A method and system for information classification retrieval

Abstract

This application relates to the field of information retrieval technology, and in particular to an information classification and retrieval method and system. The method includes: acquiring a multi-source heterogeneous data stream of futures contracts containing structured data and unstructured text; parsing the unstructured text to extract information units to be classified; matching this data with a dynamically updated futures contract-entity mapping graph to obtain preliminary contract tags and associated entity chains, which are then stored in the corresponding contract's original information pool; acquiring a personalized profile containing a user's static watchlist and dynamic behavioral characteristics; recalling candidate information units based on this profile and reordering them by relevance to generate a preliminary retrieval sequence; and finally, generating the final retrieval result through timeliness decay calculation and redundancy elimination. This application helps overcome the shortcomings of existing futures information retrieval technologies, such as inaccurate classification, insufficient personalization, and poor timeliness and redundancy control.

Description

Technical Field

[0001] This application relates to the field of information retrieval technology, and in particular to an information classification and retrieval method and system. Background Technology

[0002] With the rapid development of the futures market, futures-related information has exploded, encompassing various types of content such as exchange announcements, macroeconomic data, industry research reports, news, and trading data, forming a multi-source, heterogeneous data stream. Structured data mainly includes directly quantifiable data such as futures contract prices, trading volume, and open interest, while unstructured text includes content that cannot be directly structured, such as news flashes, in-depth analysis articles, and policy interpretations.

[0003] Existing futures information retrieval technologies mostly rely on keyword matching, which has significant drawbacks: Firstly, it cannot effectively handle implicit connections in unstructured text, resulting in low information classification accuracy and difficulty in precisely matching futures products. Secondly, the search results lack personalized adaptation, failing to push accurate information based on user preferences and behavioral habits, and are prone to information redundancy and insufficient timeliness, thus failing to meet users' needs for efficient and accurate futures information retrieval. Therefore, there is an urgent need for an information classification and retrieval method and system that can accurately classify and personally match multi-source heterogeneous information while balancing timeliness and redundancy control. Summary of the Invention

[0004] To help overcome the shortcomings of existing technologies in futures information retrieval, such as inaccurate classification, lack of personalization, and poor timeliness and redundancy control, this application provides an information classification retrieval method and system.

[0005] Firstly, this application provides an information classification and retrieval method, which adopts the following technical solution: An information classification and retrieval method, comprising: Acquire multi-source heterogeneous data streams from the futures market, wherein the multi-source heterogeneous data streams include structured data and unstructured text; Natural language parsing is performed on the unstructured text to extract multiple information units to be classified; The information units to be classified are matched with the dynamically updated futures product-entity mapping map to obtain the preliminary product label and associated entity chain corresponding to each information unit to be classified. Based on the preliminary variety label, the information unit to be classified is stored in the corresponding variety original information pool; Obtain a personalized information profile of the target user, the personalized information profile including a static follow list and dynamic behavioral characteristics; Based on the static attention list, candidate information units are recalled from the original information pool of the varieties; The candidate information units are reordered based on their relevance using the dynamic behavioral features to generate a preliminary retrieval sequence. The candidate information units in the preliminary retrieval sequence are subjected to time-sensitivity decay calculation and redundancy elimination to generate the final retrieval results.

[0006] Optionally, matching the information unit to be classified with the dynamically updated futures product-entity mapping map includes: Identify the core entity words in the information unit to be classified, including product code, contract name, and exchange abbreviation; If the core entity word is identified, the core entity word is directly mapped to a preliminary variety label, and the corresponding entity's associated entity chain in the futures variety-entity mapping graph is extracted. If the core entity words are not identified, then the edge entity words in the information unit to be classified are extracted. The edge entity words include upstream and downstream product names, macroeconomic indicators, and policy issuing entities. Calculate the semantic similarity and co-occurrence frequency between the marginal entity words and each variety node in the futures variety-entity mapping graph; The variety nodes with the highest semantic similarity and co-occurrence frequency exceeding the low-frequency threshold are selected as the initial variety labels, and the edge entity words in the matching process are stored as supplementary entity chains in the associated entity chain.

[0007] Optionally, calculating the semantic similarity and co-occurrence frequency between the marginal entity words and each variety node in the futures variety-entity mapping graph includes: Determine whether the edge entity words belong to sudden new words or out-of-vocabulary words; If it is a newly emerging word, then obtain the context of the corresponding word in recent authoritative information; Based on the aforementioned context, the range of varieties associated with the corresponding words is inferred through a real-time semantic reasoning engine; If the inferred range of varieties is a single variety, then the single variety is used as the matching result, and the update mechanism of the futures variety-entity mapping map is triggered. If the inferred range of varieties includes multiple candidate varieties, the historical frequency distribution of the corresponding words in the relevant information of different candidate varieties is obtained, and the candidate variety with the highest frequency is taken as the matching result.

[0008] Optionally, the mechanism for triggering the update of the futures contract-entity mapping map includes: Obtain the actual co-occurrence data of the newly emerging words with each variety node within a subsequent preset time period; Determine whether the actual co-occurrence data supports the inference results of the real-time semantic reasoning engine; If the inference result is supported, the sudden new word and its associated variety node will be permanently written into the futures variety-entity mapping graph and its confidence level will be marked. If the inference result is not supported, the sudden new word will be stored in a temporary observation area, and the weight of similar inference paths in the real-time semantic reasoning engine will be reduced.

[0009] Optionally, reordering the candidate information units based on the dynamic behavioral features includes: Obtain the target user's click sequence, dwell time, sharing behavior, and collection records within a preset history window; Extract user click preferences for specific information types, specific data formats, and specific publishing sources from the click sequence; The user's level of attention to in-depth analysis information, news flash information, and data table information can be identified from the dwell time. The sharing behavior is mapped to the user's first recognition label of the social value of information; The collection records are mapped as a second recognition tag of the long-term value of information by the user; The click preference, the degree of attention, the first approval tag, and the second approval tag are quantitatively fused to generate a multi-dimensional behavior weight vector; Calculate the matching degree between the feature vector of each candidate information unit and the behavior weight vector; The candidate information units are sorted in descending order based on the matching degree to generate a preliminary retrieval sequence.

[0010] Optionally, after calculating the matching degree between the feature vector of each candidate information unit and the behavior weight vector, the method further includes: Obtain the special status identifier of the current market, including the approaching delivery month, the eve of the release of major data, and the period of abnormal market fluctuations; Determine whether the special state identifier triggers a preset behavior drift rule; If the behavior drift rule is triggered, the weight values ​​of some dimensions in the behavior weight vector are temporarily adjusted to generate a temporary weight vector. The matching degree of the candidate information unit is recalculated using the temporary weight vector to generate a temporary retrieval sequence; The temporary retrieval sequence and the preliminary retrieval sequence are weighted and fused to generate a fused retrieval sequence.

[0011] Optionally, determining whether the special state identifier triggers a preset behavior drift rule includes: If the special status is identified as the delivery month is approaching, then it is determined whether the static watchlist contains varieties that are approaching delivery. If it includes commodities nearing delivery, the weighting of information types related to delivery rules, warehouse receipt reports, and position limits will be increased; If the product is not close to delivery, it is determined whether the user has shown interest in the relevant product in the same period in history. If so, the number of information recalls for the relevant product is increased.

[0012] Optionally, performing timeliness attenuation calculation and redundancy elimination on candidate information units in the preliminary retrieval sequence includes: Obtain the publication time and latest citation time of each candidate information unit in the preliminary retrieval sequence; Based on the publication time and the latest citation time, calculate the timeliness value decay curve for each candidate information unit; If the publication time of a candidate information unit exceeds a preset hard timeliness threshold, the corresponding candidate information unit will be removed from the preliminary search sequence to obtain the first filtered sequence. Semantic fingerprinting is performed on the remaining candidate information units in the first filtered sequence to identify candidate information units whose content similarity exceeds the similarity threshold, and the corresponding candidate information units are clustered into redundant information groups. In each redundant information group, retain the candidate information unit with the highest source authority or the most citations, and mark the remaining candidate information units as redundant and remove them from the first filtered sequence to obtain the second filtered sequence; The candidate information units in the second filtered sequence are used as the final retrieval results.

[0013] Optionally, before retaining the candidate information unit with the highest source authority or the most citations in each redundant information group, the following may also be included: Determine whether multiple candidate information units in the redundant information group contain contradictory core viewpoints; If the core viewpoints are contradictory, then identify the source authority and data support strength of each candidate information unit corresponding to the contradictory core viewpoints. The first candidate information unit with the highest source authority and the strongest data support in the redundant information group is retained. From the candidate information units in the same redundant information group that have a viewpoint that is opposed to the first candidate information unit, select the second candidate information unit with the second highest source authority and the second strongest data support. At the same time, the first candidate information unit and the second candidate information unit are retained, and when the candidate information units in the second filtered sequence are displayed as the final search results, viewpoint comparison identifiers are added to the first candidate information unit and the second candidate information unit.

[0014] Secondly, this application also discloses an information classification and retrieval system, which adopts the following technical solution: An information classification and retrieval system, comprising: The first acquisition module is used to acquire multi-source heterogeneous data streams from the futures market, wherein the multi-source heterogeneous data streams include structured data and unstructured text; The unit extraction module is used to perform natural language parsing on the unstructured text and extract multiple information units to be classified. The second acquisition module is used to match the information unit to be classified with the dynamically updated futures product-entity mapping map to obtain the preliminary product label and associated entity chain corresponding to each information unit to be classified. The unit storage module is used to store the information unit to be classified into the corresponding original information pool of the variety based on the preliminary variety label. The third acquisition module is used to acquire the personalized information profile of the target user, which includes a static follow list and dynamic behavioral characteristics. The unit retrieval module is used to recall candidate information units from the original information pool of the variety based on the static attention list. The sequence generation module is used to reorder the candidate information units based on the dynamic behavioral features to generate a preliminary retrieval sequence. The result generation module is used to perform time-sensitivity decay calculation and redundancy elimination on the candidate information units in the preliminary retrieval sequence to generate the final retrieval results.

[0015] In summary, this application includes the following beneficial technical effects: By matching the unclassified information units extracted from unstructured text parsing with a dynamically updated futures product-entity mapping map, and leveraging mechanisms such as direct mapping of core entity words or semantic similarity, co-occurrence frequency matching of marginal entity words, and handling of sudden new words, preliminary product tags are accurately obtained. This effectively overcomes the problem of inaccurate classification caused by the inability of existing keyword matching technologies to handle implicit related information. By obtaining personalized information profiles of target users, including static watchlists and dynamic behavioral characteristics, candidate information units are recalled and reordered based on relevance, significantly improving the matching degree between search results and user needs, and solving the problem of lack of personalized adaptation in existing technologies. At the same time, by performing timeliness decay calculations and redundancy elimination on the search sequence, expired and invalid information is removed, and duplicate and redundant content is filtered out, further improving the timeliness and conciseness of information presentation. Ultimately, this achieves accurate classification, efficient retrieval, and personalized push of multi-source heterogeneous information in the futures market, greatly improving the accuracy, practicality, and user experience of futures information retrieval. Attached Figure Description

[0016] Figure 1 This is a main flowchart of an information classification and retrieval method according to an embodiment of this application; Figure 2 This is a flowchart of the steps for matching the information units to be classified with the dynamically updated futures product-entity mapping map; Figure 3 This is a flowchart of the steps to calculate the semantic similarity and co-occurrence frequency between the marginal entity words and the nodes of each commodity in the futures commodity-entity mapping graph; Figure 4 This is a flowchart of the steps to trigger the update mechanism of the futures product-entity mapping map; Figure 5 This is a module diagram of an information classification and retrieval system according to an embodiment of this application.

[0017] Explanation of reference numerals in the attached figures: 1. First acquisition module; 2. Unit extraction module; 3. Second acquisition module; 4. Unit storage module; 5. Third acquisition module; 6. Unit retrieval module; 7. Sequence generation module; 8. Result generation module. Detailed Implementation

[0018] Firstly, this application discloses an information classification and retrieval method.

[0019] Reference Figure 1 An information classification and retrieval method, comprising steps S101 to S108: Step S101: Obtain multi-source heterogeneous data streams from the futures market, which include structured data and unstructured text.

[0020] Specifically, in this embodiment, the multi-source heterogeneous data stream refers to a collection of futures-related data obtained from multiple different sources, with different data types and inconsistent structures. Structured data refers to quantitative data with a fixed format that can be directly stored and analyzed, including opening prices, closing prices, highest prices, lowest prices, trading volumes, and open interest for various contracts published by futures exchanges, as well as client trading records and position details provided by futures companies. Unstructured text refers to textual information without a fixed format that cannot be directly quantified, including financial news, industry research reports, policy documents, analyst comments, and social media discussions. By connecting to multiple data sources such as futures exchanges, financial information platforms, and industry databases through data interfaces, and employing a combination of real-time capture and scheduled updates, a complete multi-source heterogeneous data stream is obtained, providing a data foundation for subsequent information classification and retrieval.

[0021] Step S102: Perform natural language parsing on the unstructured text to extract multiple information units to be classified.

[0022] Specifically, in this embodiment, natural language parsing refers to using Natural Language Processing (NLP) technology to process unstructured text, breaking down the chaotic text into the smallest information units with independent semantics, i.e., information units to be classified. The specific processing includes: first, preprocessing the unstructured text to remove irrelevant symbols, redundant sentences, and duplicate content, completing word segmentation, part-of-speech tagging, and stop word removal; then, extracting key information from the text using Named Entity Recognition (NER) technology, and then dividing the extracted key information into independent information units to be classified according to semantic relevance. Each information unit to be classified corresponds to a piece of content with complete semantics, and each information unit to be classified has independent information value and can be classified and matched independently.

[0023] Step S103: Match the information units to be classified with the dynamically updated futures product-entity mapping map to obtain the preliminary product label and associated entity chain corresponding to each information unit to be classified.

[0024] Specifically, in this embodiment, the futures product-entity mapping graph refers to a network-like association graph constructed with futures products as core nodes and various entities related to the products as associated nodes. The core nodes are futures products (such as copper, crude oil, soybeans, etc.), and the associated nodes include entities such as product codes, contract names, exchange abbreviations, upstream and downstream products, macroeconomic indicators, and policy-issuing entities. The graph supports dynamic updates, allowing for the supplementation and improvement of nodes and relationships based on market changes and new information. The preliminary product label refers to the core futures product identifier corresponding to the information unit to be classified, determined through matching, used to identify the futures product category to which the information unit belongs. The associated entity chain refers to a set of entities associated with the core futures product in the information unit to be classified, used to reflect the association logic between the information unit and the product. By matching the entities in the information unit to be classified with the nodes in the graph, the preliminary product classification of the information unit is achieved, and the associated entity chain is extracted, providing a classification basis for subsequent information storage and retrieval.

[0025] Step S104: Based on the preliminary variety label, store the information unit to be classified into the corresponding variety original information pool.

[0026] Specifically, in this embodiment, the original commodity information pool refers to an information storage set constructed according to the classification of futures commodities. Each futures commodity corresponds to an independent original commodity information pool, which is used to store all unclassified information units that match the commodity tag. During the storage process, the publication time, source, associated entity chain, and other attribute information of each unclassified information unit are recorded simultaneously, which facilitates quick retrieval and filtering during subsequent searches, realizing the classified storage and orderly management of information.

[0027] Step S105: Obtain the personalized information profile of the target user, which includes a static follow list and dynamic behavioral characteristics.

[0028] Specifically, in this embodiment, the target user refers to a user who uses this retrieval method to retrieve futures information, and the personalized information profile refers to a model constructed based on the user's static settings and dynamic behavior to describe the user's information needs and preferences. The static watchlist refers to a list of futures contracts that the user actively sets and follows over a long period, representing the user's fixed information needs; dynamic behavioral characteristics refer to the collection of various behavioral data generated by the user during the use of the retrieval system, including click sequences, dwell time, sharing behavior, and collection records, which can reflect the user's real-time information preferences and changes in needs, providing a dynamic basis for personalized retrieval.

[0029] Step S106: Based on the static attention list, recall candidate information units from the original information pool of varieties.

[0030] Specifically, in this embodiment, recall refers to the process of extracting all information units to be classified from the corresponding original information pool of varieties based on the user's static watchlist. Candidate information units refer to the set of information units that initially meet the user's basic needs after recall. The static watchlist enables preliminary screening of information, ensuring that all recalled candidate information units are related to the varieties the user is interested in, filtering out information about irrelevant varieties, and improving retrieval efficiency.

[0031] Step S107: Reorder candidate information units by dynamic behavioral features to generate a preliminary retrieval sequence.

[0032] Specifically, in this embodiment, relevance reordering refers to the process of prioritizing recalled candidate information units based on users' dynamic behavioral characteristics. The core is to determine the degree of matching between candidate information units and user needs based on user behavioral preferences, placing information units with high matching degrees at the top. The preliminary retrieval sequence refers to the sequence of information units arranged according to relevance priority after reordering. By analyzing users' dynamic behavioral characteristics and mining user preferences for information type, publication source, content depth, etc., candidate information units are accurately ranked, allowing users to see information that matches their preferences first.

[0033] Step S108: Perform time-sensitivity decay calculation and redundancy elimination on the candidate information units in the preliminary retrieval sequence to generate the final retrieval results.

[0034] Specifically, in this embodiment, timeliness decay calculation refers to calculating the timeliness value of information based on the publication time and latest citation time of the information unit, and filtering out information with low timeliness value, because futures market information has extremely high timeliness, and outdated information may lose its reference value; redundancy elimination refers to identifying information units with similar content in the initial search sequence, eliminating duplicate or highly similar redundant information, and avoiding users obtaining duplicate content. The final search result refers to the accurate, efficient, and redundancy-free information sequence formed after timeliness filtering and redundancy elimination, which can meet users' personalized and high-quality search needs for futures information.

[0035] Reference Figure 2 In one embodiment of this example, the step S103 of matching the information unit to be classified with the dynamically updated futures product-entity mapping map includes steps S201 to S205: Step S201: Identify the core entity words in the information unit to be classified. The core entity words include product code, contract name, and exchange abbreviation.

[0036] Specifically, in this embodiment, core entity words refer to entity words that can directly and uniquely identify futures contracts. They are the core basis for matching the information units to be classified with the futures contract-entity mapping graph. The contract code refers to the unique code assigned by the futures exchange to each futures contract, the contract name refers to the specific futures contract identifier, and the exchange abbreviation refers to the abbreviation of the futures exchange. Through named entity recognition technology, the aforementioned core entity words are accurately identified from the information units to be classified. If the core entity words can be identified, rapid matching with the graph can be directly achieved.

[0037] Step S202: If a core entity word is identified, the core entity word is directly mapped to a preliminary variety label, and the corresponding entity's associated entity chain in the futures variety-entity mapping graph is extracted.

[0038] Specifically, in this embodiment, if a core entity word is identified in the information unit to be classified, such as identifying the core entity word "Shanghai Copper 2405 Contract" in "Today's Market Price of Shanghai Copper 2405 Contract", then the futures contract "Shanghai Copper" corresponding to the core entity word is directly used as the initial contract label. At the same time, all entities associated with the "Shanghai Copper" node are extracted from the futures contract-entity mapping graph to form an associated entity chain. For example, the associated entity chain includes "CU (contract code), Shanghai Futures Exchange (exchange abbreviation), copper concentrate (upstream product), wire and cable (downstream product), LME copper price (associated indicator)", etc., fully presenting the relationship between the information unit to be classified and the Shanghai Copper contract.

[0039] Step S203: If no core entity words are identified, extract the marginal entity words from the information unit to be classified. Marginal entity words include upstream and downstream product names, macroeconomic indicators, and policy issuing entities.

[0040] Specifically, in this embodiment, marginal entity words refer to entity words that cannot directly identify futures contracts but are indirectly related to them. These are used to match futures contracts through indirect association when core entity words are not identified. Among them, upstream and downstream product names refer to the upstream raw materials and downstream processed products of the futures contract; macroeconomic indicators refer to macroeconomic data affecting the price of the futures contract; and policy issuing entities refer to the institutions that issue policies related to the futures contract. When core entity words are not identified, marginal entity words are extracted from the information units to be classified as the basis for indirect matching.

[0041] Step S204: Calculate the semantic similarity and co-occurrence frequency of marginal entity words and each commodity node in the futures commodity-entity mapping graph.

[0042] Specifically, in this embodiment, semantic similarity refers to the degree of semantic association between marginal entity words and futures commodity nodes, calculated using a semantic vector model in natural language processing. Higher semantic similarity indicates a stronger association between the marginal entity word and the commodity. Co-occurrence frequency refers to the number of times a marginal entity word and a futures commodity node appear simultaneously in historical information; higher co-occurrence frequency indicates a more stable association. By calculating the semantic similarity and co-occurrence frequency between marginal entity words and all commodity nodes in the graph, a quantitative basis is provided for subsequent commodity matching.

[0043] Step S205: Select the variety nodes with the highest semantic similarity and co-occurrence frequency exceeding the low frequency threshold as the initial variety labels, and store the edge entity words in the matching process as supplementary entity chains into the associated entity chain.

[0044] Specifically, in this embodiment, the low-frequency threshold refers to a preset co-occurrence frequency threshold, used to filter out variety nodes with too low association to avoid matching errors. First, several variety nodes with the highest semantic similarity are selected, and then variety nodes with a co-occurrence frequency exceeding the low-frequency threshold are selected as the initial variety labels for the information unit to be classified. At the same time, marginal entity words are stored in the associated entity chain as supplementary entity chains to improve the association logic between the information unit and the variety.

[0045] Reference Figure 3 In one embodiment of this example, step S204, calculating the semantic similarity and co-occurrence frequency between edge entity words and nodes of each commodity in the futures commodity-entity mapping graph, includes steps S301 to S305: Step S301: Determine whether the marginal entity words are sudden new words or out-of-vocabulary words.

[0046] Specifically, in this embodiment, "emerging neologisms" refer to words that newly emerge due to sudden market events, policy changes, or other unforeseen circumstances and are not included in the futures product-entity mapping graph; "out-of-vocabulary words" refer to words that exist but are not included in the graph, and may be niche terms, technical terms, or emerging expressions. By comparing marginal entity words with existing entity words in the graph, it is determined whether they belong to emerging neologisms or out-of-vocabulary words. If they do, they need to be matched using special methods to avoid matching failures due to the words not being included in the graph.

[0047] Step S302: If it is a newly emerging word, obtain the context of the corresponding word in recent authoritative information.

[0048] Specifically, "recent authoritative information" refers to futures-related information released by reputable financial media, industry institutions, and official platforms within the past 30 days, which carries a high degree of credibility. "Contextual context" refers to the core content of the sentence, paragraph, and entire article containing the novel term, used to analyze its meaning and related objects. This step involves capturing recent authoritative information containing the novel term, extracting its contextual context, and providing a basis for subsequent semantic reasoning. Through this contextual context, its relationship with new energy futures products can be analyzed.

[0049] Step S303: Based on the context, infer the range of varieties associated with the corresponding words through a real-time semantic reasoning engine.

[0050] Specifically, in this embodiment, the real-time semantic reasoning engine refers to an algorithmic model with real-time semantic analysis and reasoning capabilities. It can infer the association between emerging new terms and futures contracts based on contextual information, combined with industry knowledge and historical data from the futures market. The associated contract range refers to a set of one or more futures contracts that the new term may correspond to. For example, through contextual analysis, it can be inferred that "new energy storage raw material A" may be associated with futures contracts such as "thermal coal" and "lithium battery," forming a corresponding contract range.

[0051] Step S304: If the inferred range of varieties is a single variety, then the single variety is used as the matching result, and the update mechanism of the futures variety-entity mapping map is triggered.

[0052] Specifically, in this embodiment, if the real-time semantic reasoning engine infers a single commodity range, it indicates a clear association between the emerging new word and that commodity. This single commodity can then be directly used as the matching result, i.e., the initial commodity label. Simultaneously, the update mechanism of the futures commodity-entity mapping graph is triggered, establishing an association between the emerging new word as a related node and the corresponding futures commodity node. This enables dynamic updates to the graph, facilitating rapid matching of similar information units in the futures market.

[0053] Step S305: If the inferred variety range consists of multiple candidate varieties, obtain the historical frequency distribution of the corresponding words in the relevant information of different candidate varieties, and take the candidate variety with the highest frequency as the matching result.

[0054] Specifically, in this embodiment, multiple candidate varieties refer to multiple futures varieties that are inferred to be potentially associated with the emerging new term; the historical frequency distribution refers to the statistical results of the number of times the emerging new term appears in the historical information related to each candidate variety. By statistically analyzing the historical frequency of the new term in the information related to each candidate variety, the candidate variety with the highest frequency of occurrence is selected as the matching result, i.e., the preliminary variety label, to ensure the accuracy of the matching results.

[0055] Reference Figure 4 In one embodiment of this example, the mechanism for triggering the update of the futures product-entity mapping map in step S304 includes steps S401 to S404: Step S401: Obtain the actual co-occurrence data of the sudden new words with each variety node in the subsequent preset time period.

[0056] Specifically, in this embodiment, the subsequent preset time period refers to the preset observation period (such as 90 days) after the update mechanism is triggered, which is used to collect the actual correlation data between the sudden new words and each variety node; the actual co-occurrence data refers to the number of times and frequency of the sudden new words and each futures variety node appearing simultaneously in various information within the preset time period, which is used to verify whether the inference results of the real-time semantic reasoning engine are accurate.

[0057] Step S402: Determine whether the actual co-occurring data supports the inference results of the real-time semantic reasoning engine.

[0058] Specifically, in this embodiment, the judgment criterion is a preset co-occurrence support threshold. If the actual co-occurrence frequency of the sudden new word and the inferred variety node exceeds the threshold, it indicates that the actual data supports the inference result, that is, the new word and the variety do have a stable association. If the actual co-occurrence frequency does not exceed the threshold, it indicates that the inference result is inaccurate and the actual association is low.

[0059] Step S403: If the inference result is supported, the sudden new words and their associated commodity nodes are permanently written into the futures commodity-entity mapping graph and their confidence level is marked.

[0060] Specifically, in this embodiment, permanent writing refers to using emerging new words as associated nodes, establishing a fixed association with the corresponding futures product nodes, storing them in the futures product-entity mapping graph, and not easily deleting them later; confidence level refers to the reliability of the association between the new word and the product node based on actual co-occurrence data (such as high, medium, and low levels), and the higher the co-occurrence frequency, the higher the confidence level.

[0061] Step S404: If the inference result is not supported, the sudden new words are stored in the temporary observation area, and the weight of similar inference paths in the real-time semantic inference engine is reduced.

[0062] Specifically, in this embodiment, the temporary observation area refers to the area in the futures product-entity mapping graph specifically used to store new words with unconfirmed relationships. It does not establish a fixed relationship with any product node, and its co-occurrence data will continue to be collected subsequently. The similar reasoning path refers to the reasoning logic and algorithm path in the real-time semantic reasoning engine used to infer the product associated with the sudden new word. Reducing its weight is to reduce the error of subsequent similar reasoning and avoid inaccurate inference results from occurring again.

[0063] In one embodiment of this example, step S107, which reorders candidate information units based on dynamic behavioral features, includes steps S501 to S508: Step S501: Obtain the target user's click sequence, dwell time, sharing behavior, and collection records within the preset history window.

[0064] Specifically, in this embodiment, the preset history window refers to a preset time period (such as the last 60 days) used to collect user behavior data, avoiding the loss of reference value due to excessive time; the click sequence refers to the sequential record of information units clicked and viewed by the user in the retrieval system, reflecting the user's interest in different information; the dwell time refers to the duration for which the user views each information unit; the longer the dwell time, the higher the user's attention to the information; the sharing behavior refers to the user's behavior of sharing information units with others, reflecting the social value of the information; and the collection record refers to the user's behavior of adding information units to their personal favorites, reflecting the long-term reference value of the information. By collecting the above four types of user behavior data within the preset history window, a foundation is provided for extracting user preferences.

[0065] Step S502: Extract user click preferences for specific information types, specific data formats, and specific publishing sources from the click sequence.

[0066] Specifically, in this embodiment, "specific information type" refers to the content category of the information (such as in-depth analysis, breaking news, policy interpretation, data broadcasting, etc.); "specific data format" refers to the presentation form of the information (such as text, tables, charts, videos, etc.); and "specific publishing source" refers to the publishing entity of the information (such as official exchanges, well-known securities firms, authoritative financial media, etc.). By statistically analyzing the number of clicks on various types of information in the click sequence, user click preferences are extracted. For example, if users click on in-depth analysis information 70% of the time, it indicates that users prefer in-depth analysis information; if they click on table-format information more often, it indicates that users prefer information presented in a table format.

[0067] Step S503: Identify the degree of user attention to in-depth analysis information, news flash information, and data table information based on dwell time.

[0068] Specifically, in this embodiment, in-depth analysis information refers to information that provides detailed analysis and predictions of futures market trends and commodity price movements, with relatively in-depth content; news flash information refers to quickly released, concise, and clear updates on the futures market; and tabular information refers to futures data statistics presented in tabular form. By statistically analyzing the average time users spend viewing different types of information, the level of user attention can be identified.

[0069] Step S504: Map the sharing behavior to the user's first recognition label of the social value of the information.

[0070] Specifically, in this embodiment, social value refers to the attribute of information being suitable for sharing with others and having dissemination value; the first recognition tag is a tag used to quantify the degree of user recognition of the social value of information, expressed in numerical form (e.g., 1-5 points). The more times and the higher the sharing frequency, the higher the value of the first recognition tag.

[0071] Step S505: Map the collection records to a second recognition tag of the long-term value of the information by the user.

[0072] Specifically, in this embodiment, long-term value refers to the value of information for long-term reference and repeated review; the second recognition tag is a tag used to quantify the degree of user recognition of the long-term value of information, which is also expressed in numerical form (e.g., 1-5 points). The more times the information is collected and the higher the frequency of viewing it again after collection, the higher the value of the second recognition tag.

[0073] Step S506: Quantitatively integrate click preferences, attention levels, first recognition tags, and second recognition tags to generate a multi-dimensional behavior weight vector.

[0074] Specifically, in this embodiment, quantitative fusion refers to quantifying four types of indicators—click preference, attention level, first approval tag, and second approval tag—each (converting them into values ​​between 0 and 1), and then merging these four quantified indicators into a multi-dimensional behavioral weight vector according to a preset weight allocation rule. Each dimension of the behavioral weight vector corresponds to a type of user preference indicator; the higher the vector value, the more the user prefers the information features corresponding to that dimension.

[0075] Step S507: Calculate the matching degree between the feature vector and the behavior weight vector of each candidate information unit.

[0076] Specifically, in this embodiment, the feature vector of a candidate information unit refers to a multi-dimensional vector used to describe the features of the information unit. Its dimensions correspond one-to-one with the dimensions of the behavior weight vector. The value of each dimension represents the degree of conformity of the information unit with the corresponding feature (e.g., if the information unit is a deep analysis type or a table format, the value of its corresponding dimension is 1; for other types, it is below 0.5). The matching degree refers to the degree of similarity between the feature vector of the candidate information unit and the user's behavior weight vector, calculated using algorithms such as vector dot product and cosine similarity. The higher the matching degree value, the more closely the information unit matches the user's preferences.

[0077] Step S508: Arrange the candidate information units in descending order according to the matching degree to generate a preliminary retrieval sequence.

[0078] Specifically, in this embodiment, descending order means arranging all candidate information units in descending order of matching degree, with the information unit with the highest matching degree at the beginning of the sequence and the one with the lowest matching degree at the end. The preliminary search sequence is a sequence of information units formed by this sorting method, which ensures that users see the information that best matches their preferences first, thus improving the personalization of the search.

[0079] In one embodiment of this example, after calculating the matching degree between the feature vector and the behavior weight vector of each candidate information unit in step S507, steps S601 to S605 are further included: Step S601: Obtain the special status identifier of the current market. Special status identifiers include the approaching delivery month, the eve of the release of major data, and the period of abnormal market fluctuations.

[0080] Specifically, in this embodiment, the current special state of the market refers to a special period in the futures market that may affect users' information needs; the special state identifier refers to a label used to identify these special periods. Specifically, "approaching delivery month" refers to the period when a futures contract is less than one month away from expiration and delivery, at which time users may pay more attention to delivery-related information; "before the release of major data" refers to 1-3 days before the release of important macroeconomic data (such as CPI, non-farm payroll data) or industry data, at which time users may pay more attention to data forecasts and interpretations; "period of abnormal market volatility" refers to a period when futures market prices experience significant increases, decreases, or violent fluctuations, at which time users may pay more attention to market analysis and risk warnings.

[0081] Step S602: Determine whether the special status indicator triggers the preset behavior drift rule.

[0082] Specifically, in this embodiment, the behavior drift rule refers to a preset rule used to adjust the user behavior weight vector in response to special market conditions. The core principle is that when the market is in a special state, users' information needs may deviate from their daily preferences, requiring temporary weight adjustments to adapt to these changes. Determining whether to trigger the behavior drift rule involves judging whether the current special state identifier belongs to a preset state type that requires weight adjustment. If it does, the rule is triggered; otherwise, it is not triggered, and the original behavior weight vector continues to be used.

[0083] Step S603: If the behavior drift rule is triggered, the weight values ​​of some dimensions in the behavior weight vector are temporarily adjusted to generate a temporary weight vector.

[0084] Specifically, in this embodiment, temporary adjustment refers to adjusting the weight values ​​of the dimensions related to the special state in the behavior weight vector only during the existence of the current special state, and restoring the original behavior weight vector after the special state ends; some dimensions refer to the preference dimensions related to user needs in the special state.

[0085] Step S604: Recalculate the matching degree of candidate information units using temporary weight vectors to generate temporary retrieval sequences.

[0086] Specifically, in this embodiment, the temporary retrieval sequence refers to a sequence of information units formed by recalculating the matching degree of each candidate information unit based on a temporary weight vector and then arranging them in descending order of matching degree. The temporary retrieval sequence can adapt to the information needs of users in special situations. For example, when the delivery month is approaching, delivery-related information units will be prioritized in the temporary retrieval sequence to meet the user's emergency needs.

[0087] Step S605: Weight and merge the temporary search sequence with the preliminary search sequence to generate a merged search sequence.

[0088] Specifically, in this embodiment, weighted fusion refers to setting the weight ratios of the temporary search sequence and the preliminary search sequence (e.g., the temporary search sequence accounts for 60%, and the preliminary search sequence accounts for 40%), calculating the matching degree of each information unit in the two sequences according to their weight ratios, obtaining the fused matching degree, and then arranging them in descending order of the fused matching degree to generate a fused search sequence. The fused search sequence takes into account both users' daily preferences and adapts to emergency needs under special market conditions, improving the rationality and practicality of the search results.

[0089] In one embodiment of this example, step S602, determining whether a special state identifier triggers a preset behavior drift rule, includes steps S701 to S703: Step S701: If the special status is marked as delivery month approaching, then identify whether the static watch list contains varieties that are approaching delivery.

[0090] Specifically, in this embodiment, "near-delivery commodity" refers to a futures commodity that is currently in the process of approaching its delivery month, that is, a certain contract of the commodity is less than one month away from expiration and delivery; identifying whether the static watch list contains near-delivery commodities is the core basis for determining whether a user has a need for delivery-related information.

[0091] Step S702: If it includes commodities nearing delivery, increase the weight value of information types related to delivery rules, warehouse receipt reports, and position limits.

[0092] Specifically, in this embodiment, delivery rules refer to the specific regulations for the delivery of futures contracts upon expiration (such as delivery method, delivery location, delivery grade, etc.); warehouse receipt reports refer to reports issued by futures exchanges that reflect the quantity and changes of warehouse receipts for futures products; and position limits refer to the maximum position limits for a particular contract stipulated by futures exchanges. These information types are all closely related to delivery and are core needs for users as the delivery month approaches.

[0093] Step S703: If the product is not close to delivery, determine whether the user has shown interest in the relevant product in the same period in history. If so, increase the number of information recalls for the relevant product.

[0094] Specifically, in this embodiment, "user's historical period" refers to the same time period of the previous year; "related commodities" refers to futures commodities of the same type and industry chain as the current near-delivery commodity; and "historical period attention behavior" refers to user actions such as clicking, saving, and sharing related commodities during the same time period of the previous year. If a user has historical attention behavior towards related commodities, it indicates that the user may still be interested in relevant information about the current near-delivery commodity. Therefore, the number of information retrieved for related commodities is increased, that is, more candidate information units are retrieved from the original information pool of related commodities to enrich the search results and meet the user's potential needs.

[0095] In one embodiment of this example, step S108, which involves calculating the timeliness attenuation and eliminating redundancy of candidate information units in the preliminary retrieval sequence, includes steps S801 to S806: Step S801: Obtain the publication time and latest citation time of each candidate information unit in the preliminary retrieval sequence.

[0096] Specifically, in this embodiment, the publication time refers to the exact time (accurate to the minute) when the information unit is first published; the latest citation time refers to the latest time when the information unit is cited by other information or users, which is used to help judge the timeliness of the information. If the information unit is cited multiple times and the latest citation time is relatively recent, it indicates that the information still has high reference value.

[0097] Step S802: Calculate the timeliness value decay curve for each candidate information unit based on the publication time and the latest citation time.

[0098] Specifically, in this embodiment, the timeliness value decay curve refers to a curve used to describe the change in the timeliness value of an information unit over time. Its core calculation logic is: the longer the publication time, the greater the timeliness value decay; the more recent the citation time, the less the timeliness value decays. Through a preset decay algorithm, combined with the publication time and the latest citation time, the timeliness value of each information unit at the current time point is calculated. The higher the timeliness value, the higher the reference value of the information.

[0099] Step S803: If the publication time of a candidate information unit exceeds the preset hard timeliness threshold, the corresponding candidate information unit will be removed from the preliminary search sequence to obtain the first filtered sequence.

[0100] Specifically, in this embodiment, the hard timeliness threshold refers to the preset maximum time during which an information unit loses its reference value (e.g., the hard timeliness threshold for futures news is 7 days, and for in-depth research reports it is 30 days). Different types of information units correspond to different hard timeliness thresholds. By determining whether the publication time of each candidate information unit exceeds its corresponding hard timeliness threshold, expired information units are eliminated to obtain the first filtered sequence, ensuring the timeliness of the search results.

[0101] Step S804: Perform semantic fingerprint calculation on the remaining candidate information units in the first filtered sequence, identify candidate information units whose content similarity exceeds the similarity threshold, and cluster the corresponding candidate information units into redundant information groups.

[0102] Specifically, in this embodiment, semantic fingerprint calculation refers to extracting the core semantic features of each information unit using natural language processing technology to generate a unique semantic fingerprint (similar to an "information ID card"). The more similar the semantic fingerprints, the closer the content of the information units are. The similarity threshold is a preset critical value (e.g., 85%) used to determine whether an information unit is redundant. If the semantic fingerprint similarity of two information units exceeds this threshold, they are determined to be redundant information. Redundant information groups refer to grouping multiple information units whose content similarity exceeds the threshold into one group. The information units within each group have highly similar content and are considered redundant information.

[0103] Step S805: Retain the candidate information unit with the highest source authority or the most citations in each redundant information group, and mark the remaining candidate information units as redundant and remove them from the first filtered sequence to obtain the second filtered sequence.

[0104] Specifically, in this embodiment, source authority refers to the credibility and professionalism of the information publisher, with preset authority levels (e.g., official exchanges > well-known securities firms > ordinary financial media). The higher the level, the higher the source authority. Citation count refers to the total number of times the information unit is cited by other information or users; the more citations, the higher the recognition of the information. In each redundant information group, the information unit with the highest source authority is retained first. If multiple information units have the same source authority, the information unit with the most citations is retained, and the remaining redundant information units are removed to obtain the second filtered sequence, thus eliminating redundancy.

[0105] Step S806: Take the candidate information units in the second filtered sequence as the final retrieval result.

[0106] Specifically, in this embodiment, the second filtered sequence is an information unit sequence that has undergone timeliness filtering and redundancy elimination. This ensures the timeliness of the information while avoiding content redundancy, and also takes into account users' personalized preferences and the needs of special market conditions. Presenting this sequence as the final search result to the user allows them to quickly obtain accurate, efficient, and non-redundant futures information, improving the search experience and information acquisition efficiency.

[0107] In one embodiment of this example, before reserving the candidate information unit with the highest source authority or the most citations in each redundant information group in step S805, steps S901 to S905 are included: Step S901: Determine whether multiple candidate information units in the redundant information group contain contradictory core viewpoints.

[0108] Specifically, in this embodiment, the core viewpoint refers to the judgments and conclusions made by the information unit regarding the core content of futures product trends, market trends, and policy impacts; contradictory core viewpoints refer to the opposing core viewpoints of different information units within the same redundant information group (e.g., information unit A believes that the price of Shanghai copper will rise, while information unit B believes that the price of Shanghai copper will fall). By extracting the core viewpoint of each information unit, the system compares and analyzes whether there are contradictions in the viewpoints within the same redundant information group. If so, special processing is required to avoid users missing opposing viewpoints due to the removal of redundant information.

[0109] Step S902: If there are contradictory core viewpoints, identify the source authority and data support strength of the candidate information unit corresponding to each contradictory core viewpoint.

[0110] Specifically, in this embodiment, data support strength refers to the sufficiency, accuracy, and authority of the data upon which the core viewpoint of an information unit is based. The more sufficiency of the data and the more authoritative the source, the higher the data support strength (e.g., a viewpoint based on official exchange data has higher data support strength than a viewpoint based on unofficial data). After identifying contradictory core viewpoints, the source authority and data support strength of each viewpoint's corresponding information unit are statistically analyzed to provide a basis for retaining the viewpoints in the future.

[0111] Step S903: Retain the first candidate information unit with the highest source authority and the strongest data support in the redundant information group.

[0112] Specifically, in this embodiment, the first candidate information unit refers to the information unit with the highest source authority and the strongest data support in the same redundant information group, and its core viewpoint has high credibility.

[0113] Step S904: From the candidate information units in the same redundant information group that oppose the viewpoint of the first candidate information unit, select the second candidate information unit with the second highest source authority and the second strongest data support.

[0114] Specifically, in this embodiment, the second candidate information unit refers to the information unit that opposes the viewpoint of the first candidate information unit and, among the information units corresponding to the opposing viewpoint, has the second highest source authority and the second strongest data support. Retaining this information unit is to allow users to understand different opposing viewpoints, comprehensively grasp market information, and avoid the limitations of a single viewpoint.

[0115] Step S905: Simultaneously retain the first candidate information unit and the second candidate information unit, and when displaying the candidate information units in the second filtered sequence as the final search results, add viewpoint comparison labels to the first candidate information unit and the second candidate information unit.

[0116] Specifically, in this embodiment, the viewpoint comparison marker refers to a special mark (such as "opposing viewpoints" or "bullish / bearish divergence") used to identify the opposing core viewpoints of two information units. This marker is appended to the title or summary of the information unit, making it easy for users to quickly identify opposing viewpoints. Simultaneously, retaining the first and second candidate information units avoids redundant information and provides users with a comprehensive reference of viewpoints, helping them to more objectively judge market trends and improve the practicality of the search results.

[0117] Secondly, this application also discloses an information classification and retrieval system.

[0118] Reference Figure 5 An information classification and retrieval system, comprising: The first acquisition module is used to acquire multi-source heterogeneous data streams from the futures market, which include structured data and unstructured text. The unit extraction module is used to perform natural language parsing on unstructured text and extract multiple information units to be classified. The second acquisition module is used to match the information units to be classified with the dynamically updated futures product-entity mapping map to obtain the preliminary product label and associated entity chain corresponding to each information unit to be classified. The unit storage module is used to store the information units to be classified into the corresponding original information pool of the variety based on the preliminary variety label; The third acquisition module is used to acquire the personalized information profile of the target user, which includes a static follow list and dynamic behavioral characteristics. The unit retrieval module is used to retrieve candidate information units from the original information pool of varieties based on the static attention list; The sequence generation module is used to reorder candidate information units based on dynamic behavioral features to generate a preliminary retrieval sequence. The results generation module is used to perform time-sensitivity decay calculation and redundancy elimination on candidate information units in the preliminary retrieval sequence to generate the final retrieval results.

[0119] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An information classification search method characterized by, include: Acquire multi-source heterogeneous data streams from the futures market, wherein the multi-source heterogeneous data streams include structured data and unstructured text; Natural language parsing is performed on the unstructured text to extract multiple information units to be classified; The information units to be classified are matched with the dynamically updated futures product-entity mapping map to obtain the preliminary product label and associated entity chain corresponding to each information unit to be classified. Based on the preliminary variety label, the information unit to be classified is stored in the corresponding variety original information pool; Obtain a personalized information profile of the target user, the personalized information profile including a static follow list and dynamic behavioral characteristics; Based on the static attention list, candidate information units are recalled from the original information pool of the varieties; The candidate information units are reordered based on their relevance using the dynamic behavioral features to generate a preliminary retrieval sequence. The candidate information units in the preliminary retrieval sequence are subjected to time-sensitivity decay calculation and redundancy elimination to generate the final retrieval results.

2. The information classification search method according to claim 1, wherein Matching the information units to be classified with the dynamically updated futures product-entity mapping map includes: Identify the core entity words in the information unit to be classified, including product code, contract name, and exchange abbreviation; If the core entity word is identified, the core entity word is directly mapped to a preliminary variety label, and the corresponding entity's associated entity chain in the futures variety-entity mapping graph is extracted. If the core entity words are not identified, then the edge entity words in the information unit to be classified are extracted. The edge entity words include upstream and downstream product names, macroeconomic indicators, and policy issuing entities. Calculate the semantic similarity and co-occurrence frequency between the marginal entity words and each variety node in the futures variety-entity mapping graph; The variety nodes with the highest semantic similarity and co-occurrence frequency exceeding the low-frequency threshold are selected as the initial variety labels, and the edge entity words in the matching process are stored as supplementary entity chains in the associated entity chain.

3. The information classification search method according to claim 2, wherein Calculating the semantic similarity and co-occurrence frequency between the marginal entity words and each commodity node in the futures commodity-entity mapping graph includes: Determine whether the edge entity words belong to sudden new words or out-of-vocabulary words; If it is a newly emerging word, then obtain the context of the corresponding word in recent authoritative information; Based on the aforementioned context, the range of varieties associated with the corresponding words is inferred through a real-time semantic reasoning engine; If the inferred range of varieties is a single variety, then the single variety is used as the matching result, and the update mechanism of the futures variety-entity mapping map is triggered. If the inferred range of varieties includes multiple candidate varieties, the historical frequency distribution of the corresponding words in the relevant information of different candidate varieties is obtained, and the candidate variety with the highest frequency is taken as the matching result.

4. The information classification search method according to claim 3, wherein The mechanisms that trigger the update of the futures contract-entity mapping map include: Obtain the actual co-occurrence data of the newly emerging words with each variety node within a subsequent preset time period; Determine whether the actual co-occurrence data supports the inference results of the real-time semantic reasoning engine; If the inference result is supported, the sudden new word and its associated variety node will be permanently written into the futures variety-entity mapping graph and its confidence level will be marked. If the inference result is not supported, the sudden new word will be stored in a temporary observation area, and the weight of similar inference paths in the real-time semantic reasoning engine will be reduced.

5. The information classification and retrieval method according to claim 1, characterized in that, Re-ranking the candidate information units based on the dynamic behavioral features includes: Obtain the target user's click sequence, dwell time, sharing behavior, and collection records within a preset history window; Extract user click preferences for specific information types, specific data formats, and specific publishing sources from the click sequence; The user's level of attention to in-depth analysis information, news flash information, and data table information can be identified from the dwell time. The sharing behavior is mapped to the user's first recognition label of the social value of information; The collection records are mapped as a second recognition tag of the long-term value of information by the user; The click preference, the degree of attention, the first approval tag, and the second approval tag are quantitatively fused to generate a multi-dimensional behavior weight vector; Calculate the matching degree between the feature vector of each candidate information unit and the behavior weight vector; The candidate information units are sorted in descending order based on the matching degree to generate a preliminary retrieval sequence.

6. The information classification and retrieval method according to claim 5, characterized in that, After calculating the matching degree between the feature vector of each candidate information unit and the behavior weight vector, the method further includes: Obtain the special status identifier of the current market, including the approaching delivery month, the eve of the release of major data, and the period of abnormal market fluctuations; Determine whether the special state identifier triggers a preset behavior drift rule; If the behavior drift rule is triggered, the weight values ​​of some dimensions in the behavior weight vector are temporarily adjusted to generate a temporary weight vector. The matching degree of the candidate information unit is recalculated using the temporary weight vector to generate a temporary retrieval sequence; The temporary retrieval sequence and the preliminary retrieval sequence are weighted and fused to generate a fused retrieval sequence.

7. The information classification and retrieval method according to claim 6, characterized in that, Determining whether the special state identifier triggers a preset behavior drift rule includes: If the special status is identified as the delivery month is approaching, then it is determined whether the static watchlist contains varieties that are approaching delivery. If it includes commodities nearing delivery, the weighting of information types related to delivery rules, warehouse receipt reports, and position limits will be increased; If the product is not close to delivery, it is determined whether the user has shown interest in the relevant product in the same period in history. If so, the number of information recalls for the relevant product is increased.

8. The information classification and retrieval method according to claim 1, characterized in that, Performing time-sensitivity decay calculations and redundancy elimination on candidate information units in the preliminary retrieval sequence includes: Obtain the publication time and latest citation time of each candidate information unit in the preliminary retrieval sequence; Based on the publication time and the latest citation time, calculate the timeliness value decay curve for each candidate information unit; If the publication time of a candidate information unit exceeds a preset hard timeliness threshold, the corresponding candidate information unit will be removed from the preliminary search sequence to obtain the first filtered sequence. Semantic fingerprinting is performed on the remaining candidate information units in the first filtered sequence to identify candidate information units whose content similarity exceeds the similarity threshold, and the corresponding candidate information units are clustered into redundant information groups. In each redundant information group, retain the candidate information unit with the highest source authority or the most citations, and mark the remaining candidate information units as redundant and remove them from the first filtered sequence to obtain the second filtered sequence; The candidate information units in the second filtered sequence are used as the final retrieval results.

9. The information classification and retrieval method according to claim 8, characterized in that, Before retaining the candidate information unit with the highest source authority or the most citations in each redundant information group, the following is also included: Determine whether multiple candidate information units in the redundant information group contain contradictory core viewpoints; If the core viewpoints are contradictory, then identify the source authority and data support strength of each candidate information unit corresponding to the contradictory core viewpoints. The first candidate information unit with the highest source authority and the strongest data support in the redundant information group is retained. From the candidate information units in the same redundant information group that have a viewpoint that is opposed to the first candidate information unit, select the second candidate information unit with the second highest source authority and the second strongest data support. At the same time, the first candidate information unit and the second candidate information unit are retained, and when the candidate information units in the second filtered sequence are displayed as the final search results, viewpoint comparison identifiers are added to the first candidate information unit and the second candidate information unit.

10. An information classification and retrieval system, characterized in that, include: The first acquisition module is used to acquire multi-source heterogeneous data streams from the futures market, wherein the multi-source heterogeneous data streams include structured data and unstructured text; The unit extraction module is used to perform natural language parsing on the unstructured text and extract multiple information units to be classified. The second acquisition module is used to match the information unit to be classified with the dynamically updated futures product-entity mapping map to obtain the preliminary product label and associated entity chain corresponding to each information unit to be classified. The unit storage module is used to store the information unit to be classified into the corresponding original information pool of the variety based on the preliminary variety label. The third acquisition module is used to acquire the personalized information profile of the target user, which includes a static follow list and dynamic behavioral characteristics. The unit retrieval module is used to recall candidate information units from the original information pool of the variety based on the static attention list. The sequence generation module is used to reorder the candidate information units based on the dynamic behavioral features to generate a preliminary retrieval sequence. The result generation module is used to perform time-sensitivity decay calculation and redundancy elimination on the candidate information units in the preliminary retrieval sequence to generate the final retrieval results.

Images