Method for enabling large model to quickly and accurately answer related time questions

Abstract

The invention relates to the technical field of model language processing, and discloses a method for enabling a large model to quickly and accurately answer related time questions, which comprises the following steps: S1, time information preprocessing: dynamic time anchor point calibration and time context block optimization are included; s2, time logic modeling: a first stage: establishing event-time mapping; stage 2: deducing a trend based on a historical rule; and stage 3, generating a future scene. According to the method, firstly, through dynamic time anchor point calibration and time context block optimization, and through time logic modeling including event-time mapping establishment, trend deduction based on historical laws and future scene generation, a small model of 3B parameter magnitude is selected, and then through time-space two-dimensional retrieval including time axis indexing and space association, time-space two-dimensional retrieval including time axis indexing and space association is carried out. Authoritative current political / scientific news is automatically crawled every week, timeliness knowledge is updated, outdated data is automatically filed, and therefore the method has the advantages of dynamically updating the knowledge base, conducting multi-dimensional retrieval, integrating a lightweight time sequence model and the like.

Description

Technical Field

[0001] This invention relates to the field of model language processing technology, specifically to a method for enabling large models to quickly and accurately answer questions related to time. Background Technology

[0002] With the widespread application of Large Language Models (LLMs) in fields such as intelligent question answering and decision support, their shortcomings in handling time-sensitive problems have become increasingly apparent. Existing technologies mainly suffer from the following bottlenecks: The inherent limitation of time perception: The training corpus of LLM is static, which makes the model unable to dynamically perceive the changes in information brought about by the passage of time; Technical bottlenecks of dynamic updates: Although traditional retrieval augmentation generation (RAG) schemes can supplement external knowledge, they are inefficient in processing time-related data. Delayed knowledge base updates cause the model to reference outdated data, requiring frequent manual calibration. Limitations of complex time series reasoning: Existing methods (such as time series prediction or graph embedding techniques) are mostly focused on a single domain and lack general time series modeling capabilities. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for enabling large models to quickly and accurately answer time-related questions. It has advantages such as dynamically updating the knowledge base, multi-dimensional retrieval, and integrating lightweight time-series models, thus solving the aforementioned problems.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for enabling large models to quickly and accurately answer questions related to time, comprising the following steps: S1. Time information preprocessing: including dynamic time anchor point calibration and time context block optimization; S2. Time-Logic Modeling: This includes Phase 1: Establishing an event-time mapping (e.g., "New Year's Day → January 1st"); Phase 2: Extrapolating trends based on historical patterns (e.g., "10% sales growth last year → projected growth this year?"); Phase 3: Generating future scenarios. S3, Lightweight Time Series Model Integration: Selects small models with 3B parameters (such as Time-R1) and strengthens the temporal causal chain through course learning; S4. Spatiotemporal dual-dimensional search: including time axis index and spatial association; S5. Dynamic knowledge base update: Automatically crawl authoritative political / technological news every week, update timely knowledge, and automatically archive outdated data; S6. Output verification and feedback: contradiction detection mechanism and time-sensitive graded response.

[0005] Preferably, in S1, dynamic time anchor calibration automatically inserts the current timestamp as a benchmark when a question is input, and uses regular expressions or NLP tools to extract relative time words (such as "last Wednesday") from the question and convert them into absolute dates; time context block optimization divides long documents into segments according to time clues (such as by event year), retains a 10% overlap area between segments to prevent truncation of key time nodes, and adopts a sliding window verification mechanism to ensure time continuity.

[0006] Preferably, in S2, the first stage establishes a dedicated time-series knowledge base, associating events (such as "New Year's Day") with time entities (dates, cycles, etc.) and attaching semantic tags (festival type, seasonal attributes). When a user asks "What day of the week was New Year's Day last year?", the current time reference is automatically injected, and backtracking calculations are performed using time analysis tools. The second stage includes a reinforcement learning framework and multimodal trend fusion. The third stage includes causal graph constraint generation and risk perception disambiguation. Causal graph constraint generation constructs an event causal chain, such as "interest rate rise → consumption downgrade → electronic product sales decline," and the constraint generation logic conforms to economic laws. Risk perception disambiguation forces the output of probability intervals and labels key variables when the prediction involves uncertainty.

[0007] Preferably, the reinforcement learning framework includes a slow-thinking reasoning mechanism: guiding the model to decompose temporal logic step by step (e.g., "last year's sales increased by 10%" → extracting growth factors → associating market variables → probability prediction); a two-stage training strategy: SFT stage: fine-tuning the model with historical time-series data (e.g., GDP annual reports) to learn structured time-series representations; RL stage: optimizing time consistency through reward functions; the multimodal trend fusion integrates structured data such as tables and time-series graphs, and analyzes the implicit trends in the graphs through cross-modal alignment technology (e.g., LLaVA architecture).

[0008] Preferably, S3 includes model selection and compression: base: 3B parameter time-series optimized version of Transformer; expert module division of labor: time point analysis, trend fitting and anomaly detection; dynamic retrieval enhancement: for querying unknown time, automatically retrieve perpetual calendar API and inject context; time slot attention mechanism: enhance the model's sensitivity to relative time such as "last year / next month".

[0009] Preferably, in S4, the timeline index adds time metadata to the knowledge base; spatial association binds the geographical location of the event.

[0010] Preferably, the dynamic knowledge base construction mechanism in S5 includes access to authoritative data sources and intelligent segmentation and vectorization; time-aware retrieval optimization includes timestamp embedding and query time alignment; and dynamic update and archiving strategies include incremental updates and automatic archiving mechanisms.

[0011] Compared with existing technologies, this invention provides a method for large models to quickly and accurately answer questions about time, which has the following beneficial effects: 1. This invention first optimizes the time context by dynamic time anchor calibration and time context segmentation, and then uses time logic modeling, including establishing event-time mapping, inferring trends based on historical patterns, and generating future scenarios. It selects a small model with 3B parameters (such as Time-R1), strengthens the time causal chain through course learning, and then uses spatiotemporal dual-dimensional retrieval, including time axis indexing and spatial association, automatically crawls authoritative political / technological news every week, updates timely knowledge, and automatically archives outdated data. Thus, this method has the advantages of dynamically updating the knowledge base, multi-dimensional retrieval, and integrating lightweight time series models.

[0012] 2. Dynamic time anchor calibration in this invention: When a question is input, the current timestamp is automatically inserted as a benchmark. Regular expressions or NLP tools are used to extract relative time words in the question and convert them into absolute dates. Long documents are segmented according to time clues (such as by event year), and a 10% overlap area is retained between segments to prevent truncation of key time nodes. A sliding window verification mechanism is adopted to ensure time continuity.

[0013] 3. The reinforcement learning framework of this invention includes a slow-thinking reasoning mechanism: guiding the model to decompose temporal logic step by step; multimodal trend fusion integrates structured data such as tables and time series graphs, and analyzes the implicit trends in the graphs through cross-modal alignment technology (such as the LLaVA architecture). Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 A method for enabling large models to quickly and accurately answer questions about time includes the following steps: S1. Time information preprocessing: including dynamic time anchor point calibration and time context block optimization; S2. Time-Logic Modeling: This includes Phase 1: Establishing an event-time mapping (e.g., "New Year's Day → January 1st"); Phase 2: Extrapolating trends based on historical patterns (e.g., "10% sales growth last year → projected growth this year?"); Phase 3: Generating future scenarios. S3, Lightweight Time Series Model Integration: Selects small models with 3B parameters (such as Time-R1) and strengthens the temporal causal chain through course learning; S4. Spatiotemporal dual-dimensional search: including time axis index and spatial association; S5. Dynamic knowledge base update: Automatically crawl authoritative political / technological news every week, update timely knowledge, and automatically archive outdated data; S6. Output verification and feedback: contradiction detection mechanism and time-sensitive graded response.

[0017] S1 Dynamic Time Anchor Calibration: When a question is entered, the current timestamp is automatically inserted as a baseline. Regular expressions or NLP tools are used to extract relative time words (such as "last Wednesday") from the question and convert them into absolute dates. Time Context Blocking Optimization: Long documents are segmented according to time clues (such as by event year), with a 10% overlap between segments to prevent truncation of key time nodes. A sliding window validation mechanism is adopted to ensure time continuity.

[0018] Phase 1 of S2 establishes a dedicated time-series knowledge base, associating events (such as "New Year's Day") with time entities (dates, cycles, etc.) and attaching semantic tags (festival type, seasonal attributes). When a user asks "What day of the week was New Year's Day last year?", the current time reference is automatically injected, and backtracking calculations are performed using time analysis tools. Phase 2 includes a reinforcement learning framework and multimodal trend fusion. Phase 3 includes causal graph constraint generation and risk perception disambiguation. Causal graph constraint generation constructs causal chains of events, such as "interest rate rise → consumption downgrade → electronic product sales decline," and the constraint generation logic conforms to economic laws. Risk perception disambiguation forces the output of probability intervals and labels key variables when prediction involves uncertainty.

[0019] The reinforcement learning framework includes a slow-thinking reasoning mechanism: guiding the model to break down time-series logic step by step (e.g., "last year's sales increased by 10%" → extracting growth factors → associating market variables → probability prediction); a two-stage training strategy: SFT stage: fine-tuning the model with historical time-series data (e.g., GDP annual reports) to learn structured time-series representations; RL stage: optimizing time consistency through reward functions; multimodal trend fusion integrates structured data such as tables and time-series graphs, and analyzes the implicit trends in the graphs through cross-modal alignment techniques (e.g., LLaVA architecture).

[0020] S3 includes model selection and compression: Base: 3B parameter time-series optimized version of Transformer; Expert module division of labor: time point analysis, trend fitting and anomaly detection; Dynamic retrieval enhancement: for unknown time queries, automatically retrieve perpetual calendar API and inject context; Time slot attention mechanism: enhance the model's sensitivity to relative times such as "last year / next month".

[0021] In S4, timeline indexing adds time metadata to the knowledge base; spatial association binds the geographical location of events.

[0022] The dynamic knowledge base construction mechanism in S5 includes access to authoritative data sources and intelligent segmentation and vectorization; time-aware retrieval optimization includes timestamp embedding and query time alignment; dynamic update and archiving strategies include incremental updates and automatic archiving mechanisms.

[0023] When using it, the time information preprocessing includes dynamic time anchor calibration and time context block optimization; time logic modeling includes: Phase 1: establishing event-time mapping (e.g., "New Year's Day → January 1st"); Phase 2: inferring trends based on historical patterns (e.g., "10% sales growth last year → estimated growth this year?"); Phase 3: generating future scenarios; lightweight time series model integration: using small models with 3B parameters (e.g., Time-R1), and strengthening the time causal chain through course learning; spatiotemporal dual-dimensional retrieval: including time axis indexing and spatial association; dynamic knowledge base updates: automatically crawling authoritative political / technological news every week, updating timely knowledge, and automatically archiving outdated data; output verification and feedback: contradiction detection mechanism and time sensitivity graded response.

[0024] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for quickly and accurately answering time-related questions by a large model, characterized in that: Comprise the following steps: S1, time information preprocessing: including dynamic time anchor calibration and time context block optimization; S2, time logic modeling: including phase one: establishing event-time mapping; phase two: inferring trends based on historical rules; phase three: generating future scenarios; S3, lightweight time sequence model integration: select small model with 3B parameter order, through course learning to strengthen time causal chain; S4, spatio-temporal two-dimensional retrieval: including time axis index and spatial correlation; S5, dynamic knowledge base update: automatically crawl authoritative political / technology news every week, update time-sensitive knowledge, and automatically archive outdated data; S6, output verification and feedback: contradiction detection mechanism and time sensitivity grading response.

2. The method for large model to quickly and accurately answer time-related questions according to claim 1, characterized in that: The dynamic time anchor calibration in S1: automatically implant the current timestamp as the reference when inputting the problem, use regular expressions or NLP tools to extract relative time words in the question, and convert them into absolute dates; Time context block optimization: segment long documents according to time clues (such as by event year), keep 10% overlap between segments to prevent cutting off key time nodes, use sliding window verification mechanism to ensure time continuity.

3. The method for large model to quickly and accurately answer time-related questions according to claim 1, characterized in that: The first phase of S2 establishes a special time sequence knowledge base, associates events (such as "New Year's Day") with time entities (dates, periods, etc.), and adds semantic labels (holiday type, season attribute). When the user asks "What day was the New Year's Day last year?", the current time reference is automatically injected, and the time resolution tool is used to calculate back; The second phase includes a reinforcement learning framework and multi-modal trend fusion; The third phase includes causal graph constraint generation and risk perception disambiguation, where causal graph constraint generation: build event causal chain: such as "interest rate rise → consumption downgrade → electronic product sales decline", the constraint generation logic conforms to economic laws; Risk perception disambiguation: when the prediction involves uncertainty, force output probability interval and label key variables.

4. The method for large model to quickly and accurately answer time-related questions according to claim 3, characterized in that: The reinforcement learning framework includes a slow thinking reasoning mechanism: guide the model to step-by-step disassemble time sequence logic (such as: "last year's sales growth 10%" → extract growth factor → associate market variables → probability estimation); Two-stage training strategy: SFT phase: fine-tune the model with historical time series data (such as GDP annual report) to learn structured time sequence expression; RL phase: optimize time consistency through reward function; The multi-modal trend fusion integrates structured data such as tables and time series graphs, and resolves implicit trends in charts through cross-modal alignment technology (such as LLaVA architecture).

5. The method for large model to quickly and accurately answer time-related questions according to claim 1, characterized in that: S3 includes model selection and compression: base: 3B parameter time sequence optimized Transformer, expert module division: time point analysis, trend fitting and anomaly detection; Dynamic retrieval enhancement: for unknown time queries, automatically retrieve the calendar API and inject context; Time slot attention mechanism: strengthen the model's sensitivity to relative time such as "last year / next month".

6. The method of claim 1, wherein: In S4, the time axis index: add time metadata to the knowledge base; Spatial correlation: bind event geographic location.

7. The method of claim 1, wherein: The dynamic knowledge base construction mechanism in S5 includes authoritative data source access and intelligent partitioning and vectorization; time-aware retrieval optimization includes timestamp embedding and query time alignment; dynamic update and archiving strategy includes incremental update and automatic archiving mechanism.

Images