Industrial model optimization method based on pre-trained large model

Abstract

The invention relates to the field of financial industry, and discloses a pre-training large model-based industry model optimization method, which comprises the following steps of: performing modeling of time expression habits by utilizing a multi-language pre-training model aiming at financial texts in different regions, and extracting cross-culture generality characteristics of time semantics by combining context alignment and a contrast learning mechanism; after the time expression mode is recognized, a time semantic enhancement module is constructed, and modeling is carried out by aggregating time clues in a context; after the standardized time semantics are obtained, a dynamic feature fusion mechanism is introduced, and the analyzed standardized time information is effectively aligned with key entity information in the financial text; constructing an event dependency relationship graph driven by a time node by utilizing joint embedding representation, and performing logic consistency verification on an inferred event sequence through a graph structure; and according to the training feedback signal, performing fine tuning on the pre-training model in combination with specific data of the financial industry. The method has the advantage of improving the understanding ability of the model.

Description

Technical Field

[0001] The present invention relates to the field of financial industry, and specifically to an industry model optimization method based on a pre-trained large model. Background Art

[0002] In the financial industry, accurate parsing of dates and times is crucial for many critical tasks when processing large amounts of textual information, such as financial reports, news, and investment analysis reports. In recent years, natural language processing techniques based on pre-trained large models have been widely used in the analysis of financial text. These models are effective in extracting valuable information from text. However, existing technologies still face some subtle but specific challenges when processing date and time expressions in financial text, particularly regarding format differences and contextual dependencies. Because financial reports and press releases often incorporate regionally specific formats and time expression conventions, many pre-trained large models are prone to errors or ambiguities when parsing these dates and times. Although pre-trained large models can extract a certain degree of time information, they are primarily trained on general text data and are not specifically optimized for time expression in the financial sector. This results in their inability to accurately parse and understand complex date and time expressions when processing financial industry text. Therefore, it is necessary to design industry model optimization methods based on pre-trained large models to improve model comprehension capabilities. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention provides an industry model optimization method based on a pre-trained large model, which has the advantages of improving the accuracy and practicality of the model in financial text analysis, and solves the problems in the above-mentioned background technology.

[0004] To achieve the above-mentioned purpose of improving the accuracy and practicality of the model in financial text analysis, the present invention provides the following technical solution: an industry model optimization method based on a pre-trained large model, comprising the following steps:

[0005] Using multilingual pre-trained models, we model time expression habits in financial texts from different regions. By combining contextual alignment and contrastive learning mechanisms, we extract cross-cultural common features of time semantics and identify ambiguous time expressions.

[0006] After identifying the temporal expression, a temporal semantic enhancement module is constructed to model the temporal cues in the context and output a standardized temporal semantic representation.

[0007] After obtaining the standardized temporal semantic representation, a dynamic feature fusion mechanism is introduced to effectively align the parsed standardized temporal information with the key entity information in the financial text, thus constructing a joint semantic embedding representation of time and entities.

[0008] Using joint embedding representation to construct a time node-driven event dependency graph, the inferred event sequence is logically verified through the graph structure, temporal reasoning errors are automatically identified and marked, and training feedback signals are generated.

[0009] Based on the training feedback signals, the pre-training model is fine-tuned in combination with specific data from the financial industry, and time extraction and logical prediction mechanisms are introduced to ultimately generate structured industry event information.

[0010] Preferably, the process of identifying ambiguous time expressions is:

[0011] Perform language classification and preprocessing on the input multilingual financial text, unify the time format and grammatical structure, and extract candidate time phrases and context;

[0012] Semantically encode time phrases and contexts based on a multilingual pre-trained model to obtain a vector representation of time meaning.

[0013] Using a cross-language alignment algorithm, semantically similar time phrases in different languages are matched to form semantically corresponding pairs;

[0014] The contrastive learning mechanism is used to enhance the model's ability to discriminate time phrases with similar semantics but different expressions, thereby extracting cross-cultural common features of time semantics.

[0015] Analyze the contextual consistency and semantic ambiguity of each time phrase and identify ambiguous time expressions.

[0016] Preferably, the process of outputting the standardized temporal semantic representation is:

[0017] After identifying the time expression, the context window of the expression in the original text is extracted, focusing on obtaining the modifiers, entities, verbs and event information related to the time expression;

[0018] Based on the information extracted from the context, the structural types of time phrases are uniformly modeled by combining rules and models;

[0019] Map the original time expression to a standard time format and perform semantic normalization to eliminate language differences and expression ambiguity;

[0020] Based on the modeling and normalization results, a multilingual pre-trained model is used to encode the standardized time expression and its context to generate a semantic vector as a standardized time semantic representation.

[0021] Preferably, the process of constructing the joint semantic embedding representation of time and entity is:

[0022] Constructing semantic relationships between time and entities, positionally aligning and contextually pairing time phrases and entity phrases in the text, and leveraging the cross-position attention mechanism in the Transformer architecture to establish contextual interaction channels between time expressions and entity words, capturing the logical or semantic associations in the original context.

[0023] A multi-channel fusion network is introduced to dynamically assign weights to different types of semantic features, and the importance of different features is adaptively controlled through attention mechanisms or gating mechanisms.

[0024] The fused temporal representation and entity representation are concatenated or mapped into the same embedding space to generate a joint semantic representation vector.

[0025] Preferably, the process of generating the training feedback signal is:

[0026] In the constructed event dependency graph, the temporal reasoning path between event nodes is extracted to determine whether the temporal sequence between events is consistent with the semantic logic.

[0027] Set temporal logic rules, perform logical verification on the paths in the graph, and identify reasoning paths that violate temporal sequence, causal relationships, and context;

[0028] Combined with the trained temporal logic prediction model, the inference path is predicted and scored to discover the temporal relationship of the deviation from the prediction result;

[0029] If an illogical time path is found, the path will be marked as a time reasoning error, and the corresponding context, event pair, and error type will be recorded.

[0030] Preferably, the process of finally generating structured industry event information is:

[0031] The temporal reasoning errors marked in the event dependency graph are used as training feedback signals to guide the update direction of the model;

[0032] Based on the pre-trained model, two key subtasks are introduced: the temporal information extraction task, which is used to identify and standardize temporal expressions in text, and the logical relationship prediction task, which is used to determine the causal and temporal relationships between events, forming a joint optimization goal.

[0033] Leverage structured and unstructured data within the industry to train the model to better adapt to the language characteristics and semantic patterns of financial texts;

[0034] After fine-tuning, the model can automatically extract time-event-entity from the input financial text and generate structured industry event information.

[0035] Preferably, structured industry event information includes sorting event information by timeline to form an event chain, marking the occurrence time, main participating entities and roles of each event; extracting causal, conditional, and concurrent logical relationships based on the time sequence and semantic relationships between events, and constructing a structured event map; and classifying and labeling different types of industry events.

[0036] Preferably, the event graph construction process includes calculating the semantic similarity and temporal consistency between event pairs based on joint semantic embedding and standardized time information, and generating a candidate set of event associations; judging whether there is a valid association between candidate event pairs through the trained event dependency prediction model, and assigning event relationship type labels; constructing a multimodal knowledge graph to support graph visualization and query analysis functions.

[0037] Compared with the existing technology, the present invention provides an industry model optimization method based on a pre-trained large model, which has the following beneficial effects:

[0038] The present invention improves the model's ability to understand, reason and structure time information in complex financial texts through multi-layer mechanisms such as multilingual time expression modeling, semantic enhancement, dynamic fusion and graph structure reasoning. It can effectively deal with the differences and ambiguities in time expression under different language and cultural backgrounds, build a unified standardized time semantic representation, and deeply integrate it with financial entity information to generate a joint semantic embedding of time-event-entity. The event dependency graph constructed based on this embedding further realizes the automatic reasoning and consistency verification of event sequence and causal logic, can accurately identify time logic errors and generate training feedback, thereby realizing continuous optimization and iterative updating of the model. It can automatically output structured industry event information, improve the accuracy, timeliness and intelligence level of financial information processing, provide strong support for financial risk warning, event analysis and intelligent decision-making, and has broad practical application value and promotion potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0041] Example 1

[0042] See also Figure 1As shown, the industry model optimization method based on the pre-trained large model according to the embodiment of the present invention includes the following steps:

[0043] S1: Utilize a multilingual pre-trained model to model time expression habits in financial texts from different regions. By combining context alignment and contrastive learning mechanisms, we extract cross-cultural common features of time semantics and identify ambiguous time expressions.

[0044] The process of identifying ambiguous time expressions in S1 is as follows:

[0045] Perform language classification and preprocessing on the input multilingual financial text, unify the time format and grammatical structure, and extract candidate time phrases and context;

[0046] Semantically encode time phrases and their contexts based on a multilingual pre-trained model to obtain a vector representation of temporal meaning.

[0047] A cross-language alignment algorithm is used to match semantically similar time phrases in different languages to form semantic correspondences. Multilingual parallel financial text corpora containing time expressions, such as bilingual Chinese-English data in annual reports, announcements, or press releases, are collected to ensure that the same semantic content is clearly mapped across languages. A time expression recognition model is used to extract all explicit or implicit time phrases from each language text and perform preliminary semantic annotation. A multilingual pre-trained model is used to convert the extracted time phrases into high-dimensional semantic embedding vectors, preserving contextual information to reflect their true meaning in context. Similarity is calculated for time phrase vectors in different languages, and semantic alignment is enhanced through a comparative learning mechanism. Highly similar phrase pairs are screened to construct a cross-language semantic correspondence table for time expressions.

[0048] Use a multilingual pre-trained model to encode time phrases in language A and language B, and use cosine similarity to measure the semantic similarity between the two vectors. The formula is:

[0049]

[0050] Where, is the vector of the i-th time phrase in language A, is the vector of the i-th time phrase in language B, is the similarity between the two phrases;

[0051] The contrastive learning mechanism is used to enhance the model's ability to discriminate time phrases with similar semantics but different expressions, thereby extracting cross-cultural common features of time semantics.

[0052] Analyze the contextual consistency and semantic ambiguity of each time phrase and identify ambiguous time expressions.

[0053] It should be noted that the role of identifying ambiguous time expressions is:

[0054] Function 1: Ambiguous time expressions may refer to different specific points in time in different contexts. If this distinction is not made, the model may misinterpret the time of an event, leading to errors in subsequent tasks such as event sequencing, causal analysis, and trend forecasting. By identifying and processing these ambiguous expressions, the model's accuracy in parsing temporal semantics is effectively improved.

[0055] Role 2: Ambiguous expressions can cause conflicts in the model's constructed timeline. For example, within the same text, if the model mistakenly interprets "next Monday" as two different time points, it can infer an illogical sequence of events. Identifying ambiguous expressions can serve as a key input for logical validation, helping the model correct the timeline and enhance overall reasoning consistency.

[0056] S2: After identifying the temporal expression, a temporal semantic enhancement module is constructed to model the temporal cues in the aggregating context and output a standardized temporal semantic representation.

[0057] The process of constructing the temporal semantic enhancement module in S2 is as follows:

[0058] Extracting contextual vocabulary and syntactic dependency paths associated with time phrases;

[0059] Combined with the attention mechanism, these contextual time clues are weighted and aggregated to capture the temporal logical relationship and semantic orientation hidden in the context;

[0060] The semantic fusion network outputs a structured temporal representation vector, which is called the standardized temporal semantic representation.

[0061] The process of outputting the standardized temporal semantic representation in S2 is:

[0062] After identifying the time expression, the context window of the expression in the original text is extracted, focusing on obtaining modifiers, entities, verbs, and event information related to the time expression to support semantic modeling;

[0063] Based on the information extracted from the context, the structural types of time phrases are uniformly modeled by combining rules and models, taking into account the language environment, word order structure and dependency relationships;

[0064] Map the original time expression to a standard time format and perform semantic normalization to eliminate language differences and expression ambiguity. For example, "second quarter", "second quarter", and "April to June 2023" are unified into 2023Q2;

[0065] Based on the modeling and normalization results, a multilingual pre-trained model is used to encode the standardized time expression and its context, generating a semantic vector containing time type, range, and context information as a standardized time semantic representation.

[0066] It should be noted that the purpose of outputting standardized temporal semantic representation is:

[0067] Function 1: Standardized time semantic representation is language-independent, enabling the model to establish a consistent time understanding mechanism in financial texts in different languages, enhancing the model's cross-cultural adaptability and generalization capabilities.

[0068] Function 2: The standardized time vector can be used as a node to access the event dependency graph, which helps with event sorting, logical reasoning, and causal analysis, improving the model's ability to handle complex event chains.

[0069] The technical solution of this embodiment is to use a multilingual pre-trained language model to identify time expressions in input financial text. Combining a context alignment mechanism with a comparative learning strategy, this model extracts cross-cultural common features of temporal semantics and identifies time phrases with consistent semantics but different expressions. Localized expressions with ambiguous or context-dependent parsing are identified to provide prior information for subsequent modeling. Based on the completion of time expression recognition, a temporal semantics enhancement module is constructed. The output of this module not only retains the original time information but also incorporates the contextual logical sequence, temporal granularity, and potential relative / absolute time conversion information. The output can be used as a key temporal feature vector in event extraction or relationship prediction models for subsequent reasoning. This effectively addresses the problem of large pre-trained models' insufficient understanding of time phrases in financial text, improving the model's accuracy in recognizing and parsing complex time expressions. By aggregating and modeling temporal cues within the context, the model's ability to model logical relationships such as temporal linear order and causal order is enhanced, providing semantic support for event extraction and temporal reasoning tasks. This system converts previously unstructured time phrases into computationally meaningful time vectors, facilitating joint modeling with event nodes and entity information, and enabling the automatic construction of structured knowledge within financial texts. It can adapt to financial data formats in different languages and regions, possesses strong cross-domain scalability, and provides high-quality foundational semantic support for subsequent system model optimization and business logic reasoning.

[0070] Example 2

[0071] like Figure 1 As shown, the industry model optimization method based on the pre-trained large model also includes the following steps:

[0072] S3: After obtaining the standardized temporal semantic representation, a dynamic feature fusion mechanism is introduced to effectively align the parsed standardized time information with the key entity information in the financial text, and construct a joint semantic embedding representation of time and entities.

[0073] The process of constructing the joint semantic embedding representation of time and entity in S3 is as follows:

[0074] The input is pre-processed financial text, which contains two key elements: a standardized temporal semantic representation vector output by the temporal semantic enhancement module; and key entity representations identified by the entity recognition module, such as company names, products, economic indicators, and event verbs.

[0075] Constructing semantic relationships between time and entities, positionally aligning and contextually pairing time phrases and entity phrases in the text, and leveraging the cross-position attention mechanism in the Transformer architecture to establish contextual interaction channels between time expressions and entity words, capturing the logical or semantic associations in the original context.

[0076] A multi-channel fusion network is introduced to dynamically assign weights to different types of semantic features, such as semantic embedding, position encoding, and temporal logical relationships. The importance of different features is adaptively controlled through attention or gating mechanisms, thereby achieving a deep fusion of temporal and entity representations.

[0077] The fused temporal representation and entity representation are connected or mapped into the same embedding space to generate a joint semantic representation vector, which preserves both temporal context clues and entity semantic characteristics, providing an input basis for event relationship modeling.

[0078] The final joint embedding representation can be used as input features for downstream tasks, such as: temporal location of financial events; time-driven causal chain reasoning; construction of temporal event graphs; semantic node embedding in risk warning models, etc.

[0079] By introducing a dynamic feature fusion mechanism, it is possible to adaptively adjust the fusion method between time information and entity information according to the contextual semantic relationship, achieve precise alignment of standardized time semantics and key entities, and then construct a joint semantic embedding representation with stronger contextual association capabilities, effectively improving the model's understanding and modeling capabilities of the temporal relationship and semantic dependency of financial events.

[0080] S4: Use the joint embedding representation to construct a time node-driven event dependency graph, perform logical consistency verification on the inferred event sequence through the graph structure, automatically identify and mark temporal reasoning errors, and generate training feedback signals.

[0081] The process of generating the training feedback signal in S4 is as follows:

[0082] In the constructed event dependency graph, the temporal reasoning path between event nodes is extracted to determine whether the temporal sequence between events is consistent with the semantic logic.

[0083] Set temporal logic rules, perform logical verification on the paths in the graph, and identify reasoning paths that violate rules such as temporal sequence, causality, and context;

[0084] Temporal logic rules are used to constrain the temporal order and semantic relationships between events. Rule types include:

[0085] Sequential relation: If event A occurs before event B, then: A <t B , which is used when the announcement release date should be earlier than the event execution time;

[0086] Causal relationship: If A is a prerequisite for B, such as meeting held → resolution issued, then it must satisfy the occurrence (A) → possibility (B), and t A ≤t B ;

[0087] Contextual relationship: If the contextual semantics states that the event was completed last year, but the system infers that the event time is the current year, then there is a context conflict, which is resolved through syntactic analysis and text timeline consistency check;

[0088] Through the graph structure, we extract the temporal reasoning path between event nodes. Each event node is associated with the time information of its occurrence. We check the order of events by comparing timestamps or relative time series.

[0089] If the end time of event A is earlier than the start time of event B, the time order rule is met.

[0090] If the occurrence of event A requires the occurrence of event B as a condition, it is necessary to ensure that A comes after B.

[0091] Apply the defined temporal logic rules to the temporal relationship between events, and set an event sequence inference rule to determine whether the order of event A and event B is reasonable and whether there is any violation of logic;

[0092] Consider the context of the text to ensure that the timing of events is logically reasonable, for example, the release of the first quarter report should occur after the end of the company's financial year;

[0093] Automatically detect potential time reasoning errors by logically checking the time reasoning paths in the graph. Automated algorithms can be used to check event dependency graphs, including:

[0094] If the order of events does not conform to the expected timeline, mark it as an error;

[0095] If event A is judged to be a prerequisite for B, but B actually occurs after A, it is marked as an error;

[0096] If events semantically shouldn't occur at the same time, flag them as errors.

[0097] Combined with the trained temporal logic prediction model, the inference path is predicted and scored to discover the temporal relationship of the deviation from the prediction result;

[0098] If an illogical time path is found, the path will be marked as a time reasoning error, and the corresponding context, event pair, and error type will be recorded;

[0099] Generating training feedback signals includes:

[0100] The labeled error instances are constructed into training samples, with the input being the joint semantic embedding representation and the time path, and the label being correct / wrong;

[0101] The training signals used for backpropagation can include classification errors, temporal order deviations, etc., which serve as supervisory feedback for fine-tuning the model;

[0102] Design a weighting mechanism to assign higher training weights to error paths that occur frequently or seriously affect understanding.

[0103] These training feedback signals are fed back to the original pre-trained model fine-tuning module to optimize the model's processing capabilities for temporal expression and reasoning, achieving continuous iterative learning and performance improvement.

[0104] By constructing a time-node-driven event dependency graph and leveraging the graph structure to perform logical consistency checks on the order of event reasoning, this approach effectively identifies event sequence anomalies in financial texts caused by misinterpretations of time. By automatically marking these time reasoning errors and generating training feedback signals, this approach provides precise supervision information for subsequent model fine-tuning, significantly improving the accuracy and robustness of the model in temporal semantic reasoning and event sequence judgment in the financial sector.

[0105] S5: Based on the training feedback signal, the pre-trained model is fine-tuned in combination with specific data from the financial industry, and time extraction and logical prediction mechanisms are introduced to ultimately generate structured industry event information.

[0106] The process of finally generating structured industry event information in S5 is as follows:

[0107] The temporal reasoning errors marked in the event dependency graph are used as training feedback signals to guide the update direction of the model;

[0108] Based on the pre-trained model, two key subtasks are introduced: the temporal information extraction task, which is used to identify and standardize temporal expressions in text, and the logical relationship prediction task, which is used to determine the causal and temporal relationships between events, forming a joint optimization goal.

[0109] Leverage structured and unstructured data within the industry to train the model to better adapt to the language characteristics and semantic patterns of financial texts;

[0110] After fine-tuning, the model can automatically extract information elements such as time-event-entity from the input financial text and generate structured industry event information.

[0111] Structured industry event information includes but is not limited to the following:

[0112] Arrange event information in timeline to form an event chain, marking the time of occurrence of each event, the main participating entities and roles;

[0113] Based on the temporal sequence and semantic relationship between events, the logical relationships between events such as causality, conditions, and concurrency are extracted to construct a structured event graph;

[0114] Classify and label different types of industry events, including but not limited to financial announcements, policy releases, market fluctuations, compliance risks, etc., to facilitate subsequent event-driven forecasting and decision-making model calls.

[0115] The process of constructing the event graph includes:

[0116] Based on the aforementioned joint semantic embedding and normalized time information, the semantic similarity and temporal consistency between event pairs are calculated to generate event association candidate sets;

[0117] Through the trained event dependency prediction model, determine whether there is a valid association between candidate event pairs and assign event relationship type labels;

[0118] Build a multimodal knowledge graph containing event nodes, entity nodes, time nodes and their relationship edges, supporting graph visualization and query analysis functions.

[0119] By introducing training feedback signals and fine-tuning the pre-trained model with specific data from the financial industry, the model's ability to understand the complex temporal expressions and logical relationships between events in financial texts has been significantly enhanced. Through the coordinated optimization of time extraction and logical prediction mechanisms, the model can accurately identify event times, key entities, and the causal and temporal relationships between them. Ultimately, it achieves structured extraction and standardized output of financial industry event information, improving the accuracy and timeliness of event mining and providing high-quality data support for downstream tasks such as risk warning and trend analysis.

[0120] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0121] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. Industry model optimization method based on pre-trained large model, characterized by: The following steps are involved: Using multilingual pre-trained models, we model time expression habits in financial texts from different regions. By combining contextual alignment and contrastive learning mechanisms, we extract cross-cultural common features of time semantics and identify ambiguous time expressions. After identifying the temporal expression, a temporal semantic enhancement module is constructed to model the temporal cues in the context and output a standardized temporal semantic representation. After obtaining the temporal semantic representation, a dynamic feature fusion mechanism is introduced to effectively align the parsed standardized time information with the key entity information in the financial text, and construct a joint semantic embedding representation of time and entities. Using joint embedding representation to construct a time node-driven event dependency graph, the inferred event sequence is logically verified through the graph structure, temporal reasoning errors are automatically identified and marked, and training feedback signals are generated. Based on the training feedback signals, the pre-training model is fine-tuned in combination with specific data from the financial industry, and time extraction and logical prediction mechanisms are introduced to ultimately generate structured industry event information.

2. The industry model optimization method based on pre-trained large model according to claim 1 is characterized in that: The process of identifying ambiguous time expressions is: Perform language classification and preprocessing on the input multilingual financial text, unify the time format and grammatical structure, and extract candidate time phrases and context; Semantically encode time phrases and contexts based on a multilingual pre-trained model to obtain a vector representation of time meaning. Using a cross-language alignment algorithm, semantically similar time phrases in different languages are matched to form semantically corresponding pairs; The contrastive learning mechanism is used to enhance the model's ability to discriminate time phrases with similar semantics but different expressions, thereby extracting cross-cultural common features of time semantics. Analyze the contextual consistency and semantic ambiguity of each time phrase and identify ambiguous time expressions.

3. The industry model optimization method based on pre-trained large model according to claim 2 is characterized in that: The process of outputting standardized temporal semantic representation is: After identifying the time expression, the context window of the expression in the original text is extracted, focusing on obtaining the modifiers, entities, verbs and event information related to the time expression; Based on the information extracted from the context, the structural types of time phrases are uniformly modeled by combining rules and models; Map the original time expression to a standard time format and perform semantic normalization to eliminate language differences and expression ambiguity; Based on the modeling and normalization results, a multilingual pre-trained model is used to encode the standardized time expression and its context to generate a semantic vector as a standardized time semantic representation.

4. The industry model optimization method based on pre-trained large model according to claim 3 is characterized in that: The process of constructing the joint semantic embedding representation of time and entity is as follows: Constructing semantic relationships between time and entities, positionally aligning and contextually pairing time phrases and entity phrases in the text, and leveraging the cross-position attention mechanism in the Transformer architecture to establish contextual interaction channels between time expressions and entity words, capturing the logical or semantic associations in the original context. A multi-channel fusion network is introduced to dynamically assign weights to different types of semantic features, and the importance of different features is adaptively controlled through attention mechanisms or gating mechanisms. The fused temporal representation and entity representation are concatenated or mapped into the same embedding space to generate a joint semantic representation vector.

5. The industry model optimization method based on pre-trained large model according to claim 4 is characterized in that: The process of generating training feedback signals is: In the constructed event dependency graph, the temporal reasoning path between event nodes is extracted to determine whether the temporal sequence between events is consistent with the semantic logic. Set temporal logic rules, perform logical verification on the paths in the graph, and identify reasoning paths that violate temporal sequence, causal relationships, and context; Combined with the trained temporal logic prediction model, the inference path is predicted and scored to discover the temporal relationship of the deviation from the prediction result; If an illogical time path is found, the path will be marked as a time reasoning error, and the corresponding context, event pair, and error type will be recorded.

6. The industry model optimization method based on pre-trained large model according to claim 5 is characterized in that: The process of ultimately generating structured industry event information is as follows: The temporal reasoning errors marked in the event dependency graph are used as training feedback signals to guide the update direction of the model; Based on the pre-trained model, two key subtasks are introduced: the temporal information extraction task, which is used to identify and standardize temporal expressions in text, and the logical relationship prediction task, which is used to determine the causal and temporal relationships between events, forming a joint optimization goal. Leverage structured and unstructured data within the industry to train the model to better adapt to the language characteristics and semantic patterns of financial texts; After fine-tuning, the model can automatically extract time-event-entity from the input financial text and generate structured industry event information.

7. The industry model optimization method based on pre-trained large model according to claim 6 is characterized in that: Structured industry event information includes sorting event information by timeline to form an event chain, marking the occurrence time, main participating entities and roles of each event; based on the time sequence and semantic relationship between events, extracting causal, conditional, and concurrent logical relationships, and constructing a structured event map; and classifying and labeling different types of industry events.

8. The industry model optimization method based on pre-trained large model according to claim 7 is characterized in that: The construction process of the event graph includes calculating the semantic similarity and temporal consistency between event pairs based on joint semantic embedding and standardized time information, and generating a candidate set of event associations; judging whether there is a valid association between candidate event pairs through the trained event dependency prediction model, and assigning event relationship type labels; building a multimodal knowledge graph to support graph visualization and query analysis functions.

Images