Method and device for processing multi-source heterogeneous information based on large model

Abstract

The disclosure provides a large model-based multi-source heterogeneous information processing method and related device, and relates to the technical field of computers, especially the technical fields of large models, content understanding, big data processing, etc. The specific implementation scheme is: abstracting unstructured data on a time series of multiple data sources to obtain first type data; the unstructured data at least includes at least one of text, image and audio; aligning the first type data and structured second type data in time to obtain multi-source heterogeneous information corresponding to multiple time windows; the second type data is index data expressed by digital quantization; inputting prompt words containing the multi-source heterogeneous information into a target large model, so that the target large model performs the following operations: generating an initial prediction result of a target task based on the multi-source heterogeneous information; filtering a plurality of reference items from the multi-source heterogeneous information to verify the initial prediction result.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, particularly to the fields of large models, content understanding, and big data processing. Background Technology

[0002] Against the backdrop of the rapid development of information technology, the information environment processed by large models is increasingly showing complex characteristics of being multi-source and heterogeneous. Its sources cover different media, and its forms include both structured and unstructured data, with significant differences in the density of viewpoints.

[0003] The core dilemma brought about by this environment is that large models face a complex information environment, and the understanding of information content is prone to deviation. Summary of the Invention

[0004] This disclosure provides a method and related apparatus for processing multi-source heterogeneous information based on a large model.

[0005] According to one aspect of this disclosure, a method for processing multi-source heterogeneous information based on a large model is provided, comprising: Summarize unstructured data from time series data sources from multiple sources to obtain the first type of data; unstructured data includes at least one of text, images, and audio. By aligning the first type of data and the structured second type of data over time, we obtain multi-source heterogeneous information corresponding to multiple time windows; the second type of data consists of indicator data expressed using numerical quantification. Input prompts containing multi-source heterogeneous information into the target large model so that the target large model performs the following operations: Based on multi-source heterogeneous information, the initial prediction results for the target task are generated; Multiple reference items are selected from multi-source heterogeneous information to verify the initial prediction results.

[0006] According to another aspect of this disclosure, a multi-source heterogeneous information processing apparatus based on a large model is provided, comprising: The extraction module is used to extract summaries from unstructured data from time series data sources to obtain the first type of data; the unstructured data includes at least one of text, images, and audio. The alignment module is used to align the first type of data and the structured second type of data in time to obtain multi-source heterogeneous information corresponding to multiple time windows; the second type of data is indicator data expressed using numerical quantization. The prediction and validation module is used to input prompt words containing multi-source heterogeneous information into the target large model, so that the target large model can perform the following operations: Based on multi-source heterogeneous information, the initial prediction results for the target task are generated; Multiple reference items are selected from multi-source heterogeneous information to verify the initial prediction results.

[0007] According to another aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and The memory is communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform any of the methods described in the present disclosure.

[0008] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform any of the methods according to embodiments of this disclosure.

[0009] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements any of the methods according to embodiments of this disclosure.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a flowchart illustrating a multi-source heterogeneous information processing method based on a large model according to an embodiment of the present disclosure; Figure 2 This is a flowchart illustrating the process of extracting summaries from unstructured time-series data from multiple data sources according to an embodiment of this disclosure. Figure 3 This is a flowchart illustrating the process of obtaining multi-source heterogeneous information corresponding to multiple time windows according to another embodiment of the present disclosure; Figure 4 This is a flowchart illustrating the verification of initial prediction results according to an embodiment of the present disclosure. Figure 5 This is a flowchart illustrating the verification of the initial prediction result according to another embodiment of the present disclosure; Figure 6 This is a schematic diagram of the process of constructing a knowledge graph according to an embodiment of the present disclosure; Figure 7 This is a flowchart illustrating a prediction task based on a large model-based multi-source heterogeneous information processing method according to an embodiment of the present disclosure. Figure 8 This is a schematic diagram of the structure of a multi-source heterogeneous information processing device based on a large model according to another embodiment of the present disclosure; Figure 9 This is a block diagram of an electronic device used to implement the multi-source heterogeneous information processing method based on a large model according to the embodiments of this disclosure. Detailed Implementation

[0012] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] The terms “first,” “second,” etc., used in this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion, such as including a series of steps or units. A method, system, product, or apparatus is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0014] It should be noted that, unless it is explicitly stated that there is a sequential order of execution between different operations, or that there is a sequential order of execution between different operations in terms of technical implementation, the execution order between multiple operations may not be significant, and multiple operations may be executed simultaneously.

[0015] In a multi-source heterogeneous information environment, the core task of large models is to integrate, filter, analyze and interpret various types of information, thereby forming accurate understanding, judgment and prediction.

[0016] However, the inventors' research revealed that existing large-scale models still face significant technical bottlenecks when processing such complex information. On the one hand, information from different sources may contain contradictions, conflicts, or redundancy. Some information may even be false or one-sided. The format differences of heterogeneous information also increase the difficulty of information fusion and parsing in large-scale models. On the other hand, the training mechanism and reasoning logic of large-scale models have inherent limitations. When faced with massive and chaotic multi-source heterogeneous information, they are prone to cognitive biases: that is, large-scale models are easily influenced by the dominant influence of some information. After establishing a preliminary conclusion in advance, they tend to actively search, filter, and strengthen relevant information that supports the conclusion, forming a confirmation bias that favors the preliminary conclusion, while ignoring or weakening a few key pieces of information that contradict the preliminary conclusion and can refute it.

[0017] Such cognitive biases caused by interference can directly lead to a series of errors in information understanding, logical reasoning, and outcome prediction in large models. For example, in intelligent decision-making scenarios, large models may bias their acceptance of advantageous information from one source while ignoring negative information from other sources, resulting in decision recommendations that deviate from actual needs due to biased content understanding. In information retrieval and organization scenarios, they may strengthen information supporting preliminary conclusions while filtering out key counter-evidence, leading to a lack of objectivity and comprehensiveness in the output. In predictive analysis scenarios, biased conclusions may be formed due to interference from local information, thereby affecting the accuracy and reliability of prediction results.

[0018] In view of this, this disclosure provides a method for processing multi-source heterogeneous information based on a large model. The method compresses and aligns the multi-source heterogeneous data in time to complete the initial prediction. Then, highly reliable reference items are selected from the multi-source heterogeneous information to verify the initial prediction, thereby effectively verifying the initial prediction results, improving the confirmation bias problem of the large model, and thus improving the accuracy of understanding the content of the large model, thereby improving the reliability of the prediction results of the large model.

[0019] The methods provided in this disclosure can be applied to multiple fields such as autonomous driving, industrial equipment fault detection, and finance. This disclosure is applicable to fields that use large models for content understanding and prediction of multi-source heterogeneous data.

[0020] The method provided in this disclosure can be executed by a single computing device. However, if the performance of a single computing device is limited and insufficient to support at least one large model required by this disclosure, it can also be completed collaboratively by multiple computing devices. The difference lies in the case of multiple computing devices executing the method; the computing devices need to exchange their execution results.

[0021] like Figure 1 The diagram shown is a flowchart illustrating a multi-source heterogeneous information processing method based on a large model, as provided in this embodiment of the disclosure. The method includes the following: S101, extract summaries from unstructured data from time series data sources to obtain the first type of data; the unstructured data includes at least one of text, image and audio.

[0022] Among them, data sources refer to various sources that provide original information, including but not limited to: text databases, image libraries, audio streams, sensor data, business indicator libraries, media resources such as websites, announcements, and social data.

[0023] A time series is a sequence of data arranged in chronological order.

[0024] Unstructured data refers to data that lacks a fixed format, has no explicit field structure, and cannot be directly represented and managed using traditional relational database row-list tables.

[0025] In this embodiment of the disclosure, unstructured data includes at least one of text, images, and audio. For example, in a financial market data processing scenario, securities-related news, analyst research reports, and market sentiment commentary are examples of unstructured text data. In an autonomous driving scenario, data collected by sensors, such as image signals captured by a visual camera, are examples of unstructured image data. In an industrial scenario, the continuously collected audio signals of equipment operation for monitoring the equipment's status are examples of unstructured audio data.

[0026] Abstract extraction refers to the process of extracting core and key information from long or complex multi-source heterogeneous data using natural language processing or multimodal understanding technologies.

[0027] In practice, a progressive summary extraction method can be used to extract summaries from unstructured time-series data from multiple data sources, resulting in the first type of data. Progressive summary extraction can be understood as extracting summaries step-by-step within each time window. Following the total length of the time windows, after extracting the summary for the latest time window, summaries from the furthest time window are removed or further compressed. This creates a dynamically adaptive summary sequence that changes over time.

[0028] S102, align the first type of data and the structured second type of data in time to obtain multi-source heterogeneous information corresponding to multiple time windows; the second type of data is indicator data expressed using digital quantification.

[0029] The second type of data refers to indicator data expressed in numerical quantification, which can be organized, stored and parsed in a standardized format, and can directly quantify the attributes, states and relationships of things through quantitative or qualitative methods.

[0030] For example, in financial market data processing scenarios, this includes quantitative indicators related to the securities market, such as RSI (Relative Strength Index), MACD (Moving Average Convergence Divergence), Bollinger Bands, volume ratio, M2 (Money Supply 2) growth rate, government bond yield, PEG (Price / Earnings to Growth Ratio), and dividend yield. In autonomous driving scenarios, this includes real-time distance data collected by radar equipment, such as distance to obstacles ahead, safe distance to the vehicle in front, and vehicle speed, as well as obstacle coordinate data collected by lidar. In industrial scenarios, this includes real-time equipment operation quantitative indicators collected by sensors, such as specific values ​​of temperature, rotational speed, voltage, and current, and equipment operating time.

[0031] In implementation, since the first and second types of data differ in frequency and format, a unified time window is set, and time alignment is used to establish a correspondence between the first and second types of data within the same time window. A time window refers to a time interval formed by segmenting a continuous time axis according to a preset time granularity. The length of the time window can be set according to the needs of the scenario; for example, in financial market data processing scenarios, it can be divided by trading days, weeks, and months; in autonomous driving scenarios, it can be divided by seconds, minutes, and hours; and in industrial scenarios, it can be divided by minutes, hours, and days.

[0032] S103, Input prompt words containing multi-source heterogeneous information into the target large model so that the target large model performs the following operations: Operation 1: Generate initial prediction results for the target task based on multi-source heterogeneous information; The target task refers to the specific task to be completed based on multi-source heterogeneous information. It is understood that both the first and second types of data mentioned above are data collected specifically for the target task, and therefore can be used for prediction. Examples include predicting stock price trends in financial market data processing scenarios, providing early warnings of driving safety risks (such as obstacle trajectory prediction) in autonomous driving scenarios, and predicting equipment malfunctions in industrial scenarios.

[0033] The initial prediction result refers to the preliminary prediction conclusion of the target large model on the target task output based on the multi-source heterogeneous information of the input, combined with its own training parameters and inference logic.

[0034] During implementation, preset prediction prompts can be input into the target large model, which will then generate initial prediction results for the target task based on the prediction instructions in the prompts. The prediction instructions in the preset prompts can be designed in conjunction with the target task type, scenario type, and multi-source heterogeneous information, specifying at least one of the following: prediction direction, prediction method, and output format of the target large model. This ensures that the large model can accurately capture the requirements of the target task and output preliminary prediction conclusions that meet expectations.

[0035] Operation 2: Select multiple reference items from the multi-source heterogeneous information to verify the initial prediction results.

[0036] Reference items refer to information selected from multi-source heterogeneous information that is highly correlated with the initial prediction results and has high reliability. This information is key evidence or clues in multi-source heterogeneous information and can be used to verify the rationality and accuracy of the initial prediction results.

[0037] In implementation, the target big data model can first automatically select representative and highly relevant information from multi-source heterogeneous information as reference items, based on the target task requirements. Subsequently, the target big data model will perform reasoning analysis on each selected reference item according to the verification instructions (such as disproving instructions) in the prompt words, and generate verification information corresponding to each reference item that can directly verify the initial prediction result, or is related to the initial prediction result and can indirectly verify the initial prediction result. The initial prediction result will then be cross-verified using this verification information.

[0038] The large model involved in this disclosure typically refers to a large-scale pre-trained language model (LLM), which is an artificial intelligence model based on a deep learning architecture, pre-trained with massive amounts of multimodal or text data, and possessing powerful semantic understanding, text generation, logical reasoning, knowledge association, and task generalization capabilities. In the embodiments of this disclosure, the target large model can be at least one of a natural language processing large model and a multimodal large model; a suitable large model can be selected at appropriate steps.

[0039] During implementation, the aforementioned operations 1 and 2 are executed sequentially by the same target large model. This is to enable the target large model to progressively understand multi-source heterogeneous information. Specifically, through operation 1, the target large model can focus its limited attention on a small amount of key information after compression (i.e., after abstract extraction), improving the accuracy of its understanding of multi-source heterogeneous information and laying the foundation for reducing confirmation bias. Furthermore, through operation 2, the target large model selects relevant and reliable reference items as hard evidence, enabling it to understand the multi-source heterogeneous information and preliminary prediction results again, achieving a deeper understanding and verification of the preliminary prediction results, thereby further reducing confirmation bias.

[0040] In summary, this embodiment of the present disclosure obtains the first type of data by summarizing unstructured data from multiple data sources over time series. This comprehensively captures the semantic information of various types of unstructured data, including text, images, and audio. It effectively removes redundant information from large amounts of data, significantly reduces the length of information, and prevents the target large model from being distracted by a large amount of information. Furthermore, it improves the information processing efficiency of the target large model by reducing the information length. Summarizing also extracts key content from multi-source heterogeneous information, reducing cognitive biases in the target large model caused by fragmented or one-sided information, and further improving the target large model's information comprehension ability. By aligning the first type of data with structured second type of quantitative indicator data over time, unified multi-source heterogeneous information corresponding to multiple time windows can be achieved, compensating for the parsing difficulties caused by insufficient dimensionality, contradictions, or redundancy in single-type data. The prompt words containing this multi-source heterogeneous information are input into the target large model, allowing the model to generate initial prediction results for the target task based on this information. To further reduce cognitive bias, the initial prediction results are validated by selecting reference items from multi-source heterogeneous information. This allows the target large model to combine the initial prediction results and reference items to perform a secondary understanding and sorting of multi-source heterogeneous information, further suppressing cognitive interference caused by confirmation bias, improving the target large model's ability to understand the content of multi-source heterogeneous information, and thus improving the accuracy, objectivity, and reliability of the target task prediction results. This also reduces prediction bias problems caused by single data dimensions, lack of validation mechanisms, and susceptibility to local information dominance in related technologies.

[0041] In this embodiment of the disclosure, the initial prediction result for the target task is generated based on multi-source heterogeneous information, which can be achieved through the following steps: Step A1: Perform content understanding on the multi-source heterogeneous information to obtain the content understanding results; Content understanding refers to the process by which the target large model comprehensively analyzes, semantically mines, and performs association analysis on the multi-source heterogeneous information input, in order to transform multi-source heterogeneous information into structured semantic information that the target large model can directly use for reasoning.

[0042] Step A2: Generate initial prediction results for the target task based on the content understanding results.

[0043] That is, the target big model outputs preliminary prediction conclusions for the target task based on the content understanding results, combined with the requirements of the target task, its own trained knowledge, reasoning logic, and preset prediction prompts.

[0044] In this embodiment, by performing content understanding on multi-source heterogeneous information, the key features and inherent relationships within the information can be fully explored and integrated, avoiding information omissions or misunderstandings caused by dispersed information sources and significant structural differences. Based on this content understanding result, the initial prediction result for the target task is directly generated, effectively avoiding initial prediction biases caused by a single data source or incomplete information, and improving the rationality and basic accuracy of the overall prediction process.

[0045] In this embodiment of the disclosure, a summary extraction is performed on unstructured data from time series data sources from multiple data sources to obtain a first type of data, which can be as follows: Figure 2 As shown, it includes the following: S201, Perform preprocessing operations on unstructured data to obtain preprocessed data; the preprocessing operations include at least converting non-text data into text data.

[0046] Preprocessing refers to the process of transforming unstructured data with inconsistent formats into a unified data format.

[0047] During implementation, the appropriate preprocessing method can be selected based on the type of unstructured data: for text-based unstructured data, natural language processing techniques can be directly used to perform denoising, word segmentation, and preliminary screening of key information; for image-based unstructured data, corresponding image recognition techniques can be used to extract core features, key scenes, and effective information from the image, and then the extracted image information can be converted into corresponding text data; for audio-based unstructured data, audio signals can first be converted into corresponding text data using audio parsing techniques.

[0048] Preprocessing operations can also be set according to task requirements, such as data cleaning and deduplication.

[0049] S202, Extract a summary from the preprocessed data of the current time window in the time series.

[0050] During implementation, the system can extract prompts from the preset summary, call the large model to extract keywords and refine core sentences from the preprocessed data of the current time window, and obtain the summary of the current time window.

[0051] In practice, for information from multiple data sources, the information from each data source within the current time window can be classified first, and summaries can be extracted according to different category dimensions to obtain summaries for each category dimension. This will enable the target large model to understand and analyze the content of summaries that are not based on category dimensions.

[0052] S203: Concatenate the summary of the current time window with the summaries of multiple time windows preceding the current time window in chronological order to obtain a summary sequence.

[0053] In practice, the summary of the current time window can be connected sequentially with the summaries of multiple previous time windows in chronological order from early to late, thus obtaining a complete narrative summary sequence with temporal dependencies. This allows the target large model to perform content understanding and information mining based on temporal evolution.

[0054] In some embodiments, the target large model can only process a limited amount of information. If the excessively long information is simply truncated, it may result in the loss of key information and affect the accuracy of understanding the content of multi-source heterogeneous information.

[0055] S204. When the word length of the summary sequence exceeds the word threshold, a large summary extraction model is used to compress each summary in the summary sequence to obtain the first type of data.

[0056] In implementation, a word segmenter can be used to perform word segment statistics on the concatenated summary sequence and compare it with a preset word segment threshold. If the word segment length of the summary sequence does not exceed the word segment threshold, the summary sequence can be directly used as the first type of data. If the word segment length of the summary sequence exceeds the word segment threshold, a large-scale summary extraction model can be used to traverse each summary in the summary sequence. Then, each summary is rewritten and compressed in a structured manner, using more compact sentence structures to describe each summary while fully preserving key information. Finally, the compressed summaries are reassembled in chronological order, and the reassembled summary sequence is used as the first type of data.

[0057] In this embodiment, by uniformly converting non-text data into text and performing preprocessing, the format differences of multi-source heterogeneous unstructured data can be eliminated. Summarizing the current and historical time windows in chronological order creates a continuous summary sequence. This preserves temporal dependencies and continuous contextual information while ensuring the large model processes only historical and current data, preventing future data leakage. When the summary sequence exceeds a word threshold, each summary is compressed separately. This effectively controls the input length without disrupting the temporal structure and the core semantics of each window, allowing the target large model to allocate its limited attention to key information, improving the accuracy of content understanding of multi-source heterogeneous information, and ultimately enhancing the accuracy of prediction results.

[0058] In this embodiment of the disclosure, a large-scale abstract extraction model is used to compress each abstract in the abstract sequence to obtain the first type of data, which can be achieved based on the following steps: Step B1: Based on the summary extraction model, determine the importance of each summary according to the target task; Importance refers to the relevance of the information contained in each summary to the target task. For example, the higher the relevance of a summary to the target task and the greater its impact on task decision-making, the higher its importance; conversely, the lower the relevance and the smaller the impact, the lower its importance.

[0059] In implementation, the target task can be parsed based on the summary extraction model. Then, considering the semantic coherence of each summary context, semantic matching and feature extraction can be performed on each summary content in the summary sequence one by one, thereby assigning a corresponding importance level to each summary. The summary extraction model and the aforementioned target model can be different roles of the same model, or they can be models with different structures. The implementation can be set according to the task requirements, and this embodiment does not impose any restrictions on this.

[0060] Step B2: Determine the compression rate of each summary based on its importance and term threshold. Compression ratio refers to the ratio of the word length of the abstract after compression to that before compression, and is used to measure the degree of compression of the abstract.

[0061] During implementation, the total number of terms to be compressed can be determined first based on the term threshold and the total term length of the current summary sequence. After assessing the importance of each summary, a differentiated compression rate is determined independently for each summary: the higher the importance of the summary, the lower the compression rate, to retain its core information as much as possible and avoid losing key information; the lower the importance of the summary, the higher the compression rate, allowing for more simplification of redundant information to control the overall term length.

[0062] Step B3: Based on the compression ratio of each summary, the summary extraction model is used to perform information compression operations on each summary separately to obtain the first type of data.

[0063] During implementation, the abstract extraction model can be used to compress the abstracts of each time window one by one according to the compression ratio corresponding to each abstract. After all the abstracts are compressed, the compressed abstracts are spliced ​​together in the original chronological order to obtain the first type of data.

[0064] In this embodiment, the importance of each summary is first evaluated by a large-scale summary extraction model based on the target task, which can distinguish the contribution of each summary to the target task. Then, the compression ratio is determined by combining the importance and the word threshold, which can achieve differentiated compression while meeting the length limit and avoid the loss of key information caused by the same compression ratio. Finally, based on the compression ratio of each summary, information compression operation is performed on each summary to obtain the first type of data, which can improve the effectiveness and accuracy of subsequent multi-source heterogeneous information fusion and predictive inference.

[0065] In this embodiment of the disclosure, the initial indicator data of the current time window can be statistically calculated according to the indicator category of the target task to obtain the second type of data.

[0066] Among them, the indicator category refers to the quantitative indicator that is divided according to the target task and is used to distinguish at least one of the different dimensions, time granularities, and business meanings.

[0067] For example, in financial market data processing scenarios, indicators can be categorized into daily technical indicators, monthly fundamental indicators, and macro liquidity indicators. In autonomous driving scenarios, indicators can be categorized into real-time perception indicators, dynamic operation indicators, and system anomaly indicators. In industrial scenarios, indicators can be categorized into three types: real-time status indicators, trend deterioration indicators, and fault accumulation indicators.

[0068] In financial market data processing scenarios, daily technical indicators refer to quantitative indicators that reflect short-term price trends, momentum, and volatility based on daily trading data. Monthly fundamental indicators refer to indicators that reflect a company's medium- to long-term profitability, growth, and valuation based on monthly financial and operational data. Macroeconomic liquidity indicators are aggregate indicators used to describe market supply and demand for funds, interest rates, and the credit environment, and determine the discount rate for asset pricing.

[0069] In autonomous driving scenarios, real-time perception metrics refer to the accuracy, latency, and completeness of the autonomous driving system's perception layer in detecting the environment and obstacles. Dynamic operation metrics refer to the outputs of the autonomous vehicle's decision-making, planning, and control layer, reflecting the vehicle's driving stability and compliance. System anomaly metrics refer to deviations from normal operating ranges, malfunctions, and failures of autonomous driving monitoring software and hardware, serving as a safety fallback.

[0070] In industrial settings, real-time status indicators refer to the current operating conditions and physical quantities of industrial equipment, used for real-time monitoring and control. Trend degradation indicators reflect long-term trends of slow, hidden performance decline in equipment. Fault accumulation indicators are metrics that statistically measure the cumulative amounts of anomalies, overloads, and impacts to assess equipment damage and risk.

[0071] Initial indicator data refers to raw quantitative data that is collected directly without statistical calculation.

[0072] For example, in financial market data processing scenarios, raw quantitative data could be the original opening price, closing price, highest price, lowest price, real-time trading volume, and original trading value of a stock on a given day. In autonomous driving scenarios, raw quantitative data could be the vehicle's original speed, instantaneous speed, original radar distance, real-time steering angle, and instantaneous braking force collected in real time. In industrial scenarios, raw quantitative data could be real-time temperature, original rotational speed, instantaneous voltage, instantaneous current, original vibration amplitude, and real-time energy consumption directly collected by sensors.

[0073] Statistical calculation refers to the calculation methods used to sum, average, maximum / minimum, proportion, standard deviation, etc., of initial indicator data to form other characteristic indicators with analytical value.

[0074] For example, in financial market data processing scenarios, the second type of data is obtained by calculating moving averages, RSI, MACD, PEG ratios, etc., based on indicators such as real-time status indicators and trend deterioration indicators. In autonomous driving scenarios, the second type of data is obtained by statistically analyzing average vehicle speed, emergency braking frequency, etc., based on real-time perception indicators and dynamic operation indicators, based on raw vehicle speed and raw radar distance data. In industrial scenarios, the second type of data is obtained by statistically analyzing the mean, drift, and deterioration trend of raw parameters, based on real-time status indicators and trend deterioration indicators.

[0075] In this embodiment of the disclosure, the initial indicator data of the current time window is statistically calculated according to the indicator category of the target task, thereby realizing the structured regularization of the original quantitative information. This provides more standardized and reliable reference dimensions for the subsequent alignment of the second type of data with the first type of data in time and the formation of unified multi-source heterogeneous information, thereby effectively improving the accuracy of the content understanding of the target task and the accuracy of the prediction results.

[0076] In this embodiment of the disclosure, the first type of data and the structured second type of data are aligned in time to obtain multi-source heterogeneous information corresponding to multiple time windows, such as... Figure 3 As shown, it may include the following: S301, according to the time window to which the timestamp of each piece of information belongs, align the first type of data and the second type of data in time to obtain the initial data for each time window.

[0077] The first type of data uses timestamps corresponding to the time when its unstructured data was generated, such as the publication time of a text announcement, the time of image acquisition, or the time of audio recording. After abstract extraction, these timestamps can be retained as the time identifier for the first type of data. The second type of data uses timestamps corresponding to the collection time of its initial indicator data. After statistical calculation, the start / end time of the time window corresponding to the statistical calculation is retained as the time identifier for the second type of data.

[0078] In practice, the timestamps of each piece of data in the first category and each piece of data in the second category can be extracted one by one to determine the time window to which it belongs. The data in the first category and the second category belonging to the same time window can be time-aligned to obtain the data for each time window.

[0079] S302, convert the second type of data in the initial data of each time window into the corresponding text description, and obtain multi-source heterogeneous information corresponding to multiple time windows.

[0080] Since the second type of data is expressed in numerical quantification, to avoid long-tail knowledge bias when directly processed by a large model, the structured second type of data can be transformed into concise, easy-to-understand natural language text that can be integrated with the first type of data. During the transformation process, prompts can be used to constrain the generation behavior of the large model: the generated text must strictly correspond to the numerical meaning of the second type of data, without tampering with or distorting the numerical values ​​and significance of the quantification indicators, while simplifying redundant expressions, highlighting the core values ​​and indicator connotations, avoiding cumbersome expressions, and maintaining consistency with the summary style of the first type of data.

[0081] During implementation, after completing the text conversion of the second type of data, the first type of data for each time window is integrated with the textual description of the converted second type of data to form multi-source heterogeneous information corresponding to that time window.

[0082] In this embodiment, by aligning the first and second types of data according to the time window assigned by timestamps, misalignment and inconsistency in the time dimension of multi-source data can be eliminated. Furthermore, by uniformly converting the second type of data into textual representations, standardized fusion of heterogeneous data formats can be achieved, forming complete multi-source heterogeneous information within the same time window. This provides a narrative structured around time windows, enabling large models to perform content understanding and analysis based on time windows, improving the accuracy of content understanding and reducing verification bias.

[0083] In this embodiment of the disclosure, target data is selected from multi-source heterogeneous information to verify the initial prediction results, such as... Figure 4 As shown, it includes the following: S401, based on the attention mechanism of the target large model, selects multiple reference terms from multi-source heterogeneous information to falsify the initial prediction results.

[0084] Falsification refers to finding evidence that can refute or overturn the initial prediction results.

[0085] During implementation, the attention mechanism of the target large model can be used to automatically identify information from multi-source heterogeneous information that may contradict the initial prediction result and refute the prediction result, and then filter it as a reference item.

[0086] For example, in financial market data processing scenarios, multi-source heterogeneous information can be categorized into three types: technical, fundamental, and macroeconomic, each requiring falsification from different dimensions. This includes: Technical divergence screening: Check high-frequency volume and price indicators to identify whether there is a risk of short-term technical collapse, such as severely overbought RSI, MACD momentum exhaustion, or low-volume bull trap. Fundamental Overvaluation Check: Utilize low-frequency structured financial data to assess the current PEG ratio and dividend yield. If valuation metrics have already factored in future earnings expectations, then the positive news has lost its effectiveness. Macroeconomic liquidity suppression assessment: Examine M2 growth and the risk-free interest rate environment. If systemic liquidity tightens, then individual stock-level positive factors cannot support independent market rallies.

[0087] In autonomous driving scenarios, such as obstacle recognition and trajectory prediction, the integrated data from multiple radar sensors along the obstacle's direction can be analyzed to disprove the initial prediction results.

[0088] In industrial settings, candidate options can be selected based on historical data and expert experience to disprove the initial prediction results.

[0089] In summary, in any application scenario, at least one of the three types of data—technical, fundamental, and macroeconomic—can be extracted from historical data based on expert knowledge to falsify the initial prediction results.

[0090] Specifically, in the financial sector: technical analysis provides evidence of market behavior; fundamental analysis provides evidence of a company's intrinsic value; and macroeconomic analysis provides evidence of the overall economic environment.

[0091] In the field of autonomous driving: From a technical perspective, this refers to the dynamic data stream generated in real time by the vehicle's own perception and control system. This includes, for example, real-time calculated target positions / velocities, real-time execution commands from controllers (steering, braking, throttle), and instantaneous vehicle speed and yaw rate on the vehicle's CAN bus; it reflects the system's current "behavior" and "response." It can be used to verify the correctness of the system's instantaneous decisions. For example, a predicted path may be safe, but technical data shows a sudden sensor failure or a significant mismatch between braking commands and obstacle distances.

[0092] Fundamentals refer to the inherent hardware and software configuration and long-term health status of a vehicle. They represent the system's "intrinsic capabilities" and "physical condition." Data used may include, for example, the version of the autonomous driving algorithm, sensor calibration parameters, vehicle dynamics model parameters, historical fault code logs, and wear and life data of key components (such as brake pads and tires). Fundamentals can be used to disprove the assumption that the system has the ability to reliably perform a certain task. For example, assuming the system performs normally, but fundamental data shows severe camera calibration drift or unrepaired hidden fault codes in the braking system.

[0093] The macroscopic aspect refers to the dynamic external physical and regulatory environment in which the vehicle operates. It constitutes the "context" and "constraints" of the system's operation. Data used includes high-precision map data, real-time traffic flow and road conditions (V2X), weather (rain, snow, fog) and lighting conditions, traffic signs, and temporary traffic control information. It can be used to disprove the rationality of planning based on environmental assumptions. For example, a high-speed driving route may be planned, but macroscopic data shows that the route is currently under traffic control or experiencing extreme fog.

[0094] For equipment monitoring scenarios in industrial settings: The technical aspect refers to the real-time physical signals directly collected by sensors during equipment operation. It reflects the equipment's current "operating vital signs." Data used includes real-time readings from sensors such as vibration spectrum, temperature, pressure, current, noise, and lubricating oil particle size. It is used to determine whether the equipment is currently in a healthy state. For example, if the equipment is predicted to be running smoothly, but the technical aspect data shows a sudden exceedance of the dominant vibration frequency amplitude or an abnormal rise in bearing temperature.

[0095] Fundamentals refer to the static attributes, design specifications, and cumulative historical status of equipment. They represent the equipment's "inherent identity" and "health record." Data used includes equipment model, design life, material strength, maintenance records, historical failure case database, and mean time between failures (MTBF) for similar equipment. Fundamentals can be used to disprove expectations that equipment should theoretically be in good condition. For example, a new piece of equipment might be expected to be trouble-free, but fundamentals indicate it has exceeded its overhaul cycle or has batch problems due to design flaws.

[0096] The macroscopic aspect refers to the operating conditions of the production system to which the equipment belongs and the external environment. It constitutes the "load background" and "environmental stress" of the equipment's operation. Data used includes production plans and load rates, changes in raw material characteristics, ambient temperature and humidity, voltage stability of the factory power grid, and the linkage status of upstream and downstream equipment. It is used to disprove conclusions that attribute failures to the equipment itself. For example, the initial judgment may be a pump failure, but macroscopic data shows that the abnormal operating conditions are caused by a sudden increase in the viscosity of the inlet material or a sharp drop in the power grid voltage.

[0097] S402, using a target large model, infers the reference prediction results of each reference item for the target task based on each reference item.

[0098] Reference prediction results refer to the prediction conclusions for the target task derived independently by the target large model based solely on information from a single reference item, aiming to reflect the development trend of the task pointed to by that reference item.

[0099] During implementation, the target large model performs inference operations for each selected falsification reference item and independently outputs the reference prediction result corresponding to a single reference item.

[0100] For example, in a financial market data processing scenario, reasoning is performed separately for each disproving reference item: For technical divergence checks, the target large model calls corresponding high-frequency price and volume indicators, focusing on analyzing whether the RSI indicator is in a severely overbought zone, whether the MACD indicator shows momentum exhaustion signals, and whether the trading volume exhibits characteristics of a volume-induced bull trap. It independently derives a reference prediction result of "there is a short-term technical collapse risk" or "there is no short-term technical collapse risk," used to disprove the initial prediction result. For fundamental overvaluation checks, the large model calls low-frequency structured financial data to accurately assess the current PEG ratio and dividend yield levels, comparing them with industry averages and historical reasonable ranges to determine whether valuation indicators have already over-priced future performance expectations, independently outputting a reference prediction result of "news benefits have expired (Priced-in)" or "news benefits have not expired." For macro liquidity suppression checks, the large model examines M2 growth data and the risk-free interest rate environment to determine whether there is currently a systemic liquidity crunch, independently deriving a reference prediction result of "individual stock benefits cannot support independent market performance" or "individual stock benefits can support independent market performance."

[0101] S403, cross-validate the initial prediction results based on the reference prediction results of each reference item.

[0102] Cross-validation refers to the process of comparing and analyzing the reference prediction results corresponding to multiple reference items one by one with the initial prediction results, determining whether the initial prediction results are consistent with each reference prediction result and whether there are any contradictions, and then evaluating the accuracy and reliability of the initial prediction results.

[0103] During implementation, the initial prediction results are compared and analyzed one by one with the reference prediction results for each reference item, and the consistency between the two is recorded.

[0104] S404, if the reference prediction result of any reference item contradicts the initial prediction result, reduce the confidence score of the initial prediction result to the first score interval.

[0105] The confidence score is a quantitative score that measures the reliability of the initial prediction results by the target large model. The higher the score, the stronger the reliability of the initial prediction results; the lower the score, the weaker the reliability of the initial prediction results. Its value range can be set according to actual needs.

[0106] The first score range is a preset range of scores used to characterize the lower reliability of the prediction results. Its specific value can be flexibly set according to business needs and matches the range of confidence scores. Its core function is to intuitively reflect the reliability of the initial prediction results after verification through scores, and to provide a basis for subsequent decision-making.

[0107] During implementation, if the reference prediction result of any reference item is found to contradict the initial prediction result during cross-validation, the confidence score of the initial prediction result will be reduced from the initial value to the first score interval.

[0108] In this embodiment of the disclosure, when there is a large amount of data content in multi-source heterogeneous data, the target large model's attention mechanism is used to filter falsification reference items. This allows the target large model to flexibly select hard evidence based on its current understanding, improving the flexibility and accuracy of information understanding. Reference prediction results are obtained through independent reasoning based on each reference item, and then cross-validated with the initial prediction results. When contradictions occur, the confidence score is reduced. This allows for reverse verification of the reliability of the initial prediction from multi-source heterogeneous information, timely detection of potential biases and contradictory evidence, and quantification of the credibility of the prediction results, effectively improving the reliability of the target task's output.

[0109] During implementation, the target big model can screen based on the conditions of hard indicators. For example, it can use attention mechanisms to understand multi-source heterogeneous information and combine the knowledge of the target big model itself to evaluate the correlation and reliability of each potential candidate indicator, and select highly correlated and highly reliable candidate indicators as reference items.

[0110] In other embodiments, multiple reference items are selected from multi-source heterogeneous information to verify the initial prediction results, which can also be implemented as follows: Figure 5 As shown, it includes the following: S501, based on the understanding of the target large model, extracts multiple reference items to falsify the prediction results from the pre-built knowledge graph; the knowledge graph is constructed based on the original data of the first type of data and the second type of data.

[0111] Knowledge graphs refer to a structured knowledge structure that integrates and links the original data of the first and second categories of data, forming a network of entities, attributes, and relationships. It can be used to sort out the inherent connections and logical relationships between various types of data.

[0112] During implementation, the knowledge graph can be analyzed through the target large model. Based on the entities, attributes and relationships in the knowledge graph, the logical constraints, contradictory relationships and reverse support information between each node can be identified, thereby extracting node information from the knowledge graph as reference items.

[0113] S502 uses a target large model to infer the reference prediction results of each reference item for the target task based on each reference item.

[0114] Reference prediction results refer to the prediction conclusions for the target task derived by the target large model through independent analysis and calculation of the relevant relationships in the knowledge graph for each reference item.

[0115] It is understandable that the reference items are divided into categories, and each reference item includes at least one parameter belonging to the same category.

[0116] S503, cross-validate the initial prediction results based on the reference prediction results of each reference item.

[0117] During implementation, the initial prediction results are compared and analyzed one by one with the reference prediction results of each reference item, and the consistency between the two is recorded in conjunction with the relationships in the knowledge graph.

[0118] S504, if the reference prediction result of any reference item contradicts the initial prediction result, reduce the confidence score of the initial prediction result to the first score interval.

[0119] During implementation, if the reference prediction result of any reference item is found to contradict the initial prediction result during cross-validation, and the contradiction is consistent with the knowledge graph logic, the confidence score of the initial prediction result will be reduced from the initial value to the first score range.

[0120] In this embodiment, the target large model extracts falsification reference items from the knowledge graph constructed based on the first and second types of original data. Then, it independently infers reference prediction results based on each reference item and cross-validates them with the initial prediction results. When any reference prediction result contradicts the initial prediction result, its confidence score is reduced to the first score range. This approach breaks through the cognitive limitations of single forward reasoning from a reverse falsification perspective, effectively suppresses confirmation bias and local information dominance issues, and enhances the accuracy of the initial prediction results through independent verification based on multiple reference items, further ensuring the reliability and stability of the final prediction results.

[0121] In this embodiment of the disclosure, a knowledge graph is constructed, such as... Figure 6 As shown, it may include the following: S601, obtain the original data of the first type of data and the second type of data in the current analysis period to obtain the data to be processed.

[0122] The current analysis cycle refers to the pre-defined time range used to build the knowledge graph. It can be defined independently from the time window mentioned above and can be flexibly set according to the target task requirements (such as 1 week, 1 month, 1 quarter). The core purpose is to ensure that the data to be processed is timely and relevant, and can support the construction and subsequent verification of the knowledge graph.

[0123] The first type of data refers to unstructured raw data before abstract extraction, which is then preprocessed for use in constructing knowledge graphs.

[0124] For example, in financial market data processing scenarios, the raw data for the first type of data can be company announcements, financial news, research report summaries, market sentiment text, etc. In autonomous driving scenarios, the raw data for the first type of data can be camera images, obstacle event records, driving behavior logs, etc. In industrial scenarios, the raw data for the first type of data can be audio data, fault logs, maintenance records, anomaly description text, etc.

[0125] The second type of raw data refers to the initial indicator data before statistical calculation, that is, the raw quantitative data collected directly from the data source without being summarized.

[0126] During implementation, the original data of the first and second categories of data obtained in the current analysis period can be summarized, sorted, and invalid and duplicate data can be removed to obtain the dataset to be processed.

[0127] S602, based on the preset graph pattern, extract information from the data to be processed to obtain an initial graph containing the initial node and the attribute information of the initial node.

[0128] The preset graph pattern refers to the predefined knowledge graph structure specification, including preset rules for node types, attribute types, and relationship types between nodes. Its core function is to standardize the direction and content of information extraction, ensure that the initial node and attribute information extracted meets the construction requirements of the knowledge graph, and avoid the extraction of disorganized content.

[0129] Information extraction refers to the process of extracting initial nodes and node attribute information that meet the requirements from the data to be processed based on a preset graph pattern. In practice, information extraction methods can be adopted according to different types of data to be processed.

[0130] An initial node is a basic unit extracted from the data to be processed that constitutes the knowledge graph. Each initial node corresponds to a type of core information in the data. The attribute information of a node refers to the specific information used to describe the characteristics and state of the initial node.

[0131] For example, in a financial market data processing scenario, preset node types may include stock entities, indicator nodes, event nodes, and data slice nodes. Stock entity nodes are entities based on specific stock targets; indicator nodes are independent units based on time-dimensional indicators, with attributes including date, value, and calculation window; event nodes are independent units based on important announcements or news, with attributes including release date, event type, and sentiment polarity; data slice nodes are units based on time summary intervals, used to carry the statistical characteristics of raw data within the corresponding time period. Preset relationship types include: indicator-at-date-value, indicator-leading-indicator, indicator-with-indicator-intraday correlation, event-occurring-on-date, event-causing-indicator anomaly, and indicator-belonging-to-statistical slice, respectively used to describe the temporal correlation between stock and indicator values, the lead-lag relationship between indicators, the intraday correlation of indicators, the timing of event occurrence, the abnormal impact of events on indicators, and the attribution relationship between daily indicators and time slices.

[0132] In autonomous driving scenarios, preset node types include vehicle entity, perception node, operational indicator node, event node, and data slice node. Vehicle entity nodes are defined by specific vehicles; perception nodes use single-time-dimension perception data as independent units, with attributes including acquisition time, perception type, perception content, and perception accuracy; operational indicator nodes use single-time-dimension operational parameters as independent units, with attributes including acquisition time, indicator name, value, and safety threshold; event nodes use driving-related events as independent units, with attributes including occurrence time, event type, and event level; and data slice nodes use time-summarized intervals as units, used to carry driving statistical characteristics for corresponding time periods. Preset relationship types include perception-at-time-acquisition, operational indicator-at-time-value, perception-and-operation indicator-related, event-at-time-occurrence, event-causing-operational anomaly, and operational indicator-belonging to-statistical slice, respectively used to describe the association between vehicle and perception data acquisition, the temporal association between vehicle and operational indicators, the correlation between perception data and operational indicators, the occurrence sequence of driving events, the abnormal impact of events on vehicle operation, and the attribution relationship between real-time indicators and time slices.

[0133] In industrial scenarios, preset node types include equipment entity nodes, operating parameter nodes, fault event nodes, and data slice nodes. Equipment entity nodes are specific devices; operating parameter nodes are single-time-dimensional operating parameters as independent units, with attributes including acquisition time, parameter name, value, and normal range; fault event nodes are equipment fault-related events as independent units, with attributes including occurrence time, fault type, fault level, and alarm information; and data slice nodes are time-summarized intervals used to carry the statistical characteristics of equipment operation during the corresponding time period. Preset relationship types include operating parameter - at - time - value, operating parameter - precedes - fault event, operating parameter - related to - operating parameter, fault event - occurs - at - time - occurrence, fault event - triggers - parameter anomaly, and operating parameter - belongs to - statistical slice. These are used to describe the temporal association between equipment and operating parameters, the leading relationship between operating parameters and fault events, the correlation between different operating parameters, the occurrence sequence of fault events, the abnormal impact of faults on parameters, and the attribution relationship between real-time parameters and time slices.

[0134] S603 performs leader relationship analysis and event influence relationship analysis on multiple initial nodes to obtain the leader relationship and event influence relationship among multiple initial nodes.

[0135] A leading relationship refers to the sequential association between multiple initial nodes in a time series. That is, an indicator or event corresponding to one node occurs before an indicator or event corresponding to another node, and there is a certain temporal correlation between the two. For example, in financial market data processing scenarios, the leading relationship between indicator nodes can be seen as the "increased trading volume" indicator node occurring before the "price breakout" indicator node; that is, a significant change in trading volume occurs first, subsequently driving the stock price to break through the previous range. In autonomous driving scenarios, the leading relationship between perception nodes and operational indicator nodes can be seen as the "radar detects an obstacle ahead" perception node occurring before the "vehicle speed decreases" operational indicator node; that is, the radar captures obstacle information first, and then the vehicle decelerates. In industrial scenarios, the leading relationship between operational parameter nodes and fault event nodes can be seen as the "continuously increasing equipment vibration frequency" operational parameter node occurring before the "equipment fault alarm" fault event node; that is, abnormal equipment vibration occurs first, and only after accumulating to a certain level will a fault alarm be triggered.

[0136] Event impact relationship refers to the causal relationship between multiple initial nodes, that is, a change in an event or indicator corresponding to one node will directly lead to a change in an event or indicator corresponding to another node.

[0137] S604 updates the initial graph based on the leading relationship and the event influence relationship to obtain the knowledge graph.

[0138] Graph updating refers to adding the leading relationships between nodes and the influence relationships of events obtained in step S603 to the initial graph, improving the structure of the initial graph, and upgrading the initial graph from a basic structure containing only nodes and attributes to a complete knowledge graph containing nodes, attributes, and relationships between nodes.

[0139] In practice, all the leading and causal relationships obtained can be added to the initial graph one by one, and the specific meaning of the relationship can be marked. Then, the updated graph can be verified to remove contradictory and erroneous relationships to ensure the relevance and logic of nodes and relationships, thereby obtaining the knowledge graph.

[0140] In some embodiments, leader relationship analysis can be performed on multiple initial nodes, such as... Figure 6 As shown, it includes: S6031 converts the temporal node information sequence of each of the multiple initial nodes into a stationary dataset.

[0141] A stationary dataset is a dataset whose statistical properties do not change significantly over time after processing. In practice, a stationary dataset can be specifically understood as a dataset in which at least one of the core statistical properties of the initial nodes, such as the mean, variance, and autocovariance, does not change systematically over time.

[0142] For each initial node, multiple pieces of information about that initial node in time sequence (including information name, attributes, etc.) are obtained and constructed into a node information sequence.

[0143] S6032, based on the stationary dataset of the two initial nodes in the pair of indicators to be tested, performs Granger causality test on the two initial nodes in the pair of indicators to be tested, and obtains the test results.

[0144] The Granger causality test is a statistical test method used to determine the primacy relationship between two initial nodes. The test logic is as follows: if the node sequence of one initial node over time can effectively predict the node sequence of the other initial node over time, then the former is considered a Granger cause of the latter, meaning the former occurs before the latter, and a primacy relationship exists between them.

[0145] The pair of indicators to be tested refers to a pre-defined combination of two initial nodes for which it is necessary to determine whether a leading relationship exists.

[0146] In implementation, the stationary datasets of the two nodes in each pair of indicators to be tested can be input into the Granger causality test model, and the test operation can be performed to obtain the test statistic, p-value, optimal lag number, and other test results for each pair of indicators. The test statistic measures the strength of the time-series correlation between the two nodes in the pair of indicators, determining whether the sequence of one node can effectively predict the sequence of the other node. The p-value is the core parameter used to determine the significance of the test results, reflecting the probability of the test statistic occurring. By comparing it with a preset significance level, if the p-value is less than the preset significance level, it indicates that the leading relationship between the two nodes is statistically significant; if the p-value is not less than the preset significance level, it indicates that the leading relationship is not significant. The optimal lag number is the time lag length that makes the test results most reliable, i.e., the optimal time interval at which the sequence of one node affects the sequence of another node.

[0147] S6033, determine whether there is a leading relationship between the two initial nodes in the pair of indicators to be tested based on the test results.

[0148] During implementation, the test statistic, p-value, and optimal lag number of each pair of indicators to be tested can be analyzed one by one to determine and record whether there is a leading relationship, the leading direction, and the leading duration between the two nodes in each pair of indicators.

[0149] In this embodiment of the disclosure, event influence relationship analysis is performed on multiple initial nodes, as follows: Figure 6 As shown, it includes: S6034, based on the timestamp of the event corresponding to each initial node, generate an event window and an estimation window; the event window is a first time range determined based on the reference time in the timestamp; the estimation window is a second time range determined based on the reference time; in terms of timing, the second time range is located before the first time range.

[0150] The baseline time refers to the time with the timestamp of the event corresponding to the initial node as the core reference point, and is used to define the range of the event window and the estimation window.

[0151] An event window is a time frame set around a baseline time to observe the impact of an event on target metrics. It encompasses the period before and after an event occurs, using language to capture changes in metrics before and after the event.

[0152] An estimation window is a time range set around a baseline time and preceding the event window. It is used to capture the normal variation patterns of target indicators before the event occurs, providing a reference benchmark for subsequent comparative analysis of the event's impact.

[0153] During implementation, for each event corresponding to the initial node, its timestamp is extracted as the base time. Then, according to the preset window length, the event window and the estimation window are set respectively to ensure that the estimation window is ahead of the event window in terms of timing.

[0154] S6035 generates an estimation model based on the indicator data of multiple target indicators within the estimation window; the estimation model is used to estimate the target indicators of the target task; the target indicators are nodes in the knowledge graph.

[0155] The target indicator is the indicator that the estimation model needs to predict, and the indicator data is the data related to the target indicator. The indicator data of the target indicator can be determined according to actual needs.

[0156] An estimation model is a predictive model obtained through fitting and regression learning based on multiple indicator data within an estimation window. It is used to estimate the numerical value or trend of the target indicator that needs to be predicted in the target task based on the multiple indicator data within the estimation window. The target indicator refers to the quantitative indicator that ultimately needs to be predicted, analyzed, or verified in the target task.

[0157] S6036, generates a first reference value sequence of the target indicator within the event window based on multiple indicator data of the event window, and generates a second reference value sequence of the target indicator within the event window based on the estimation model.

[0158] The first reference value sequence refers to the actual observation sequence of the target indicator within the event window. That is, the sequence formed by arranging the real change data of the target indicator in chronological order after the event occurs, which is used to reflect the actual change of the target indicator after the event occurs.

[0159] The second reference value sequence refers to the sequence of target index values ​​obtained by inferring and predicting the event window through the estimation model, reflecting the model estimation performance of the target index within the event window.

[0160] S6037, based on the difference between the first reference value sequence and the second reference value sequence, analyze the significance of the event's impact on the target instruction to obtain the significance parameter.

[0161] The degree of difference refers to the numerical deviation, fluctuation range, or trend deviation between the first reference value series and the second reference value series on the same time axis.

[0162] Impact significance refers to the degree to which an event affects a target indicator, that is, whether the actual fluctuation of the target indicator significantly deviates from the normal trend estimated by the estimation model. The higher the significance, the greater the impact of the event on the target indicator; conversely, the smaller the impact, the less significant the impact. The analysis process needs to be carried out in accordance with the logic of calculating test statistics and judging significance. The significance parameter refers to the numerical value after quantifying the impact significance; the higher the value, the more significant the impact of the event on the target indicator.

[0163] During implementation, the significance of the event's impact on the target instruction is analyzed to obtain significance parameters, which can be achieved based on the following steps: Step C1: Calculate the individual difference (AR) between the first reference value sequence and the second reference value sequence at each time point within the event window. Then, sum the individual difference values ​​at all time points to obtain the cumulative difference (CAR) within the event window. The cumulative difference can comprehensively reflect the overall deviation level between the actual value and the model estimate of the target indicator within the event window.

[0164] Step C2: Set up the hypothesis testing logic. This involves determining whether the event has a significant abnormal impact on the target indicator, and setting the following hypothesis: (1) Null hypothesis The cumulative difference within the event window is equal to 0, meaning that the event has no significant abnormal impact on the target indicator, and the actual fluctuation of the target indicator is not substantially different from the normal trend estimated by the model. (2) Alternative Hypothesis If the cumulative difference within the event window is not equal to 0, it means that the event has a significant abnormal impact on the target indicator, and the actual fluctuation of the target indicator deviates significantly from the normal trend estimated by the model.

[0165] Step C3: Calculate the significance parameter. That is, after the hypothesis is set, based on the pre-defined correspondence between the difference and the significance parameter, and combined with the calculation logic of the test statistic (such as the t-statistic), the cumulative difference is converted into a significance parameter (denoted as t_CAR), which can be specifically described by expression (1): (1) Expression (1), Indicates the significance parameter; This indicates the cumulative difference within the event window; The standard deviation represents the cumulative variance and is used to measure the volatility of the cumulative variance estimate.

[0166] This standard deviation can be described by expression (2): (2) In expression (2), L represents the length of the event window, that is, the number of time points contained in the event window; This represents the variance of the individual differences within the estimation window. This variance represents the natural variation level of the target indicator under normal operating conditions without event impact.

[0167] S6038, Determine whether an event has an impact on a target indicator based on the significance parameter.

[0168] In implementation, based on the significance parameter obtained above, and by combining it with a preset significance level and referring to a two-tailed t-distribution table, the critical t-value for the corresponding degrees of freedom can be found. Then, the absolute value of the calculated significance parameter is compared with the found critical t-value to determine whether the event affects the target indicator. If the significance parameter is greater than the critical t-value, the event is determined to have a significant impact on the target indicator, meaning there is a causal relationship between the two. If the significance parameter is not greater than the critical t-value, the event is determined to have no significant impact on the target indicator, meaning there is no causal relationship between the two.

[0169] In this embodiment, an initial knowledge graph is constructed by acquiring multi-source raw data. Then, stabilization processing and Granger causality tests are used to mine the leading relationships between nodes. The causal impact of events is determined by combining event window and estimation model difference analysis. Finally, the temporal leading and causal relationships are integrated into the knowledge graph. This enables the structured extraction of entities, attributes and deep logical connections from the original heterogeneous data, and the construction of a knowledge graph with temporal dependence and causal reasoning capabilities. This provides solid knowledge support and logical basis for the subsequent large model to accurately extract falsification reference items and achieve content understanding and prediction verification that can reduce confirmation bias.

[0170] In this embodiment of the disclosure, if the reference prediction results of all reference items are consistent with the initial prediction results, the confidence score of the initial prediction results is increased to the second score interval.

[0171] The second score interval is a preset score range used to characterize the high reliability of the prediction results. Its specific value can be flexibly set according to business needs, and it must be higher than the first score interval and match the range of confidence scores.

[0172] During implementation, after completing the cross-validation step, it is confirmed one by one whether the reference prediction results of all reference items are consistent with the initial prediction results, and whether the consistency conforms to the association logic in the knowledge graph; if it is confirmed that all reference items are consistent with the initial prediction results without any contradictions or conflicts, the confidence score enhancement operation is initiated.

[0173] In this embodiment, by verifying the consistency between the reference prediction results of all reference items and the initial prediction results, it is possible to fully verify whether the initial prediction conclusion has multi-dimensional and multi-source data support, forming a reliable and quantifiable basis for confidence judgment. Furthermore, by raising the confidence score of the initial prediction results that meet the full consistency condition to the second score range, the confidence assessment can be based on sufficient verification, avoiding the unreliability caused by relying solely on single information or unverified reasoning, thus making the prediction results more reliable.

[0174] In practice, in some scenarios, a neutral conclusion cannot be given by the target large model. Therefore, in this embodiment, the total score interval constructed from the first score interval and the second score interval is discontinuous. That is, the first score interval is biased towards the first result, and the second score interval is biased towards the second result.

[0175] Taking short-term target stock price fluctuation prediction in a financial market data processing scenario as an example, the multi-source heterogeneous information processing method based on a large model provided in this disclosure can be effectively applied to this prediction task. For example... Figure 7 As shown, it can be implemented as follows: S701, Data Acquisition and Input.

[0176] During implementation, one can first acquire massive amounts of multi-source heterogeneous data on the target stock daily. This includes unstructured data, such as news and information related to the target stock, analyst research reports, and market sentiment commentary; as well as multi-dimensional structured data, such as technical data like RSI, MACD, Bollinger Bands, and volume ratio; fundamental data like PEG ratio and dividend yield; and macroeconomic data like M2 growth rate and government bond yield.

[0177] S702, Data Cleaning and Alignment.

[0178] In implementation, the acquired unstructured data can first be processed by extracting prompt words from preset summaries. Then, the target large model can be invoked to perform progressive semantic refinement through a rolling time window to obtain a summary sequence. If the word length of the summary sequence exceeds the word threshold, word-aware adaptive compression is automatically triggered to extract the core semantics, compress the context, and obtain the corresponding first-class data.

[0179] By setting a unified time window and aligning the time, a correspondence is formed between the first type of data and the second type of data corresponding to the structured data within the same time window, so as to obtain multi-source heterogeneous information corresponding to multiple time windows.

[0180] S703, intuitive deduction.

[0181] That is, the target language model receives the multi-source heterogeneous information output in step S703 as context input, and then guides the target big model to comprehensively analyze the current market environment and the correlation of multi-source data through preset prediction prompts, and infers and generates the initial prediction result of the short-term trend of the target stock, such as short-term bullish or short-term bearish.

[0182] S704, Counterfactual falsification.

[0183] During implementation, multiple reference items are selected from the multi-source heterogeneous information to verify the initial prediction results. For example, the multi-source heterogeneous information can be divided into three categories: technical, fundamental, and macroeconomic, and each category is falsified along different dimensions. Technical divergence screening: Check high-frequency volume and price indicators to identify whether there is a risk of short-term technical collapse, such as severely overbought RSI, MACD momentum exhaustion, or low-volume deception.

[0184] Fundamental Overvaluation Check: Utilize low-frequency structured financial data to assess the current PEG ratio and dividend yield. If valuation metrics have already factored in future earnings expectations, then the positive news has lost its effectiveness.

[0185] Macroeconomic liquidity suppression assessment: Examine M2 growth and the risk-free interest rate environment. If systemic liquidity tightens, then individual stock-level positive factors cannot support independent market rallies.

[0186] S705, two-way confidence check.

[0187] If the reference prediction result of any reference item contradicts the initial prediction result, the confidence score of the initial prediction result is reduced to the first score range. That is, if the target model extracts any substantial bearish hard evidence from the fundamentals, technicals, or macroeconomics, a veto mechanism is triggered, which means it tends to favor the bullish expectation in the initial prediction result and outputs a low confidence score (e.g., 1-4 points) for the bearish direction, in order to strictly control the maximum drawdown of the portfolio.

[0188] If the reference predictions for all reference items are consistent with the initial prediction, the confidence score of the initial prediction is increased to the second score range. That is, if the target model has not extracted any substantial bearish hard evidence from the fundamental, technical, or macroeconomic aspects, i.e., no substantial falsification evidence is found, the trading signal logic is deemed extremely robust. At this point, the target model will overcome its defensive bias and assign the prediction a very high bullish confidence score (e.g., 6-10 points).

[0189] S706 outputs the final decision.

[0190] The system generates a final risk-adjusted confidence score (e.g., 1-4 points for bearish, 6-10 points for bullish) and outputs a retrospective analysis report containing complete self-falsification logic to directly guide the execution of quantitative trading.

[0191] Therefore, the multi-source heterogeneous information processing method based on a large model provided in this disclosure can effectively overcome the bias traps of traditional models under strong emotional news stimulation and significantly reduce the maximum drawdown risk of quantitative investment portfolios when performing short-term target stock prediction tasks in financial market data processing scenarios.

[0192] Based on the same technical concept, this disclosure also provides a multi-source heterogeneous information processing device 800 based on a large model, such as... Figure 8 As shown, it includes: Extraction module 801 is used to extract summaries from unstructured data from time series data sources from multiple data sources to obtain the first type of data; the unstructured data includes at least one of text, image and audio. Alignment module 802 is used to align the first type of data and the structured second type of data in time to obtain multi-source heterogeneous information corresponding to multiple time windows; the second type of data is indicator data expressed using digital quantization. The prediction and validation module 803 is used to input prompt words containing multi-source heterogeneous information into the target large model so that the target large model can perform the following operations: Based on multi-source heterogeneous information, the initial prediction results for the target task are generated; Multiple reference items are selected from multi-source heterogeneous information to verify the initial prediction results.

[0193] In some embodiments, the extraction module includes: The preprocessing unit is used to perform preprocessing operations on unstructured data to obtain preprocessed data; the preprocessing operations include at least converting non-text data into text data. The extraction unit is used to extract a summary from the preprocessed data of the current time window in the time series. The splicing unit is used to splice the summary of the current time window with the summaries of multiple time windows preceding the current time window in chronological order to obtain a summary sequence; The compression unit is used to compress each summary in the summary sequence separately using a summary extraction model when the word length of the summary sequence exceeds the word threshold, so as to obtain the first type of data.

[0194] In some embodiments, the compression unit is used for: Based on the summary extraction model, the importance of each summary is determined according to the target task; The compression rate of each summary is determined based on its importance and the term threshold. Based on the compression ratio of each summary, a large-scale summary extraction model is used to perform information compression operations on each summary separately to obtain the first type of data.

[0195] In some embodiments, an acquisition module is also included, configured to obtain the second type of data based on the following method: Based on the indicator categories of the target task, the initial indicator data of the current time window are statistically calculated to obtain the second type of data.

[0196] In some embodiments, the alignment module includes: The alignment unit is used to align the first type of data and the second type of data according to the time window to which the timestamp of each piece of information belongs, so as to obtain the initial data of each time window.

[0197] The transformation unit is used to convert the second type of data in the initial data of each time window into the corresponding text representation, so as to obtain multi-source heterogeneous information corresponding to multiple time windows.

[0198] In some embodiments, the prediction verification module includes: The filtering unit is used to filter out multiple reference items from multi-source heterogeneous information to falsify the initial prediction results based on the attention mechanism of the target large model. The first inference unit is used to infer the reference prediction results of each reference item for the target task based on each reference item using the target large model; The first verification unit is used to perform cross-verification of the initial prediction results based on the reference prediction results of each reference item. The first score reduction unit is used to reduce the confidence score of the initial prediction result to the first score interval when the reference prediction result of any reference item contradicts the initial prediction result.

[0199] In some embodiments, the prediction verification module includes: The extraction unit is used to extract multiple reference items to falsify the initial prediction results from the pre-built knowledge graph based on the target large model understanding; the knowledge graph is constructed based on the original data of the first type of data and the second type of data; The second inference unit is used to infer the reference prediction results of each reference item for the target task based on each reference item using the target large model; The second verification unit is used to perform cross-verification of the initial prediction results based on the reference prediction results of each reference item. The second score reduction unit is used to reduce the confidence score of the initial prediction result to the first score interval when the reference prediction result of any reference item contradicts the initial prediction result.

[0200] In some embodiments, a construction module is also included for constructing a knowledge graph based on the following methods: Obtain the original data of the first and second categories of data within the current analysis period to obtain the data to be processed; Information is extracted from the data to be processed based on the preset graph pattern to obtain an initial graph containing the initial node and its attribute information. Lead relationship analysis and event influence relationship analysis are performed on multiple initial nodes to obtain the lead relationship and event influence relationship among multiple initial nodes; The initial graph is updated based on the leading relationship and the event influence relationship to obtain the knowledge graph.

[0201] In some embodiments, a score enhancement unit is also included, for: If the reference predictions for all reference items are consistent with the initial predictions, increase the confidence score of the initial predictions to the second score range.

[0202] In some embodiments, the prediction verification module includes: The understanding unit is used to perform content understanding on multi-source heterogeneous information and obtain content understanding results; The prediction unit is used to generate initial prediction results for the target task based on the content understanding results.

[0203] The specific functions and examples of each module and submodule of the apparatus in this disclosure can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0204] The acquisition, storage, and application of any type of information, such as user personal information, involved in the technical solutions disclosed herein comply with relevant laws and regulations and do not violate public order and good morals.

[0205] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0206] Figure 9A schematic block diagram of an example electronic device 900 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0207] like Figure 9 As shown, device 900 includes a computing unit 901, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 902 or a computer program loaded from storage unit 908 into random access memory (RAM) 903. RAM 903 may also store various programs and data required for the operation of device 900. The computing unit 901, ROM 902, and RAM 903 are interconnected via bus 904. Input / output (I / O) interface 905 is also connected to bus 904.

[0208] Multiple components in device 900 are connected to I / O interface 905, including: input unit 906, such as keyboard, mouse, etc.; output unit 907, such as various types of monitors, speakers, etc.; storage unit 908, such as disk, optical disk, etc.; and communication unit 909, such as network card, modem, wireless transceiver, etc. Communication unit 909 allows device 900 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0209] The computing unit 901 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 901 performs the various methods and processes described above, such as the large-model-based multi-source heterogeneous information processing method. For example, in some embodiments, the large-model-based multi-source heterogeneous information processing method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on device 900 via ROM 902 and / or communication unit 909. When the computer program is loaded into RAM 903 and executed by the computing unit 901, one or more steps of the large-model-based multi-source heterogeneous information processing method described above can be performed. Alternatively, in other embodiments, computing unit 901 may be configured by any other suitable means (e.g., by means of firmware) to perform a multi-source heterogeneous information processing method based on a large model.

[0210] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0211] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0212] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0213] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0214] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0215] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0216] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0217] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for processing multi-source heterogeneous information based on a large model, comprising: Summarizing unstructured data from time series data sources from multiple data sources yields a first type of data; the unstructured data includes at least one of text, images, and audio. By aligning the first type of data and the structured second type of data in time, we obtain multi-source heterogeneous information corresponding to multiple time windows; the second type of data is indicator data expressed using numerical quantification. Input prompt words containing the multi-source heterogeneous information into the target large model, so that the target large model performs the following operations: Based on the multi-source heterogeneous information, an initial prediction result for the target task is generated; Multiple reference items are selected from the multi-source heterogeneous information to verify the initial prediction results.

2. The method according to claim 1, wherein, The process of extracting summaries from unstructured time-series data from multiple data sources yields a first type of data, including: The unstructured data is preprocessed to obtain preprocessed data; the preprocessing operation includes at least converting non-text data into text data. From the time series, a summary is extracted from the preprocessed data of the current time window; The summary of the current time window and the summaries of multiple time windows preceding the current time window are concatenated in chronological order to obtain a summary sequence; If the word length of the summary sequence exceeds the word threshold, a large summary extraction model is used to compress each summary in the summary sequence to obtain the first type of data.

3. The method according to claim 2, wherein, The method employs a large-scale abstract extraction model to compress each abstract in the abstract sequence to obtain the first type of data, including: Based on the aforementioned summary extraction model, the importance of each summary is determined according to the target task. The compression rate of each summary is determined based on the importance of each summary and the term threshold. Based on the compression ratio of each summary, the summary extraction model is used to perform information compression operations on each summary to obtain the first type of data.

4. The method according to any one of claims 1-3, further comprising obtaining the second type of data based on the following method: According to the indicator categories of the target task, the initial indicator data of the current time window are statistically calculated to obtain the second type of data.

5. The method according to claim 1, wherein, The step of aligning the first type of data and the structured second type of data in time to obtain multi-source heterogeneous information corresponding to multiple time windows includes: Based on the time window to which the timestamp of each piece of information belongs, the first type of data and the second type of data are time-aligned to obtain the initial data for each time window. The second type of data in the initial data of each time window is converted into the corresponding text description to obtain the multi-source heterogeneous information corresponding to the multiple time windows.

6. The method according to claim 1, wherein, The step of filtering target data from the multi-source heterogeneous information to verify the initial prediction result includes: Based on the attention mechanism of the target large model, multiple reference items are selected from the multi-source heterogeneous information to falsify the initial prediction result; The target large model is used to infer the reference prediction results of each reference term for the target task based on each reference term; The initial prediction results are cross-validated based on the reference prediction results of each reference item; If the reference prediction result of any reference item contradicts the initial prediction result, the confidence score of the initial prediction result is reduced to a first score interval.

7. The method according to claim 1, wherein, The step of filtering multiple reference items from the multi-source heterogeneous information and verifying the initial prediction result includes: Based on the target large model, multiple reference items are extracted from the pre-constructed knowledge graph to falsify the initial prediction result; the knowledge graph is constructed based on the original data of the first type of data and the second type of data. The target large model is used to infer the reference prediction results of each reference term for the target task based on each reference term; The initial prediction results are cross-validated based on the reference prediction results of each reference item; If the reference prediction result of any reference item contradicts the initial prediction result, the confidence score of the initial prediction result is reduced to a first score interval.

8. The method of claim 7, further comprising constructing the knowledge graph based on the following method: Obtain the original data of the first type of data and the second type of data within the current analysis period to obtain the data to be processed; Information is extracted from the data to be processed based on a preset graph pattern to obtain an initial graph containing the initial node and the attribute information of the initial node. Lead relationship analysis and event influence relationship analysis are performed on multiple initial nodes to obtain the lead relationship and event influence relationship among the multiple initial nodes; The initial graph is updated based on the leading relationship and the event influence relationship to obtain the knowledge graph.

9. The method according to claim 6 or 7, further comprising: If the reference prediction results for all reference items are consistent with the initial prediction results, the confidence score of the initial prediction results is increased to the second score interval.

10. The method according to claim 1, wherein, The process of generating initial prediction results for the target task based on the multi-source heterogeneous information includes: Content understanding is performed on the multi-source heterogeneous information to obtain content understanding results; The initial prediction result for the target task is generated based on the content understanding result.

11. A multi-source heterogeneous information processing device based on a large model, comprising: The extraction module is used to extract summaries from unstructured time-series data from multiple data sources to obtain the first type of data; The unstructured data includes at least one of text, images, and audio. The alignment module is used to align the first type of data and the structured second type of data in time to obtain multi-source heterogeneous information corresponding to multiple time windows; The second type of data consists of indicator data expressed using numerical quantification; The prediction and verification module is used to input prompt words containing the multi-source heterogeneous information into the target large model, so that the target large model performs the following operations: Based on the multi-source heterogeneous information, an initial prediction result for the target task is generated; Multiple reference items are selected from the multi-source heterogeneous information to verify the initial prediction results.

12. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method of any one of claims 1-10.

13. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-10.

14. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-10.

Images