is used for initializing a query graph set, then traversing all entity query nodes found from the dependency tree, finding out other entity query nodes closest to the entity query node for each entity query node, obtaining the shortest paths of the other entity query nodes in the dependency tree, and finally adding head and tail nodes and paths between the two nodes into the query graph set to obtain a query graph;
the graph matching sub-module is used for traversing each node in the query graph in a breadth-first mode, converting an edge starting from each node in the query graph into a graph database query statement, and if a result can be queried in a graph database, indicating that the edge is successfully matched; if the edge matching is successful, adding the edge into the query subgraph; judging whether the query subgraph is the generation subgraph of the query graph; if the query subgraph is generated, the query subgraph is successfully matched with the knowledge graph, edges corresponding to the relation nodes identified by the intention identification submodule are converted into query sentences, and query results are used as matching answers of the query subgraph and stored in a result set; generating subgraphs, wherein the generated subgraphs comprise all vertex subgraphs in the query graph;
and the answer ranking module is used for ranking the matched answers obtained by the graph matching sub-module, adding the number of the query graph edges and the sum of the similarity of all the nodes of the query graph to calculate a score, and taking the first N answers as a final result.
Preferably, in the above intelligent question-answering system for the automobile field, the template matching module includes:
the template predefining module is used for predefining templates with some problems, setting a trigger word for each template, setting a template slot position to be filled, and if the trigger word is identified in the problems input by the user, considering that the corresponding template is matched;
the trigger word and key information identification module is used for carrying out word segmentation and semantic analysis operations on the problems input by the user, identifying the trigger words and key information in the problems and matching the trigger words and key information with corresponding templates;
the template slot filling module is used for filling the trigger words and the key information identified by the key information identification module into preset template slots according to the templates matched with the trigger words to generate complete template query sentences;
and the query execution module is used for generating a graph database query statement by using the complete template query statement, and matching the knowledge graph by using the graph database query statement to obtain a question answer.
Preferably, in the above intelligent question-answering system for the automobile field, the retrieval module includes:
the index file establishing module is used for splitting original data in the corpus, storing the split data into a vocabulary table, obtaining an index value of the data in the splitting process, and establishing an index file according to the index value;
the word segmentation module is used for segmenting words input by a user according to Chinese words by adopting an IKAnalyzer word segmentation device to obtain keywords in the problems;
the retrieval ordering module is used for searching keywords obtained by the word segmentation result in the index file by adopting a Lucene full-text information retrieval engine, and retrieving the keywords from the vocabulary firstly during the searching and then entering the original data for retrieving after the retrieval is successful; and sorting the different search results according to the relevancy scores, taking the top 5 results with the highest scores as answers to the questions, and returning the answers to the visual interaction module.
Preferably, in the above intelligent question-answering system for the automobile field, the end-to-end module includes:
the model building module is used for building a deep learning model by adopting a word embedding layer and a bidirectional gating cyclic unit network;
the word frequency dictionary generating module is used for taking the corpus as a training data set in the training stage of the deep learning model, and performing word segmentation and word frequency statistics on the training data set in sequence to generate a word frequency dictionary;
the serialization operation module is used for generating a serialization dictionary and an anti-serialization dictionary according to the word frequency dictionary, carrying out serialization operation on the training data by using the serialization dictionary, converting the text of the training data into a digital sequence, and inputting the digital sequence of the training data into the deep learning model; or in the testing stage of the deep learning model, the problems input by the user are serialized, the texts of the problems are converted into digital sequences, and the digital sequences of the problems are input into the deep learning model for recognition;
and the deserialization operation module is used for deserializing the question answer sequence identified and output by the deep learning model, converting the digital sequence into a text sequence and obtaining the question answer in a text form.
Preferably, in the above intelligent question-answering system for the automobile field, the deep learning model comprises an encoder and a decoder; both the encoder and the decoder are constructed based on a word embedding layer and a bidirectional gating cyclic unit network; attention mechanisms have also been introduced into the encoder.
Preferably, in the above intelligent question-answering system for the automobile field, the deep learning model generates a loss value between the currently output answer and the real answer through loss function calculation in a training stage, and the convergence rate of the deep learning model is increased by adopting an optimization algorithm to adjust parameters of the deep learning model and combining a Teacher Forcing mechanism until the loss value is smaller than a set threshold, and the training of the deep learning model is finished.
Compared with the prior art, the invention discloses and provides an intelligent question-answering system for the automobile field, which has the following beneficial effects:
1. according to the method, the intention of the user problem is identified through the deep learning model, compared with the traditional method, the deep learning model is higher in classification precision, and the accuracy of judgment on the intention of the user can be improved.
2. The graph matching method used by the invention firstly converts the question of the user into the dependency tree, then converts the structure of the dependency tree into the structure of the query graph, and finally matches the knowledge graph with the query graph to obtain the answer of the question in the knowledge graph. The step of converting into the dependency tree can fully consider semantic information of the user, can effectively answer complex questions, and the step of converting into the query graph and subsequent graph matching can accurately obtain answers of the user questions by utilizing huge knowledge information of a knowledge graph while retaining the semantic information of the user.
3. The image matching module and the template matching module of the invention rely on the knowledge graph, the answers are required to be obtained from the knowledge graph, the retrieval module retrieves the answers from the corpus, and the end-to-end module uses the trained deep learning model to directly generate the answers according to the input contents of the user. The four modules are different in the question-answering method and the knowledge base on which the question-answering depends and complement each other, so that the capability and the range of the system for answering the questions are greatly improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.
FIG. 1 is a block diagram of an overall structure of an intelligent question answering system for the automobile field according to the present invention;
FIG. 2 is a schematic diagram of the structure of a knowledge-graph provided by the present invention;
FIG. 3 is a diagram illustrating an example of a web-based interactive interface provided by the present invention;
FIG. 4 is a diagram illustrating an exemplary WeChat interaction interface provided by the present invention;
FIG. 5 is a flow diagram illustrating the identification of an intent recognition module provided by the present invention;
FIG. 6 is a diagram matching flow diagram of a diagram matching module provided by the present invention;
FIG. 7 is a diagram illustrating a structure of a trie according to the present invention;
FIG. 8 is a diagram illustrating a structure of a dependency tree according to the present invention;
FIG. 9 is a schematic diagram of a query graph according to the present invention;
FIG. 10 is a flow chart illustrating the matching of the template matching module provided by the present invention;
FIG. 11 is a retrieval flow diagram of a retrieval module provided by the present invention;
FIG. 12 is a flow chart illustrating the training and testing of an end-to-end module according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
As shown in fig. 1, an embodiment of the present invention discloses an intelligent question-answering system for the automobile field, which includes: the system comprises a knowledge base module, a visual interaction module, an intention identification module, a graph matching module, a template matching module, a retrieval module and an end-to-end module;
the knowledge base module is used for storing a knowledge map and a corpus of the automobile field;
the visual interaction module is used for inputting questions and feeding back answers to the questions by a user;
the intention identification module is used for judging whether the type of the user input question is an automobile field question or a chatting question, inputting the automobile field question to the graph matching module, and inputting the chatting question to the end-to-end module;
the map matching module is used for matching the automobile field questions with the knowledge map in the knowledge base module by using a map matching method, and if the matching is successful, the answers to the questions are fed back to the visual interaction module; if the matching fails, inputting the automobile field problem into a template matching module;
the template matching module is used for matching the automobile field problems with the problem templates predefined in the knowledge base module by using a template matching method, and if the matching is successful, the problem answers are fed back to the visual interaction module; if the matching fails, inputting the automobile field problem into a retrieval module;
the retrieval module is used for retrieving answers of the automobile field questions in the corpus, and if the retrieval is successful, the answers of the questions are fed back to the visual interaction module; if the retrieval fails, inputting the automobile field problem into an end-to-end module;
the end-to-end module is used for identifying automobile field questions or chatting questions according to a pre-established deep learning model, generating question answers and feeding back the question answers to the visual interaction module.
Specifically, the knowledge base module comprises two parts, namely a knowledge graph and a corpus. The data is mainly obtained by crawling a vertical field website in the automobile field through a crawler technology, and specifically comprises structured data and unstructured text corpora related to the automobile field. The structured data is processed to generate triple knowledge, the triple knowledge is stored in a map database to form a knowledge map, and the knowledge map is used by a map matching module and a template matching module as shown in fig. 2. The unstructured text data is stored in a relational database and used as text corpora for training a retrieval module and an end-to-end module. The corpus also contains some open-source chatting corpora, data sets and the like, and the data are mainly used for training an end-to-end module and an intention recognition module.
The embodiment of the invention designs two visual interfaces for interaction, namely a web interactive interface based on a Django framework and a WeChat interactive interface based on itChat. The web interactive interface may allow the user to enter questions and present system-generated answers to the user, as shown in FIG. 3; the WeChat interactive interface can access the personal WeChat to the intelligent question-answering system by calling the itChat interface, and provides a function similar to a chat robot, as shown in FIG. 4.
The intention recognition module is used for performing labeling and word segmentation operations on the current input problem of the user by using a pre-trained FastText text classification model, and judging the type of the input problem as an automobile field problem or a chatting problem. The training and testing process is shown in fig. 5, and specifically includes:
1. in the training process of the text classification model, the questions of a question-answer board of a website in the automobile field are used as positive example data, chatting linguistic data are used as negative example data, the data are labeled and participled, and then the data are input into a FastText text classification model for training to obtain a trained model;
2. in the testing stage, when a user inputs a question, the trained model is used for identifying the intention of the user input, and the user question is submitted to the graph matching module or the visual interaction module for processing according to the classification result.
The matching process of the graph matching module to the user input problem is shown in fig. 6, and the graph matching module includes: the device comprises a dictionary module, a dictionary tree module, a preprocessing module, a matching module, a word segmentation and part of speech tagging module, a dependency tree construction module, a node type judgment module, a pronoun resolution module, an intention recognition sub-module, a query graph construction module, a graph matching sub-module and an answer ordering module.
The dictionary building and the dictionary tree building are carried out before the problem input, belong to an off-line working part, and other modules are carried out after the problem input by a user, and belong to an on-line working part. The method comprises the following specific steps:
1. constructing a dictionary: the dictionary module extracts and constructs dictionaries of different types from the knowledge graph; the dictionary includes: the system comprises an entity dictionary, a relation dictionary, a property value dictionary, a type dictionary generated manually in advance, a query word dictionary and a stop word dictionary.
2. Constructing a dictionary tree: the trie module generates a trie from the contents of the different types of dictionaries constructed in step 1, as shown in fig. 7, which is a trie composed of the character strings "speed S level", "speed E level", "speed", "bmw", "bme X5", and "bme X6".
3. Problem pretreatment: from here on the online part is entered, i.e. a question input by the user is required. When a user inputs a question, the preprocessing module preprocesses the input question, such as removing punctuation marks, converting English into lower case, and the like.
4. Longest matching based on dictionary trees: and the matching module performs longest matching on the user question based on the dictionary tree obtained in the step 2. Specifically, the nodes of the dictionary tree are matched downwards according to the character sequence from the root node until the nodes cannot be matched.
5. Word segmentation and part-of-speech tagging: in order to implement the subsequent graph matching method, word segmentation needs to be carried out on the question of the user, in the process, the word segmentation and part-of-speech tagging module transmits the matching result obtained in the step 4 as an additional word segmentation dictionary into the word segmentation device, the word segmentation accuracy is improved, and part-of-speech tagging is carried out on the word segmentation result after word segmentation.
6. Building a dependency tree: the dependency tree is also called semantic dependency analysis, and the dependency tree construction module analyzes semantic association among all parts of the sentence and presents the semantic association in a tree structure. And building a dependency tree for the result obtained in the step 5. As shown in fig. 8, a dependency tree corresponding to the question "how much the displacement of the galloping car is".
7. Judging the node type: the node type judging module traverses each node of the dependency tree in the dictionary obtained in the step 1, and if the word corresponding to the node appears in the stop word dictionary, the node is set as a stop word node correspondingly; and if the editing distance similarity between the word corresponding to the node and the word in the entity dictionary or the relation dictionary is greater than a set threshold value, setting the node as an entity query node or a relation query node correspondingly.
8. And (3) pronoun resolution: and finding out the specific object referred by the pronoun by utilizing the dependency tree. If the input question contains pronouns, the step is executed, otherwise, the step is skipped. The specific method comprises the following steps: firstly, finding out a node with a node type as a pronoun in a dependency tree according to a part-of-speech tagging result obtained by a word segmentation and part-of-speech tagging module, then finding out an entity node closest to the node, and taking a word corresponding to the entity node as a specific object referred by the pronoun;
9. intention recognition: for identifying the intent of the user to input a question. The specific method comprises the following steps: the intention identification module firstly finds out the query word nodes in the dependency tree and then finds out the attribute nodes or relationship nodes closest to the nodes as the intention;
10. constructing a query graph: the query graph construction module initializes a set of query graphs. First, all entity query nodes found in the dependency tree are traversed, for each entity query node, all other entity query nodes closest to the node are found, and then the shortest paths of the other entity query nodes in the dependency tree are obtained. And finally adding the head node and the tail node and the path between the two nodes into the query graph set. As shown in fig. 9, the dependency tree corresponding to the question "how much the displacement of the car is" is converted into the result of the query graph;
11. and (3) matching the graphs: the graph matching module traverses the nodes in the query graph in a breadth-first mode, converts the edges starting from each node in the graph into graph database query statements, and if the results can be queried in a database, the edges are successfully matched. And if the edge matching is successful, adding the edge into the query subgraph. And judging whether the query subgraph is a generated subgraph of the query graph, wherein the generated subgraph contains all the vertexes in the query graph. If the generated subgraph is the query subgraph, the query subgraph is successfully matched with the knowledge graph. Converting the edges corresponding to the relation nodes identified by the intention graph in the step 9 into query statements, taking the query results as matching answers of the query subgraphs, and storing the query results into a result set;
12. and (3) answer sorting: and the answer sorting module sorts the answers generated by the graph matching module, adds the number of the query graph edges and the sum of the similarity of all the nodes of the query graph to calculate a score, and takes the first N answers as a final result.
The template matching module comprises: the template pre-defining module, the trigger word and key information identifying module, the template slot filling module and the query executing module, wherein the implementation steps of the modules are as shown in fig. 10, and specifically are as follows:
1. pre-defining a template: templates with problems are predefined by a template predefining module, a trigger word is set for each template, and a template slot position to be filled is set. If a trigger word is identified in the user question, the corresponding template is considered to be matched. For example "[ brand ]? Which are "(and? ", respectively, corresponding to a question asking for a brand vehicle family and a question asking for a car selling price;
2. identifying trigger words and key information: and performing operations such as word segmentation, semantic analysis and the like on the question of the user by using a trigger word and key information identification module, identifying the trigger words and key information in the question, and further matching the trigger words and the key information to a corresponding template, wherein the key information is also used for filling the template slot positions. What are the question "car series that is speeding? ", the trigger" train "and the key information" speed "will be identified;
3. filling template slots: the template slot filling module fills the key information identified in the step 2 into a preset template slot according to the template matched with the trigger word, so as to generate a complete template query statement;
4. and executing by using a query statement: and the query execution module generates and executes the graph database query statement by using the complete template query statement, and obtains the question answer in the knowledge graph.
The retrieval module comprises: the system comprises an index file establishing module, a word segmentation module and a retrieval sequencing module. The retrieval module adopts a Lucene full-text information retrieval engine, and Lucene is a full-text retrieval engine toolkit of an open source code issued by the Apache software foundation. As shown in fig. 11, the specific implementation steps of each module are as follows:
1. establishing an index file: the index file establishing module splits original data in the material library, stores the split data into a vocabulary table, obtains an index value of the data in the splitting process, and establishes an index file. During searching, firstly, searching from the vocabulary, and entering the original data for searching after the searching is successful;
2. word segmentation: the word segmentation module is used for segmenting words of the user problem, and comprises operations of removing stop words, converting English letters into lower case words and the like, and specifically an IKAnalyzer word segmentation device is used, so that words can be segmented according to Chinese words, and keywords in the problem can be obtained after the words are segmented;
3. and (3) retrieval and sorting: the retrieval sorting module can set the relevancy scores in the Lucene to sort different results, and the invention is set to sort according to a single field. And searching keywords obtained by the word segmentation result in the index file, and returning the top 5 results with the highest scores as answers.
The end-to-end module comprises: the device comprises a model building module, a word frequency dictionary generating module, a serialization operation module and an deserialization operation module.
The model building module is used for building a deep learning model by adopting a word embedding layer and a bidirectional gating cyclic unit network. The deep learning model comprises an encoder and a decoder; both the encoder and the decoder are constructed based on a word embedding layer and a bidirectional gating cyclic unit network; attention mechanisms have also been introduced in the encoders.
And the word frequency dictionary generating module is used for taking the corpus as a training data set in the training stage of the deep learning model, and performing word segmentation and word frequency statistics on the training data set in sequence to generate a word frequency dictionary.
The serialization operation module is used for generating a serialization dictionary and an anti-serialization dictionary according to the word frequency dictionary, carrying out serialization operation on the training data by using the serialization dictionary, converting the text of the training data into a digital sequence, and inputting the digital sequence of the training data into the deep learning model; or in the testing stage of the deep learning model, the problems input by the user are serialized, the texts of the problems are converted into digital sequences, and the digital sequences of the problems are input into the deep learning model for recognition.
And the deserialization operation module is used for deserializing the question answer sequence identified and output by the deep learning model, converting the digital sequence into a text sequence and obtaining the question answer in a text form.
Before using an end-to-end module, the building of a deep learning model and the training of the model need to be completed, and the training and testing process of the deep learning model is as shown in fig. 12, which specifically includes the following steps:
1. building a deep learning model: an encoder and a decoder of a deep learning model (hereinafter referred to as a model) are both constructed by using a word embedding layer and a bidirectional gating cyclic unit, an attention mechanism is introduced into the decoder, specifically, a Luong attention mechanism is used, an attention weight calculation method is used, an Adam optimization algorithm is used as an optimization algorithm of the model, and a loss function uses negative log likelihood loss.
2. In the training phase of the deep learning model, a corpus in a knowledge base module is used as a training data set. The method comprises the steps of preprocessing data in a data set, filtering noise, removing blank lines and special symbols in the data, segmenting the data, designing a jieba module for loading a stop word dictionary to perform Chinese segmentation, and segmenting sentences in a corpus into a form of words for subsequent processing.
3. And after word segmentation, carrying out word frequency statistics on the word segmentation result to generate a word frequency dictionary, and generating a serialization dictionary and an anti-serialization dictionary according to the word frequency dictionary. Specifically, the word frequency dictionary takes words appearing in the corpus as keys and the number of times of appearance of the corresponding words as a value, the serialization dictionary selects words with the number of words in the word frequency dictionary larger than a certain threshold value, the words are used as the keys of the dictionary, the numerical sequence number is used as the value of the dictionary, and the text can be converted into the numerical sequence through the serialization dictionary. And reversing the key values of the serialization dictionary to obtain an anti-serialization dictionary, wherein the digital sequence can be converted into a text by using the anti-serialization dictionary.
4. And performing serialization operation on the processed training data by using a serialization dictionary, converting the text into a digital sequence, and then inputting the digital sequence into a model for training. The input of the model is a serialized question sequence, and the output is a generated answer sequence. And calculating a loss value between the generated answer and the real answer through a loss function, and adjusting parameters of the model by using an optimization algorithm. And introducing a Teacher Forcing mechanism in the training process to accelerate the convergence of the model, and finishing the training of the model when the loss value of the model is smaller than a set threshold value. The correlation calculation formula is as follows:
attention weight calculation formula:
in the formula
Which represents the hidden state of the decoder,
representing the output of the encoder.
Adam optimization algorithm:
w represents the parameters of the model, α represents the learning rate, vwRepresenting the momentum after the calculated decay, swRepresents the average of the historical squared gradient of the decay, δ being a smoothing term used to prevent the divisor from being 0.
Negative log-likelihood loss:
loss(p，x)＝-∑x*log(p) (3)
x represents the true value of the corresponding sample and p represents the probability of classification of the corresponding sample.
After the model training is finished, the end-to-end module can be used. When a question of a user is input, preprocessing and word segmentation are carried out on the question of the user, the processed data are serialized, a model is input, the answer sequence is output by the model, and the answer to the question in a text form is obtained by deserializing the output sequence.
According to the invention, the deep learning model is used for identifying the user intention, so that the accuracy of judging the user intention is improved. The graph matching method used by the invention can fully utilize semantic information in the user questions and improve the quality of answers; the invention combines a plurality of question answering methods, which complement each other, and improves the capability and range of the system for answering questions.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. The device disclosed by the embodiment corresponds to the method disclosed by the embodiment, so that the description is simple, and the relevant points can be referred to the method part for description.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims (10)
1. An intelligent question-answering system for the automobile field is characterized by comprising: the system comprises a knowledge base module, a visual interaction module, an intention identification module, a graph matching module, a template matching module, a retrieval module and an end-to-end module;
the knowledge base module is used for storing a knowledge map and a corpus of the automobile field;
the visual interaction module is used for a user to input questions and feed back answers to the questions;
the intention identification module is used for judging whether the type of the user input question is an automobile field question or a chatting question, inputting the automobile field question to the graph matching module and inputting the chatting question to the end-to-end module;
the map matching module is used for matching the automobile field questions with the knowledge map in the knowledge base module by using a map matching method, and if the matching is successful, the answers to the questions are fed back to the visual interaction module; if the matching fails, inputting the automobile field problem into the template matching module;
the template matching module is used for matching automobile field problems with problem templates predefined in the knowledge base module by using a template matching method, and if the matching is successful, the problem answers are fed back to the visual interaction module; if the matching fails, inputting the automobile field problem into the retrieval module;
the retrieval module is used for retrieving answers of the automobile field questions in the corpus, and if the retrieval is successful, the answers of the questions are fed back to the visual interaction module; if the retrieval fails, inputting the automobile field problem into the end-to-end module;
the end-to-end module is used for identifying automobile field questions or chatting questions according to a pre-established deep learning model, generating question answers and feeding the question answers back to the visual interaction module.
2. The intelligent question-answering system for the automobile field according to claim 1, wherein the knowledge graph comprises: triple knowledge generated after processing of structured data related to the automobile field; the corpus at least comprises chatting corpus, data sets and unstructured text data related to the automobile field.
3. The intelligent question-answering system for the automobile field according to claim 1, wherein the visual interaction module is a web interaction interface based on a Django framework or a WeChat interaction interface based on itChat.
4. The system of claim 1, wherein the intention recognition module is configured to perform tagging and word segmentation operations on a current input question of the user by using a pre-trained FastText text classification model, and determine that the type of the input question is an automobile field question or a chat question.
5. The intelligent question-answering system for the automobile field according to claim 1, wherein the graph matching module comprises:
the dictionary module is used for extracting and constructing dictionaries of different types from the knowledge graph in advance; the dictionary includes: the system comprises an entity dictionary, a relation dictionary, an attribute value dictionary, a type dictionary generated manually in advance, a query word dictionary and a stop word dictionary;
the dictionary tree module is used for correspondingly generating a dictionary tree according to the content of each dictionary;
the preprocessing module is used for removing punctuation marks and carrying out capital and lower English case conversion processing on the problems input by the user;
the matching module is used for matching each node of the dictionary tree downwards from a root node according to the character sequence until the preprocessed user input problem cannot be matched;
the word segmentation and part-of-speech tagging module is used for segmenting words of the matching result output by the matching module and tagging the word segmentation result in terms of speech;
the dependency tree construction module is used for analyzing semantic association among all parts in the word segmentation result and establishing a dependency tree according to the semantic association;
the node type judging module is used for traversing each node of the dependency tree in the dictionary of each type and determining the type of each node; if the word corresponding to the node appears in the stop word dictionary, setting the node as a stop word node; if the editing distance similarity between the words corresponding to the nodes and the words in the entity dictionary or the relation dictionary is larger than a set threshold value, setting the nodes as entity query nodes or relation query nodes;
the pronoun resolution module is used for finding out a node of which the node type is a pronoun in the dependency tree according to the part of speech tagging result, finding out an entity node closest to the node, and taking a word corresponding to the entity node as a specific object referred by the pronoun;
the intention identification submodule is used for finding out the query word nodes in the dependency tree and finding out the attribute nodes or the relationship nodes which are closest to the nodes as the intention;
the query graph building module is used for initializing a query graph set, then traversing all entity query nodes found from the dependency tree, finding out other entity query nodes closest to the entity query node for each entity query node, obtaining the shortest paths of the other entity query nodes in the dependency tree, and finally adding head and tail nodes and paths between the two nodes into the query graph set to obtain a query graph;
the graph matching sub-module is used for traversing each node in the query graph in a breadth-first mode, converting an edge starting from each node in the query graph into a graph database query statement, and if a result can be queried in a graph database, indicating that the edge is successfully matched; if the edge matching is successful, adding the edge into the query subgraph; judging whether the query subgraph is the generation subgraph of the query graph; if the query subgraph is generated, the query subgraph is successfully matched with the knowledge graph, edges corresponding to the relation nodes identified by the intention identification submodule are converted into query sentences, and query results are used as matching answers of the query subgraph and stored in a result set; generating subgraphs, wherein the generated subgraphs comprise all vertex subgraphs in the query graph;
and the answer ranking module is used for ranking the matched answers obtained by the graph matching sub-module, adding the number of the query graph edges and the sum of the similarity of all the nodes of the query graph to calculate a score, and taking the first N answers as a final result.
6. The intelligent question-answering system for the automobile field according to claim 1, wherein the template matching module comprises:
the template predefining module is used for predefining templates with some problems, setting a trigger word for each template, setting a template slot position to be filled, and if the trigger word is identified in the problems input by the user, considering that the corresponding template is matched;
the trigger word and key information identification module is used for carrying out word segmentation and semantic analysis operations on the problems input by the user, identifying the trigger words and key information in the problems and matching the trigger words and key information with corresponding templates;
the template slot filling module is used for filling the trigger words and the key information identified by the key information identification module into preset template slots according to the templates matched with the trigger words to generate complete template query sentences;
and the query execution module is used for generating a graph database query statement by using the complete template query statement, and matching the knowledge graph by using the graph database query statement to obtain a question answer.
7. The intelligent question-answering system for the automobile field according to claim 1, wherein the retrieval module comprises:
the index file establishing module is used for splitting original data in the corpus, storing the split data into a vocabulary table, obtaining an index value of the data in the splitting process, and establishing an index file according to the index value;
the word segmentation module is used for segmenting words input by a user according to Chinese words by adopting an IKAnalyzer word segmentation device to obtain keywords in the problems;
the retrieval ordering module is used for searching keywords obtained by the word segmentation result in the index file by adopting a Lucene full-text information retrieval engine, and retrieving the keywords from the vocabulary firstly during the searching and then entering the original data for retrieving after the retrieval is successful; and sorting the different search results according to the relevancy scores, taking the top 5 results with the highest scores as answers to the questions, and returning the answers to the visual interaction module.
8. The intelligent question-answering system for the automobile field according to claim 1, wherein the end-to-end module comprises:
the model building module is used for building a deep learning model by adopting a word embedding layer and a bidirectional gating cyclic unit network;
the word frequency dictionary generating module is used for taking the corpus as a training data set in the training stage of the deep learning model, and performing word segmentation and word frequency statistics on the training data set in sequence to generate a word frequency dictionary;
the serialization operation module is used for generating a serialization dictionary and an anti-serialization dictionary according to the word frequency dictionary, carrying out serialization operation on the training data by using the serialization dictionary, converting the text of the training data into a digital sequence, and inputting the digital sequence of the training data into the deep learning model; or in the testing stage of the deep learning model, the problems input by the user are serialized, the texts of the problems are converted into digital sequences, and the digital sequences of the problems are input into the deep learning model for recognition;
and the deserialization operation module is used for deserializing the question answer sequence identified and output by the deep learning model, converting the digital sequence into a text sequence and obtaining the question answer in a text form.
9. The intelligent question-answering system for the automobile field according to claim 8, wherein the deep learning model comprises an encoder and a decoder; both the encoder and the decoder are constructed based on a word embedding layer and a bidirectional gating cyclic unit network; attention mechanisms have also been introduced into the encoder.
10. The intelligent question-answering system for the automobile field according to claim 8, wherein in a training stage, the deep learning model generates a loss value between a currently output answer and a real answer through loss function calculation, an optimization algorithm is adopted to adjust parameters of the deep learning model, and a Teacher Forcing mechanism is combined to accelerate the convergence speed of the deep learning model until the loss value is smaller than a set threshold value, so that the training of the deep learning model is finished.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110778221.1A CN113505209A (en) | 2021-07-09 | 2021-07-09 | Intelligent question-answering system for automobile field |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110778221.1A CN113505209A (en) | 2021-07-09 | 2021-07-09 | Intelligent question-answering system for automobile field |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113505209A true CN113505209A (en) | 2021-10-15 |
Family
ID=78012493
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110778221.1A Pending CN113505209A (en) | 2021-07-09 | 2021-07-09 | Intelligent question-answering system for automobile field |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113505209A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210237757A1 (en) * | 2020-01-31 | 2021-08-05 | Toyota Jidosha Kabushiki Kaisha | Information processing device, information processing method, and storage medium storing information processing program |
| CN114189740A (en) * | 2021-10-27 | 2022-03-15 | 杭州摸象大数据科技有限公司 | Video synthesis dialogue construction method and device, computer equipment and storage medium |
| CN114328880A (en) * | 2022-01-19 | 2022-04-12 | 重庆长安汽车股份有限公司 | Intelligent question and answer method and system for automobile field |
| CN114357139A (en) * | 2022-01-07 | 2022-04-15 | 深圳市易流科技股份有限公司 | Method, device, storage medium and equipment for responding to guard checking |
| CN114691831A (en) * | 2022-03-31 | 2022-07-01 | 彩虹无线（北京）新技术有限公司 | Task-type intelligent automobile fault question-answering system based on knowledge graph |
| CN114756663A (en) * | 2022-03-29 | 2022-07-15 | 税友信息技术有限公司 | Intelligent question answering method, system, equipment and computer readable storage medium |
| CN115238101A (en) * | 2022-09-23 | 2022-10-25 | 中国电子科技集团公司第十研究所 | Multi-engine intelligent question-answering system oriented to multi-type knowledge base |
| CN115658852A (en) * | 2022-12-28 | 2023-01-31 | 北京百车宝科技有限公司 | Automobile intelligent question-answering system based on knowledge base |
| CN116450858A (en) * | 2023-06-14 | 2023-07-18 | 辰风策划(深圳)有限公司 | Sales system for electronic product |
| CN117573847A (en) * | 2024-01-16 | 2024-02-20 | 浙江同花顺智能科技有限公司 | Visualized answer generation method, device, equipment and storage medium |
| CN118069816A (en) * | 2024-04-17 | 2024-05-24 | 深圳市前海数据服务有限公司 | Intelligent data service robot system based on NLP technology |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110110062A (en) * | 2019-04-30 | 2019-08-09 | 贝壳技术有限公司 | Machine intelligence answering method, device and electronic equipment |
| CN110399457A (en) * | 2019-07-01 | 2019-11-01 | 吉林大学 | A kind of intelligent answer method and system |
| CN111125316A (en) * | 2019-12-26 | 2020-05-08 | 北京工业大学 | Knowledge base question-answering method integrating multiple loss functions and attention mechanism |
| CN111708869A (en) * | 2020-05-12 | 2020-09-25 | 北京明略软件系统有限公司 | Man-machine conversation processing method and device |
| CN112364147A (en) * | 2020-12-01 | 2021-02-12 | 四川长虹电器股份有限公司 | Cross-domain multi-turn dialogue method based on knowledge graph and implementation system |
| CN112597287A (en) * | 2020-12-15 | 2021-04-02 | 深圳市优必选科技股份有限公司 | Statement processing method, statement processing device and intelligent equipment |
-
2021
- 2021-07-09 CN CN202110778221.1A patent/CN113505209A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110110062A (en) * | 2019-04-30 | 2019-08-09 | 贝壳技术有限公司 | Machine intelligence answering method, device and electronic equipment |
| CN110399457A (en) * | 2019-07-01 | 2019-11-01 | 吉林大学 | A kind of intelligent answer method and system |
| CN111125316A (en) * | 2019-12-26 | 2020-05-08 | 北京工业大学 | Knowledge base question-answering method integrating multiple loss functions and attention mechanism |
| CN111708869A (en) * | 2020-05-12 | 2020-09-25 | 北京明略软件系统有限公司 | Man-machine conversation processing method and device |
| CN112364147A (en) * | 2020-12-01 | 2021-02-12 | 四川长虹电器股份有限公司 | Cross-domain multi-turn dialogue method based on knowledge graph and implementation system |
| CN112597287A (en) * | 2020-12-15 | 2021-04-02 | 深圳市优必选科技股份有限公司 | Statement processing method, statement processing device and intelligent equipment |
Non-Patent Citations (3)
| Title |
|---|
| MINH-THANG LUONG 等: "Effective Approaches to Attention-based Neural Machine Translation", 《ARXIV:1508.04025V5》 * |
| 刘泽华: "基于知识图谱的事实型智能问答方法研究", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
| 刘连喜: "基于深度学习的高校招生问答服务系统的研究及实现", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑》 * |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210237757A1 (en) * | 2020-01-31 | 2021-08-05 | Toyota Jidosha Kabushiki Kaisha | Information processing device, information processing method, and storage medium storing information processing program |
| US11577745B2 (en) * | 2020-01-31 | 2023-02-14 | Toyota Jidosha Kabushiki Kaisha | Information processing device, information processing method, and storage medium storing information processing program |
| CN114189740A (en) * | 2021-10-27 | 2022-03-15 | 杭州摸象大数据科技有限公司 | Video synthesis dialogue construction method and device, computer equipment and storage medium |
| CN114189740B (en) * | 2021-10-27 | 2022-11-11 | 杭州摸象大数据科技有限公司 | Video synthesis dialogue construction method and device, computer equipment and storage medium |
| CN114357139A (en) * | 2022-01-07 | 2022-04-15 | 深圳市易流科技股份有限公司 | Method, device, storage medium and equipment for responding to guard checking |
| CN114328880A (en) * | 2022-01-19 | 2022-04-12 | 重庆长安汽车股份有限公司 | Intelligent question and answer method and system for automobile field |
| CN114756663A (en) * | 2022-03-29 | 2022-07-15 | 税友信息技术有限公司 | Intelligent question answering method, system, equipment and computer readable storage medium |
| CN114691831B (en) * | 2022-03-31 | 2024-07-16 | 彩虹无线（北京）新技术有限公司 | Task type automobile fault intelligent question-answering system based on knowledge graph |
| CN114691831A (en) * | 2022-03-31 | 2022-07-01 | 彩虹无线（北京）新技术有限公司 | Task-type intelligent automobile fault question-answering system based on knowledge graph |
| CN115238101A (en) * | 2022-09-23 | 2022-10-25 | 中国电子科技集团公司第十研究所 | Multi-engine intelligent question-answering system oriented to multi-type knowledge base |
| CN115658852A (en) * | 2022-12-28 | 2023-01-31 | 北京百车宝科技有限公司 | Automobile intelligent question-answering system based on knowledge base |
| CN115658852B (en) * | 2022-12-28 | 2023-10-24 | 北京百车宝科技有限公司 | Automobile intelligent question-answering system based on knowledge base |
| CN116450858B (en) * | 2023-06-14 | 2023-09-05 | 辰风策划(深圳)有限公司 | Sales system for electronic product |
| CN116450858A (en) * | 2023-06-14 | 2023-07-18 | 辰风策划(深圳)有限公司 | Sales system for electronic product |
| CN117573847A (en) * | 2024-01-16 | 2024-02-20 | 浙江同花顺智能科技有限公司 | Visualized answer generation method, device, equipment and storage medium |
| CN117573847B (en) * | 2024-01-16 | 2024-05-07 | 浙江同花顺智能科技有限公司 | Visualized answer generation method, device, equipment and storage medium |
| CN118069816A (en) * | 2024-04-17 | 2024-05-24 | 深圳市前海数据服务有限公司 | Intelligent data service robot system based on NLP technology |
| CN118069816B (en) * | 2024-04-17 | 2024-07-19 | 深圳市前海数据服务有限公司 | Intelligent data service robot system based on NLP technology |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113505209A (en) | Intelligent question-answering system for automobile field | |
| CN108287822B (en) | Chinese similarity problem generation system and method | |
| CN112989005B (en) | Knowledge graph common sense question-answering method and system based on staged query | |
| CN104050256B (en) | Initiative study-based questioning and answering method and questioning and answering system adopting initiative study-based questioning and answering method | |
| CN107798140B (en) | Dialog system construction method, semantic controlled response method and device | |
| CN110377715A (en) | Reasoning type accurate intelligent answering method based on legal knowledge map | |
| CN108681574B (en) | Text abstract-based non-fact question-answer selection method and system | |
| CN112800170A (en) | Question matching method and device and question reply method and device | |
| CN109271506A (en) | A kind of construction method of the field of power communication knowledge mapping question answering system based on deep learning | |
| CN109460459B (en) | Log learning-based dialogue system automatic optimization method | |
| CN111241294A (en) | Graph convolution network relation extraction method based on dependency analysis and key words | |
| CN113505200B (en) | Sentence-level Chinese event detection method combined with document key information | |
| CN108538294B (en) | Voice interaction method and device | |
| CN118170894B (en) | Knowledge graph question-answering method, knowledge graph question-answering device and storage medium | |
| CN109614620B (en) | HowNet-based graph model word sense disambiguation method and system | |
| CN108509409A (en) | A method of automatically generating semantic similarity sentence sample | |
| CN110348024A (en) | Intelligent identifying system based on legal knowledge map | |
| CN112328800A (en) | System and method for automatically generating programming specification question answers | |
| CN112364132A (en) | Similarity calculation model and system based on dependency syntax and method for building system | |
| CN111460114A (en) | Retrieval method, device, equipment and computer readable storage medium | |
| CN108763355B (en) | User-based intelligent robot interactive data processing system and method | |
| CN113971394A (en) | Text repeat rewriting system | |
| CN112800184A (en) | Short text comment emotion analysis method based on Target-Aspect-Opinion joint extraction | |
| CN115017268A (en) | Heuristic log extraction method and system based on tree structure | |
| CN107562907B (en) | Intelligent lawyer expert case response device |
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 |