A method and system for generating thematic pulses

Abstract

This invention discloses a method and system for generating thematic contexts, including acquiring a set of topics, constructing the information structure of the topics, extracting basic information from the topic texts, generating multi-angle descriptive information of the topic texts, constructing a graph network based on the basic information and multi-angle descriptive information, sorting the graph network by topics, obtaining an ordered set of valid topics, and generating thematic contexts based on the ordered set of valid topics. According to the above technical solution, text content under the same topic can be accurately categorized according to the same events, and the chronological order of the events and topics can be analyzed to form thematic contexts. The process is clear, flexible, efficient, and scalable.

Description

Technical Field

[0001] This invention relates to the field of computer applications, and more specifically, to a method and system for generating subject matter plots. Background Technology

[0002] The explosive growth of internet information, with its numerous sources, presents problems of disorganization, redundancy, and excessive data. Therefore, a technological method is needed to automatically process this information using computers, grouping information on similar topics together. This would effectively assist information workers in the use, production, and post-processing of information. For example, the generation of event timelines for news information primarily aims to identify which news articles in a news and public opinion system point to the same event, their chronological order, and thus understand how the event unfolded.

[0003] Currently, common methods for topic summarization include neural network text classification and text vector analysis. However, neural network text classification, considering the entire text, introduces significant interference from interjections, modal particles, and adjectives, leading to unclear classification boundaries. Furthermore, its reliance on neural networks results in poor fine-tuning. Text vector analysis, on the other hand, is prone to misassociations; for example, terms like "brother-in-law" and "sister's husband" might be considered similar, resulting in matching more irrelevant topics.

[0004] Therefore, a method for generating event context is needed that can clearly, accurately, and flexibly summarize topic information and generate thematic context. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides a method for generating a topic cluster, comprising the following steps:

[0006] Get a collection of topics;

[0007] Constructing the information structure of a topic: The information structure of a topic includes multiple topics, and each topic includes multiple topic texts;

[0008] Extract basic information from the topic text and generate multi-angle descriptive information of the topic text;

[0009] A graph network is constructed based on basic information and multi-angle descriptive information; the graph network is used to sort out themes and obtain an ordered set of effective topics;

[0010] Thematic threads are generated from an ordered set of valid topics.

[0011] The multi-dimensional descriptive information generated for the topic text includes: extracting triples from the topic text and selecting valid triples; the elements of a triple include a subject, a predicate, and an object.

[0012] Furthermore, the graph network structure includes: title nodes, element nodes, and node paths, where the node path represents the relationship between the title and element nodes;

[0013] Furthermore, a graph network is constructed based on basic information and multi-angle descriptive information, that is, elements of basic information and multi-angle descriptive information are filled into the structure of the graph network.

[0014] Further, the thematic analysis includes:

[0015] Retrieve related topic sets through graph path search;

[0016] Calculate the relevance of related topics to obtain a valid set of topics;

[0017] Filter the valid topic set by time;

[0018] The valid topics are sorted according to time to generate an ordered set of valid topics;

[0019] The effective topics are systematically categorized into themes to generate thematic threads.

[0020] Furthermore, the basic information in the topic text includes, but is not limited to: location and object name, where the object name includes person name and organization name.

[0021] Furthermore, the steps for extracting triples include:

[0022] The topic text is segmented to obtain text keywords, which include verbs and nouns;

[0023] Extract verbs from keywords to serve as the predicates of triples;

[0024] Using the predicate as the standard, the subject is determined by searching for nouns in the context.

[0025] Use the predicate as the standard and look up nouns in the context to determine the object.

[0026] Furthermore, determining the subject by searching for nouns in context also includes performing complete semantic processing on the subject; determining the object by searching for nouns in context also includes performing complete semantic processing on the object.

[0027] The process of selecting valid triplets includes:

[0028] Quantize triplet pairs to obtain triplet component values;

[0029] Sort the triples by their integrals and filter out the triples.

[0030] On the other hand, the present invention provides a system for generating topic clusters, characterized in that it includes:

[0031] Topic Acquisition Module: Used to acquire a set of topics and output the content of the elements in the topic set;

[0032] Information Structure Processing Module: This module processes the text content of the topic collection elements, performs information structuring on the text content, parses the text content into topic text, and outputs basic information and multi-angle descriptive information based on the topic text.

[0033] Graph Network Application Module: Based on the output of the Information Structure Processing Module, a graph network is constructed, and an ordered set of valid topics is output. The graph network structure includes: title nodes, element nodes, and node paths; where the node path represents the relationship between the title and element nodes.

[0034] Topic Pulse Output Module: Generates topic pulsations based on an ordered set of valid topics.

[0035] Furthermore, the information processing module includes:

[0036] Information extraction submodule: Used to extract basic information from topic text, including location and object name;

[0037] The triple processing submodule is used to extract and improve triples from topic text. The elements of a triple include subject, predicate, and object. Improving triples includes completing the content of the triple and filtering out invalid triples.

[0038] The graph network application module includes:

[0039] Graph network generation module: Generates a graph network based on the basic information and multi-angle description information output by the information processing module and the structure of the graph network;

[0040] Graph network computation module: Calculates the relevance of a specified topic and outputs a set of valid topics;

[0041] Set sorting module: Used to perform time filtering and sorting on the valid topic set output by the graph network calculation module, and output an ordered set of valid topics.

[0042] According to the present invention, text content under the same topic is accurately classified according to the same events, and the chronological order of the events is analyzed to form a thematic context, achieving a clear, flexible, efficient, and scalable process. Attached Figure Description

[0043] Figure 1 This is an original information image of a Weibo topic related to the same event provided in an embodiment of the present invention;

[0044] Figure 2 This is a flowchart of a method for generating a topic context according to an embodiment of the present invention;

[0045] Figure 3 This is a topic information structure diagram provided according to an embodiment of the present invention;

[0046] Figure 4 This is a graph network structure diagram provided according to an embodiment of the present invention;

[0047] Figure 5 This is a microblogging topic mapping diagram provided according to an embodiment of the present invention;

[0048] Figure 6 This is a structural diagram of a topic pulse generation system provided according to an embodiment of the present invention. Detailed Implementation

[0049] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings.

[0050] The problem this invention aims to solve, such as generating event timelines for news information, is primarily to identify which news articles in a news and public opinion system refer to the same event, and what the chronological order of these articles is, thereby understanding how the event unfolded. In other words, it uses technical means to categorize text content under the same topic according to similar events and analyze the chronological order of these events.

[0051] This article uses a case study from Weibo to illustrate this point: Weibo content has a topic-based nature, and information creators embed the topic title with # symbols. Therefore, each topic naturally aggregates the information published under that topic. In this example, the content is as follows... Figure 1 As shown. Although the topics are about the same event and constitute the whole event in a certain chronological order, there is no connection between them on Weibo.

[0052] The steps to connect the topics of an event are to first look at the structure of the topic, such as... Figure 3As shown: A topic includes multiple subtopics, each containing a title. Each subtopic contains content posted by creators (i.e., topic text) and the posting time. Clearly, the event described by each subtopic exists within the topic text. Therefore, analyzing the topic text reveals the connections between them. On the other hand, it's necessary to identify the commonalities in the topic texts of subtopics related to the same event. When describing an event, we typically mention time (when), place (where), people (who), and descriptions of the event from various angles (what). These descriptions can be further divided into subject-verb-object triplets. For example, phrases like "CCTV reporter visits a village in Yunnan where wild elephants stayed last night" and "wild elephants in Yunnan ate corn in a village" can both be used to extract triplets. By extracting these time, place, people, and descriptions of the event from various angles, and using specific techniques to calculate the scores between subtopics, we can determine whether two subtopics belong to the same event.

[0053] The topic context generation method provided by this invention summarizes the basic elements of topic text, extracts the description triplets of events from various angles, and judges the relationship between multiple topics. It is efficient and flexible.

[0054] Figure 2 This is a flowchart of a method for generating a topic cluster according to an embodiment of the present invention, as shown in the figure, including the following steps:

[0055] S100: Get the topic collection;

[0056] First, obtain the set of topics that need to be analyzed and define the symbols for the elements in the set. A single topic text under a topic is represented by x, a topic element is represented by S, n represents the number of topic texts under that topic, and the set of all topics is represented by E.

[0057] S110: Constructing the information structure of the topic: its structure is as follows Figure 3 As shown, the information structure of a topic includes multiple topics (i.e., topics in the topic set E), one topic includes multiple topic texts, and one topic text contains multiple times. In this example, time is an important element in constituting the topic context.

[0058] S120: Extract the basic information of the topic text; the basic information of the topic text includes time, location, and object name, where the object name includes person name and organization name. The specific implementation method is as follows:

[0059] 1. Time Processing: There are various ways to express time in topic texts, such as directly stating the time, estimating the time from the publication time when the content does not contain a time, and using multiple times. Time processing standardizes the format of time data.

[0060] The topic posting times generate a set Pt. S The set of times when the topic was generated, Pt. S Perform calculations to obtain the minimum value, min_Pt S The calculation rules for the time when a topic is generated are as follows:

[0061] Pt S =[pt x0 ,pt x1 ,...pt xn-1 ]

[0062] min_pt S =Min(Pt) S )

[0063] For events related to a topic, a set of event-related timestamps is obtained through time processing. x , and the time of publication (pt) x Combined together, they form the time set t of this content. x The time set of all content under a topic is combined into the topic's time set. S And the topic time set Time S By finding the maximum and minimum values, we can mark the topic's time boundaries:

[0064] min_time S =Min(Time) S )

[0065] max_time S =Max(Time) S )

[0066] 2. Location Processing: Processing location descriptions within topics and providing complete supplementary information, including country, province, city, and county (district) in the regional information, represented as region l. x The set of regions for all content under a topic is merged together to form the topic's region set (Loc). S , means as follows:

[0067] Loc S =[l x0 ,l x1 ,...l xn-1 ].

[0068] 3. Processing of Person and Organization Names: The names of persons and organizations in the topic text are generally text fragments such as "Beijing Hospital," "CCTV," "Ministry of Foreign Affairs," and "Henan Provincial People's Congress." This invention uses a Natural Language Processing (NLP) model to extract the names of persons and organizations. Furthermore, the sets of person and organization names from all topic texts under the topic set are merged, resulting in the topic person and organization name set Ner. S , represents the following:

[0069] Ner S =[p x0 ,p x1 ,...p xn-1 ].

[0070] S125: Generate multi-faceted descriptive information for topic text;

[0071] Multi-faceted descriptive information refers to a complete and structured explanation of various aspects of an event. A text contains a subject, predicate, object, attributive, adverbial, and other noun modifiers and interjections. In this step, words not closely related to the core of the text are removed, retaining only the core words while maintaining the contextual meaning. In this invention, subject-verb-object triples are extracted through syntactic analysis and part-of-speech tagging, and an algorithm is used to select valid triples. Subject-verb-object triples are one form of representing information from various perspectives. The execution method in this example is as follows:

[0072] ① Extracting subject-verb-object triplets

[0073]

[0074] A triple consists of three basic elements: subject, predicate, and object. According to Chinese syntax, sometimes the subject or object may be missing, but the predicate is usually not missing. Therefore, you can start with the predicate and find the nouns in the context to find the three elements.

[0075] Therefore, in this method, the topic text is segmented to obtain text keywords. Generally, keywords can be verbs or nouns.

[0076] After obtaining the keywords, the verbs are extracted as the predicates of the triples. Then, using the predicates as the standard, the nouns are searched in the context to determine the subject and object. In semantic environments where the subject or object is missing, after obtaining the predicate, this step is used to obtain the subject or object. The subject and object can also be further processed according to different situations, with complete semantic processing for each subject and object.

[0077] Complete word meaning processing adds a step to more accurately express the subject and object. It takes the noun as the center and finds the modifiers in the context to form a more accurate subject and object.

[0078] In practice, both the word segmentation (sentences) and the part-of-speech tag (POS) of the text content x can be obtained from the word segmentation tool. For example:

[0079] ["CCTV", "Reporter", "Visit", "Yunnan", "Wild Elephant Herd", "Last Night", "Stop", "Village"]

[0080] ["NN","NN","VV","NR","NN","NT","VV","NN"]

[0081] There are two verbs, 'visit' and 'stay'. When searching for contextual nouns centered on 'stay', the results are ('wild elephant herd', 'stay', 'village'). When searching for contextual nouns centered on 'visit', the results are ('reporter', 'visit', 'wild elephant herd'). Performing a complete semantic processing on the subject "reporter"—that is, finding modifiers through nouns centered on 'reporter'—results in "CCTV reporter," ultimately yielding ('CCTV reporter', 'visit', 'wild elephant herd').

[0082] This method extracts two triples, which are descriptions of the event from two different perspectives.

[0083] The pseudocode for syntactic analysis is as follows:

[0084]

[0085] Through the above processing, multiple sets of triples related to the events described by content x are obtained, forming a topic triple set `Spo`. S , `Spo S =[e x0 ,e x1 ,...e xn-1 ].

[0086] ② Triple sorting and filtering

[0087] The set of topic triples is too large, requiring sorting to select only the most important and representative triples. This invention employs a quantified triples approach to achieve this sorting, i.e., obtaining the score of each triple.

[0088] In this invention, the TF-IDF weight values ​​of all texts under a topic (which can be obtained from word segmentation tools) are used to quantify and score the triples, as follows:

[0089] First, input all the text S under the topic into the TF-IDF function of the word segmentation tool to obtain the weight values ​​of the word segmentation, in the following format: {'CCTV':0.62,'Reporter':0.76,'Visit':0.85,'Yunnan':0.79,'Wild Elephant Herd':0.93,'Last Night':0.53,'Stay':0.77,'Village':0.58};

[0090] Then iterate through the topic triple set `Spo` S If the triple contains tfidf S For each word in the triplet, its weight value is added to that triplet. Each triplet will then have a combined weight value. The set of combined weight values ​​for all triplets is represented by w. S This indicates that the set of triples `Spo` represents... S According to the score of the triplet, i.e., the weight value w S Sort the results and extract the top 10 largest triples; these represent the topic's description of the event from various angles. (Using Spo...) S express.

[0091] The pseudocode for sorting and filtering triples is as follows:

[0092]

[0093] In this invention, triples are filtered, retaining only the main triples. This further narrows the scope of associations for the next step of constructing the graph network, improving the effectiveness and correlation of each element.

[0094] S130: Construct a graph network based on basic information and multi-dimensional descriptive information;

[0095] Through steps S120 and S125, a large amount of data information conforming to the specifications was obtained. In this step, the attributes of the topic set are defined, and the information obtained in the above steps is assigned to the corresponding attributes, as shown in the table below:

[0096]

[0097] The structure of the graph network consists of: title nodes (represented by boxes), element nodes (represented by ellipses), and node paths, where the node paths represent the relationships between title and element nodes.

[0098] The graph network is constructed based on the attributes of each element in the topic set, which consists of basic information and multi-angle descriptive information. This involves filling the title node, element nodes, and node paths of the graph network with elements from the basic information and multi-angle descriptive information. The specific implementation method is as follows:

[0099] a. Treat the title of each topic S in E as a title node (represented by a box);

[0100] b. A set of topic regions, a set of topic people and organizations names, and a triplet describing the event from various angles, with each text element in the triplet as an element node (represented by an ellipse).

[0101] c. The relationship between topic S and text elements is represented as the path between nodes;

[0102] In this context, the path for countries within a region is "country", the path for provinces within a region is "province", and the path for cities within a region is "city".

[0103] For the names of individuals and organizations, the unified path is "person".

[0104] In this context, the subject and object of the triple describing the event from different perspectives are uniformly set to "subject and object". This is because the subject and object will be interchanged when describing different perspectives.

[0105] The predicate of a triple follows the path "predicate".

[0106] For example: There are three topics.

[0107] Topic 1: CCTV reporter visits village in Yunnan where wild elephants stayed last night

[0108]

[0109] Topic 2: Wild elephants in Yunnan devour corn in a village

[0110]

[0111] Topic 3: A post-90s girl recreates a Yunnan map using 16 dishes.

[0112]

[0113]

[0114] The generated graph network is constructed based on the above attributes and the graph structure is as follows: Figure 4 .

[0115] S140: Perform thematic sorting on the graph network to obtain an ordered set of valid topics;

[0116] First, extract any one of the title nodes (represented by root) from the graph network. Using the attributes of the title node as search criteria, find other title nodes (i.e. topics) that belong to the same event as the title node in the graph network of all topics, and sort these topics in chronological order.

[0117] 1) Obtain a set of related topics through graph path search:

[0118] 2) Calculate the relevance of the related topics to obtain a valid topic set E. score ;

[0119] 3) Effective topic set E score Time-based filtering will further refine the effective topic set. final ;

[0120] 4) Sort the set of valid topics to obtain an ordered set of valid topics.

[0121] S150: Generate thematic threads based on the ordered set of valid topics.

[0122] In this example, `E` final Topics are sorted by time within the set, forming a valid ordered set of topics.

[0123] `E final The topic set is based on the min_Pt of each topic. S Sort by time (sort), and label the result as E. final This topic set is the set of events E that belong to the same event as root and are arranged in chronological order. final =sort(`E final ,min_Pt S ).

[0124] After sorting the valid topics into an ordered set, they are displayed chronologically and categorized by theme, such as... Figure 5 As shown, the topic pulse is generated.

[0125] Figure 6 This is a structural diagram of the topic pulse generation system provided by the present invention. As shown in the figure, the generation system includes:

[0126] P610 Topic Acquisition Module: Used to acquire a collection of topics and output the content of the elements in the topic collection.

[0127] This module allows for the acquisition of topics by supporting the creation of new collections from multiple elements, as well as the addition of new elements to existing collections. For example, it can extract multiple Weibo posts at once and combine their text content into a new collection, or add new Weibo text content to an existing topic collection. The newly added content elements or newly created content collections are output to the P620 information structure processing module.

[0128] The P620 Information Structure Processing Module: This module processes the text content of the topic set elements, performs information structuring on the text content, parses the text content into topic text, and outputs basic information and multi-angle descriptive information based on the topic text. In summary, it acquires the content of each element in the topic set output by the P610 Topic Acquisition Module, extracting and splitting the content according to a standardized structure. In the sub-modules of the generation system, different processing environments are encapsulated based on the different forms of data existence in the process from text information to graph networks to pulse information, to ensure the independence of sub-modules and improve flexibility and maintainability. The specific execution structure includes the following parts:

[0129] The P621 information extraction submodule is used to extract basic information about the topic text from the topic collection elements, such as location, object name, and time, and to process information with different characteristics in different ways.

[0130] 1. Time Processing Function Block: Collects one or more times from the publication time and text content, and unifies the format of the time data. It supports processing multiple time formats and outputs time data under various conditions, such as specific time values, including the maximum and minimum values ​​within a time range, and time sets.

[0131] 2. Location Processing Function Block: Processes the descriptions of locations in the content under the topic and provides complete supplements to the descriptions.

[0132] 3. Name Processing Function Block: Extracts and merges the names of people and organizations from all topic texts under the topic set. The extraction process supports model training.

[0133] The P622 triplet processing submodule extracts and refines triples from topic text. Triples consist of a subject, verb, and object. It completes and refines the elements of triples, including adding subjects and objects when they are missing, and removing triples that are not closely related to the core of the text. The triplet processing submodule encapsulates functionality in its implementation, including:

[0134] 1. Triple Extraction Function Block: This function segments the input topic text to obtain keywords, extracts the predicate based on semantic analysis, and analyzes and refines the subject and object based on the predicate.

[0135] 2. Triple processing function block: Quantifies, filters, and sorts triples.

[0136] The P620 information structure processing module also provides an output interface to output basic information and triplet information according to the interface specification. This information is used by the next module in the process to generate a graph network. The data format supported by this interface is not limited to text and can be extended to images, audio and video, etc.

[0137] The P630 Graph Network Application Module: Based on the output of the P620 Information Structure Processing Module, this module constructs a graph network and outputs an ordered set of valid topics. The graph network structure in this module includes: title nodes and element nodes. The relationship between title nodes and element nodes is represented by node paths.

[0138] This module implements the generation and application of graph networks, including the process from generating graph networks to calculating related relationships based on graph elements. The module is encapsulated according to its functions and includes the following three sub-modules:

[0139] The P631 graph network generation submodule generates a graph network based on the basic information and multi-angle descriptive information output by the information processing module, combined with the graph network structure. The structure can be adjusted according to requirements, and the types of elements can be expanded. For example, in this invention, to protect the accuracy of associations, the graph network only considers text elements such as person / organization names, locations, times, and triples. As requirements change, other elements such as images can be added.

[0140] P632 Graph Network Computation Submodule: Calculates the relevance of a specified topic and outputs a set of valid topics.

[0141] In this invention, a title node of the graph network is designated as the title node of the root topic. The element nodes of this node are used as search criteria to find other topics belonging to the same event within the graph network corresponding to the set E of all topics. The computation submodule also supports the implementation of other algorithms, such as searching based on other elements corresponding to a certain element node, defining computation methods, and defining output methods.

[0142] The P633 set sorting submodule is used to perform time filtering and sorting on the valid topic set output by the graph network calculation module, and output an ordered set of valid topics.

[0143] The organization of effective topic sets is not limited to filtering and sorting by time; it can also be expanded according to needs.

[0144] P640: Topic Pulse Output Module: Generates topic pulsations based on an ordered set of valid topics.

[0145] This invention collects event information from various sources, categorizes content under the same type of event according to certain classification rules, and applies it to the event context generation scenario of news information. It can control the accuracy of classification, ensure the computational efficiency of the classification process, improve scalability, and clearly, accurately, and flexibly summarize topic information to generate thematic contexts.

[0146] The above-disclosed embodiments are merely a few specific examples of the present invention. However, the present invention is not limited thereto, and any variations that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A method for generating a thematic context, characterized in that, Includes the following steps: Retrieve a set of topics; the information structure of one topic includes multiple topics, and one topic includes multiple topic texts; Extract the basic information from the topic text; the basic information includes time, location, and object name, wherein the object name includes person name and organization name; Generating multi-dimensional descriptive information of the topic text includes: segmenting the topic text to obtain text keywords, the keywords including verbs and nouns; extracting verbs from the keywords as predicates of triples; using the predicates as criteria to find nouns in the context to determine the subject; using the predicates as criteria to find nouns in the context to determine the object; and performing complete semantic processing on the subject and / or object; quantifying the triples to obtain the triple component values; sorting the triples according to the triple component values, filtering out invalid triples, and selecting valid triples; The quantification of the triplet and the acquisition of the triplet group values ​​include: performing TF-IDF calculation on all text under the topic to obtain the weight value of each word segment; traversing the triplet and summing the weight values ​​of the words contained in the triplet as the comprehensive weight value of the triplet; the sorting according to the triplet group values ​​and filtering out the triplets that can be eliminated includes: sorting according to the comprehensive weight value from largest to smallest, and extracting the valid triplets according to the sorting result; The structure of a graph network is constructed, which includes: title nodes, element nodes, and node paths; wherein, the node path is the relationship between the title and element nodes. Based on the basic information and the selected valid triples as multi-angle description information, the elements of the basic information and multi-angle description information are filled into the structure of the graph network to construct the graph network; The graph network is thematically organized to obtain an ordered set of valid topics; this includes: obtaining a set of related topics through graph path search; calculating the relevance of related topics to obtain a set of valid topics; filtering the set of valid topics by time; and sorting the valid topics according to time to generate an ordered set of valid topics. Thematic threads are generated based on the ordered set of valid topics.

2. The method for generating a thematic cluster according to claim 1, characterized in that, The element nodes in the graph network structure include: region nodes based on the location, person / organization nodes based on the object name, subject / object nodes based on the subject and object of the triple, and predicate nodes based on the predicate of the triple; the node paths include: the path between the title node and the region node is labeled as a region hierarchy relationship; the path between the title node and the person / organization node is uniformly labeled as "person"; the path between the title node and the subject / object node is uniformly labeled as "subject / object"; the path between the title node and the predicate node is uniformly labeled as "predicate"; constructing the graph network based on the basic information and multi-angle description information refers to filling the elements of the basic information and multi-angle description information into the structure of the graph network.

3. The method for generating a thematic cluster according to claim 1, characterized in that, The ordered set of valid topics is categorized by topic to generate topic lines; The method of obtaining the set of related topics through graph path search includes: taking any title node in the graph network as the search starting point and taking the element nodes associated with that title node as the search conditions.

4. The method for generating a thematic cluster according to claim 1, characterized in that, The time includes the topic publication time and / or the event time extracted from the topic text content, and the time is formatted in a unified manner; the location includes regional information after the location description is fully supplemented, and the regional information includes country, province, city, county / district.

5. The method for generating a thematic cluster according to claim 1, characterized in that, The complete semantic processing of the subject and / or object includes: taking the noun of the subject or object as the center, finding the modifiers in the context, and combining the noun and the modifiers to form a complete subject or object.

6. A system for generating thematic threads, characterized in that, The method for performing claim 1 includes: Topic Acquisition Module: Used to acquire a set of topics and output the content of the elements in the topic set; Information Structure Processing Module: This module processes the text content of the topic set elements, performs information structuring processing on the text content, parses the text content into topic text, and outputs basic information and multi-angle description information based on the topic text. The basic information includes time, location, and object name, and the multi-angle description information includes effective triples extracted from the topic text and filtered by quantization. Graph Network Application Module: Based on the output of the information structure processing module, a graph network is constructed, and an ordered set of valid topics is output. The graph network structure includes: title nodes, element nodes, and node paths; wherein, the node path is the relationship between the title and element nodes. Topic Pulse Output Module: Generates topic pulsations based on the ordered set of valid topics.

Images