A method and system for processing multi-language investment research content based on an overseas large model
Patent Information
- Application Number
- CN202610741695.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-27
- Publication Date
- 2026-10-27
AI Technical Summary
[0062]相较于现有技术，本发明采用以上技术方案的有益技术效果如下：本发明为金融投研领域，尤其是海外国际化投研服务，构建了一套多模态、高效率、广适配的智能处理方案。通过时间同步转录技术实现毫秒级文本与音视频时序对齐，解决了传统转写无法精准定位的问题，提升视频内容检索效率；采用重叠式文本分块技术，避免语义信息断裂，保障核心观点传递的完整性；借助多模型可插拔架构，实现不同任务与最优模型的动态适配，提升处理效果与迭代灵活性；拓展多语种翻译支持等，打破语言障碍，适配国际化投研场景；通过时间戳驱动的视频导航系统，构建 “文本 - 时间 - 视频” 三维索引，实现投研信息的快速定位与交互。本发明大幅减少人工干预，降低投研信息处理成本，提升海外投资者的信息获取效率与体验，同时基于境外云基础设施搭建完整服务链路，助力公司国际化战略布局，增强海外投研服务的核心竞争力。
Smart Images
Figure CN122986129A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of financial technology and natural language processing, specifically involving a method and system for intelligent processing of multilingual investment research content based on a large overseas model. Background Technology
[0002] With the deepening of global investment, overseas investment research activities are becoming increasingly frequent, requiring investment research personnel to process a large number of research reports and roadshow videos from different markets and languages. However, existing investment research information processing technologies have the following shortcomings:
[0003] (1) Data format heterogeneity and low processing efficiency: Research report text and roadshow video are two different modalities of data. Traditionally, they need to be processed separately. The audio information in the roadshow video cannot be uniformly retrieved and associated with the text content. Researchers need to spend a lot of time manually browsing the video and reading the research report.
[0004] (2) Semantic segmentation is prone to losing context: When processing long texts into blocks, traditional methods use simple segmentation, which can easily lead to the breakage of semantic information at the block boundaries, affecting the accuracy of subsequent vector retrieval and model inference.
[0005] (3) The retrieval and recall methods are too simple and the accuracy is insufficient: relying solely on vector similarity retrieval may lead to semantic drift, while relying solely on keyword retrieval makes it difficult to capture deep semantic relationships. A single recall strategy is difficult to balance recall and precision.
[0006] (4) The model is rigid and cannot be dynamically switched: Different tasks (such as transcription, translation, summarization, and question answering) have different requirements for model capabilities. Existing technologies usually use fixed models to handle all tasks and cannot dynamically select the optimal model according to the task type, which makes it difficult to balance reasoning efficiency and effectiveness.
[0007] (5) Lack of a closed-loop optimization mechanism: Existing technologies are mostly one-way processing flows, which cannot continuously optimize the model effect based on user feedback, and it is also difficult to trace the source of the output results, making compliance auditing difficult.
[0008] Therefore, there is an urgent need for an intelligent processing method and system for investment research content that can integrate multimodal investment research data, support multilingual processing, possess intelligent retrieval and dynamic model switching capabilities, and form a closed-loop optimization. Summary of the Invention
[0009] This invention addresses the problems existing in the prior art by providing a method and system for intelligent processing of multilingual investment research content based on a large overseas model. It can achieve accurate transcription of roadshow videos, efficient multilingual translation, intelligent Q&A of research reports, and rapid navigation of video content, thus meeting the investment research information acquisition needs of overseas investors with different language backgrounds.
[0010] To address the above technical problems, this invention provides the following technical solution: a multilingual investment research content intelligent processing method based on a large overseas model, comprising the following steps:
[0011] S1. Synchronize publicly and compliantly obtained investment research data to AWS S3 storage. The investment research data includes research reports and roadshow videos.
[0012] S2. Separate audio and video from the roadshow video, use ASR technology to convert speech to text, and simultaneously record timestamps to establish a time-series mapping; use overlapping block technology to cut the research report text and transcribed text, retaining the N character / word overlap areas of adjacent blocks to avoid semantic breaks;
[0013] S3. Vectorize the preprocessed text from step S2 and store it in VectorDB. Combine vector similarity retrieval and keyword retrieval to recall contextual information in a dual-pathway manner.
[0014] S4. Through a multi-model pluggable architecture, the differences in calling different models are shielded by a unified API interface layer, and the optimal model is dynamically switched according to the task type to achieve accurate and efficient model inference.
[0015] S5. Based on the above processing results, output timestamped subtitle transcription, multilingual translation, core viewpoint summary, intelligent Q&A for research reports, and timestamped video navigation services to form a complete closed loop for investment research information processing.
[0016] Furthermore, the aforementioned step S1 includes the following sub-steps:
[0017] S1.1 Data Screening: Select research reports and roadshow video data that can be legally obtained by Xingzhi Overseas Edition clients, and exclude sensitive information such as client data and core company data to ensure data compliance;
[0018] S1.2 Data Transmission: Data originates from the Xingzhi system and is transmitted to AWS S3 storage;
[0019] S1.3 Data Classification and Storage: The research report text and roadshow video are classified and stored separately, and the data source and format information are recorded to provide data support for subsequent preprocessing.
[0020] Furthermore, step 2 described above includes the following sub-steps:
[0021] S2.1 Audio and video separation: The roadshow video is split into audio and video files. Pure audio files are extracted for speech-to-text processing, while video files are retained for subsequent navigation association.
[0022] S2.2 Time-Synchronized Transcription: ASR technology is used to convert audio to text, and the precise timestamp of each text segment / sentence / paragraph is recorded in real time to establish the temporal alignment relationship between text and audio / video, ensuring that the time mapping accuracy reaches the millisecond level;
[0023] S2.3 Overlapping text segmentation: The research report text and video transcription text are segmented into blocks, and an overlapping area design is introduced. N characters / words are retained between adjacent text blocks to avoid semantic information being broken at the block boundaries and to maintain contextual coherence.
[0024] S2.4 Text Vectorization Storage: The segmented text is vectorized and stored in VectorDB. At the same time, the source, timestamp, and segment number of each text segment are recorded to support retrieval and source tracing.
[0025] Furthermore, step 3 described above includes the following sub-steps:
[0026] S3.1 Vector Retrieval Recall: After vectorizing user queries or task requirements, the nearest neighbor search vector similarity retrieval algorithm is used to calculate the similarity with the stored text blocks and recall the content with the highest semantic relevance.
[0027] S3.2 Keyword Retrieval: To address the potential issues of insufficient accuracy or information loss in vector retrieval, a keyword retrieval strategy is added to extract core keywords from the query, match key information in text blocks, and supplement and optimize the retrieval results to improve the comprehensiveness of the retrieval.
[0028] Furthermore, step 4 described above includes the following sub-steps:
[0029] S4.1 Build a multi-model pluggable architecture: Design a unified API interface layer to shield the differences in calling different large language models and support dynamic switching of the optimal model corresponding to different tasks such as transcription, translation, summarization and question answering;
[0030] S4.2 Task Type Identification: Upon receiving user requests or system-triggered tasks, automatically identify the task type and determine the corresponding model selection strategy;
[0031] S4.3 Model Invocation and Inference: Based on the task type, the appropriate model is invoked, and combined with the context information of the retrieval and recall, model inference processing is performed to generate preliminary results.
[0032] Furthermore, the aforementioned step S5 includes the following sub-steps:
[0033] S5.1 Subtitle Transcription Output: Outputs video subtitle text with millisecond-level timestamps to ensure that the subtitles are aligned with the audio and video timing, and supports users to view or edit them directly.
[0034] S5.2 Multilingual Translation Output: Enables multilingual translation between research report texts and video transcription texts, supports accurate multilingual translation, optimizes professional terminology recognition, and improves translation accuracy;
[0035] S5.3 Summary of core viewpoints: Based on the transcribed text or research report content, the core viewpoints and recommended questions are automatically extracted to generate concise summary content, improving the efficiency of investment research information acquisition;
[0036] S5.4 Intelligent Q&A for Research Reports: For research questions raised by users, the system generates professional and accurate answers by combining retrieved research report paragraphs, and supports tracing the source of the answers to the corresponding research reports or video clips.
[0037] S5.5 Video Navigation Service: Based on text search results, the corresponding timestamps are automatically extracted to construct a three-dimensional index of "text-time-video", enabling direct jump from text query to video clip, and supporting users to quickly locate the core content of the video.
[0038] Furthermore, the dual-path recall in step S3 mentioned above also includes result fusion and reordering optimization:
[0039] S3.3 Dual-path result fusion: The result set retrieved by vector retrieval and the result set retrieved by keyword retrieval are merged, and the highest similarity score is used as the final score for the text blocks that are repeatedly retrieved.
[0040] S3.4, Time Proximity Weighting: For recall results derived from video-to-text transcription, a time proximity bonus is added to the original similarity score. That is, when a user query involves a continuous time interval, the score of consecutive text blocks within the query time interval is multiplied by a preset time weighting coefficient. To enhance the continuity of time sequence;
[0041] S3.5 Reordering: The fused result set is reordered using a lightweight cross-encoder, and the Top-K text blocks with the highest semantic matching degree with the user query are output as the final context.
[0042] Furthermore, in step S4 above, the multi-model pluggable architecture also includes a model quality assessment and automatic degradation mechanism, which further includes:
[0043] S4.4 Model Call Monitoring: Real-time recording of response time, inference result integrity, and abnormal status for each model call;
[0044] S4.5 Model Health Score: Based on three indicators—response success rate, average inference latency, and output format standardization—a weighted scoring method is used to calculate the current health of each available model.
[0045] S4.6 Automatic Degradation Switching: When the health score of the currently invoked optimal model falls below the preset threshold for three consecutive times, the system automatically switches subsequent inference tasks of the same type to the backup model with the highest health score and records the switching event in the log for operation and maintenance analysis.
[0046] Furthermore, the aforementioned intelligent processing method for multilingual investment research content based on overseas large-scale models also includes step S6: constructing a quality closed-loop optimization process based on feedback learning.
[0047] S6.1 User Feedback Collection: Embed an explicit feedback control in the output results interface to collect user ratings for three indicators: translation accuracy, summary completeness, and question-and-answer usefulness, as well as the corrected text after user corrections.
[0048] S6.2 Feedback Labeling: The collected feedback is labeled according to the task type. Feedback data with outliers below the preset threshold is marked as "low-quality sample" and enters the manual review queue.
[0049] S6.3 Model Fine-tuning Trigger: When the number of low-quality samples accumulated under a certain task type exceeds the preset sample size threshold, the system automatically triggers the model fine-tuning task, using the accumulated feedback samples to perform LoRA or full parameter fine-tuning on the currently deployed model.
[0050] S6.4 Fine-tuning Model Gray Release: The fine-tuned model is first deployed to the shadow environment and inferred in parallel with the production model in a bypass manner. The consistency between the output of the two models and the historical feedback samples is compared. After passing the consistency assessment, the model is gradually rolled out in a gray-scale manner with a traffic ratio of 5%→20%→50%→100%.
[0051] Furthermore, the aforementioned intelligent processing method for multilingual investment research content based on overseas large-scale models also includes step S7: multimodal data embedding and context enhancement.
[0052] S7.1 Keyframe Extraction from Roadshow Video: Extract keyframe images from the roadshow video at fixed time intervals or using scene transition detection algorithms;
[0053] S7.2 Image Feature Vectorization: The keyframe image is converted into a visual feature vector through a pre-trained visual encoder, stored in an independent vector database ImageVectorDB, and associated with the corresponding timestamp text vector.
[0054] S7.3 Enhanced Cross-Modal Retrieval: When a user's query involves charts, data boards, or presentation content, the system simultaneously performs text vector retrieval and image feature retrieval. The system merges the image description text retrieved by visual recall with the text blocks retrieved by textual recall and inputs them into the large model as multimodal context-assisted reasoning.
[0055] This invention also provides a multilingual investment research content intelligent processing system based on an overseas large-scale model for the aforementioned multilingual investment research content intelligent processing method based on an overseas large-scale model, comprising:
[0056] The data storage module is configured to synchronize publicly and compliantly obtained investment research data to AWS S3 storage, including research reports and roadshow videos.
[0057] The audio and video preprocessing module is configured to: separate audio and video in the roadshow video, convert speech to text using ASR technology, and simultaneously record timestamps to establish a time-series mapping.
[0058] The text segmentation module is configured to use overlapping segmentation technology to cut the research report text and transcribed text, retaining N overlapping characters / words in adjacent segments to avoid semantic breaks; the vectorization and storage module is configured to vectorize the segmented text and store it in VectorDB.
[0059] The dual-path recall module is configured to combine vector similarity retrieval and keyword retrieval, and to recall contextual information in a dual-path manner.
[0060] The multi-model pluggable inference module is configured to: shield the differences in calling different models through a unified API interface layer, dynamically switch the optimal model according to the task type, and achieve accurate and efficient model inference;
[0061] The output and service module is configured to output, based on the processing results of the above modules, timestamped subtitle transcription, multilingual translation, core viewpoint summary, intelligent Q&A for research reports, and timestamped video navigation services, forming a complete closed loop for investment research information processing.
[0062] Compared to existing technologies, the beneficial technical effects of the above technical solutions adopted in this invention are as follows: This invention constructs a multimodal, highly efficient, and widely adaptable intelligent processing solution for the financial investment research field, especially for international investment research services overseas. It achieves millisecond-level text and audio / video temporal alignment through time-synchronized transcription technology, solving the problem of inaccurate positioning in traditional transcription and improving video content retrieval efficiency; it employs overlapping text segmentation technology to avoid semantic information breaks and ensure the integrity of core viewpoints; it leverages a multi-model pluggable architecture to achieve dynamic adaptation between different tasks and the optimal model, improving processing effectiveness and iteration flexibility; it expands multilingual translation support, breaking down language barriers and adapting to international investment research scenarios; and it constructs a "text-time-video" three-dimensional index through a timestamp-driven video navigation system, enabling rapid location and interaction of investment research information. This invention significantly reduces manual intervention, lowers the cost of investment research information processing, improves the efficiency and experience of information acquisition for overseas investors, and simultaneously builds a complete service chain based on overseas cloud infrastructure, supporting the company's international strategic layout and enhancing the core competitiveness of overseas investment research services. Attached Figure Description
[0063] Figure 1 This is a framework diagram of the present invention. Detailed Implementation
[0064] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.
[0065] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.
[0066] Example 1, as Figure 1 As shown in the figure, this embodiment provides a method for intelligent processing of multilingual investment research content based on a large overseas model.
[0067] Step S1: Data Storage: Taking the Xingzhi Overseas Edition system of an overseas securities firm as an example, the system first performs data filtering, selecting research reports (PDF format) and roadshow videos (MP4 format) that clients can legally access, and excluding sensitive files containing client privacy data or core company data. Data filtering is completed through file metadata tags and a whitelist mechanism.
[0068] The filtered data is securely transferred from the Xingzhi system to an AWS S3 bucket via the AWS Transfer Family service. The S3 bucket has two root directories set up according to data type: / reports / for storing research report text and / videos / for storing roadshow videos. Metadata is synchronously recorded for each file during storage, including: data source (Xingzhi system), file format (pdf / mp4), upload timestamp, and original filename.
[0069] Step S2: Audio and video preprocessing and text segmentation: For the roadshow video, the system calls the FFmpeg library to perform audio and video separation, extracts the pure audio file (AAC format), and retains the original video file for subsequent navigation association.
[0070] The audio file is fed into an ASR engine (such as the Whisper large model), which converts speech into text with millisecond precision. The ASR output is in JSON format, and each text fragment contains the following field: {"text": "Today we will discuss investment opportunities in the new energy sector", "start_time": 12345, "end_time": 12890}, where the timestamp is in milliseconds.
[0071] For the research report text, the system first uses a PDF parsing library to extract the plain text content. Then, it uniformly performs overlapping chunking on both the research report text and the aforementioned video transcription text. The chunking strategy is as follows: each chunk is set to a length of 512 tokens, with an overlap of 128 tokens. This means that the i-th chunk shares the content of the last 128 tokens with the (i+1)-th chunk. This design ensures that key semantic information located at chunk boundaries (such as paragraph transition sentences and question-answer pairs) can appear simultaneously in two adjacent chunks, avoiding information breaks.
[0072] After the text is divided into blocks, each block records the following metadata: source type (research report / video), source identifier (research report ID or video ID), timestamp start and end positions (video text only), and block number.
[0073] Step S3: Dual-path recall: Text blocks are converted into 768-dimensional vectors through an embedding model (such as text-embedding-3-small) and stored in a vector database (such as Pinecone or Milvus).
[0074] When a user enters a query, such as "the impact of new energy subsidy policies on the industrial chain", the system performs a dual-path recall:
[0075] Vector retrieval recall: User queries are also vectorized through an embedding model, and the cosine similarity nearest neighbor search algorithm is used in the vector database to recall the top 20 text blocks with the highest similarity.
[0076] Keyword retrieval and recall: Use word segmentation tools to extract core keywords ("new energy", "subsidy policy", "industrial chain") from the query, match text blocks containing these keywords in the Elasticsearch index, and recall the top 20 results.
[0077] Dual-path result fusion: The two result sets are merged. For text blocks that appear in both result sets, the highest similarity score between the two sets is taken as the final score.
[0078] Time proximity weighting: If the user query contains a time range limit (e.g., "third quarter of 2024"), the system extracts the time interval and multiplies the score of the text block in the video transcribed text whose timestamp falls within this interval by the time weight coefficient αt=1.5 to enhance the continuity of time sequence.
[0079] Reordering: The weighted result set is reordered by a lightweight cross encoder (such as Cross-Encoder / ms-marco-MiniLM-L-6-v2) and the top-5 text blocks are output as the context of the large model.
[0080] Step S4: Multi-model pluggable inference: The system adopts a multi-model pluggable architecture, managing multiple large language models through a unified API interface layer, for example:
[0081] Translation task: Prioritize using GPT-4o or Claude-3-Opus;
[0082] Summary of tasks: Prioritize calling GPT-4o-mini or Claude-3-Haiku;
[0083] Question answering task: Prioritize calling GPT-4o or DeepSeek-V3;
[0084] After step S3 is completed, the system automatically identifies the current task type (transcription, translation, summary, Q&A). Assuming the user request is "Please summarize the core points of the above roadshow in English," the system identifies it as a "summary + translation" composite task and successively calls the optimal model for summarizing ability and the optimal model for translating ability.
[0085] During model invocation, the system records the response time, inference result completeness, and abnormal status of each invocation in real time. Based on three metrics—response success rate, average inference latency, and output format standardization—a weighted scoring method is used to calculate the health of each model.
[0086] For example, if a model's health score is below 0.6 (out of 1.0) for three consecutive times, the system will automatically switch subsequent tasks of the same type to the backup model with the highest health score (such as switching from GPT-4o to Claude-3-Opus) to ensure uninterrupted service.
[0087] Step S5: Multimodal Output: Based on the above processing results, the system outputs the following services to the user:
[0088] Subtitle transcription output: The video transcription results are output in SRT subtitle format. Each subtitle includes a sequence number, start timestamp (format: HH:MM:SS,mmm), end timestamp, and subtitle text content to ensure complete synchronization with audio and video.
[0089] Multilingual translation output: Supports mutual translation between Chinese, English, Japanese, Korean and other languages. For professional investment research terms (such as "PEratio" and "DCF valuation"), the system performs accurate matching and translation through the glossary.
[0090] Key takeaways summary: Automatically generates a summary of no more than 200 words, extracts 3-5 key takeaways from the roadshow or research report, and recommends 3 follow-up questions (such as "What is the company's revenue guidance for 2025?").
[0091] Intelligent Q&A for research reports: After the user enters a question, the system generates a professional answer by combining the retrieved paragraphs of research reports, and includes source annotations, such as: "Based on paragraph 2 of page 3 of 'Outlook for the New Energy Industry in 2025'...".
[0092] Video navigation service: When a user finds a segment of the video-to-text by searching for keywords, the system automatically extracts the timestamp corresponding to that segment and generates a jump link. After the user clicks on the link, the video will start playing directly from that timestamp.
[0093] As a preferred option, step S6 is also included: feedback learning loop:
[0094] Below the output, the system embeds three rating controls (1-5 points), corresponding to translation accuracy, summary completeness, and question-and-answer usefulness, respectively. Users can also submit corrected text after error correction.
[0095] When the accumulated low-quality samples (score ≤ 2 points) for a certain task type (such as question answering) exceed 200, the system automatically triggers a model fine-tuning task, using the accumulated feedback samples to perform LoRA fine-tuning on the currently deployed model. The fine-tuned model is first deployed to a shadow environment for parallel inference with the production model in a bypass manner. The consistency of the outputs is compared with 100 test samples. After evaluation, it is gradually rolled out to the production environment in a phased, gradual manner, with a traffic ratio of 5% → 20% → 50% → 100%.
[0096] In a preferred embodiment, step S7 is also included: multimodal enhancement.
[0097] For research reports or roadshow videos containing charts, the system extracts keyframes every 10 seconds, converts the images into 512-dimensional visual feature vectors using the CLIP-ViT model, and stores them in ImageVectorDB. When a user queries "show the trend chart of new energy installed capacity," the system simultaneously performs text retrieval and image feature retrieval. It merges the visually retrieved image description text with the text retrieved text blocks and inputs the result into a multimodal large model (such as GPT-4V) to generate an answer that integrates chart information.
[0098] Example 2
[0099] This embodiment provides a multilingual investment research content intelligent processing system based on an overseas large model, used to implement the method described in Embodiment 1.
[0100] The system includes the following modules: Data storage module: configured to synchronize publicly and compliantly obtained investment research data to AWS S3 storage, including research reports and roadshow videos.
[0101] Audio and video preprocessing module: configured to separate audio and video in roadshow videos, realize speech-to-text conversion through ASR technology, and synchronously record timestamps to establish time-series mapping.
[0102] Text segmentation module: Configured to use overlapping segmentation technology to cut research report text and transcribed text, retaining N overlapping areas of adjacent segments to avoid semantic breaks.
[0103] Vectorization and storage module: Configured to vectorize the segmented text and store it in VectorDB.
[0104] Dual-path recall module: configured to combine vector similarity retrieval and keyword retrieval, with dual-path recall associated with contextual information.
[0105] Multi-model pluggable inference module: Configured to shield the differences in calling different models through a unified API interface layer, dynamically switch the optimal model according to the task type, and achieve accurate and efficient model inference.
[0106] Output and Service Module: Configured to output timestamped subtitle transcription, multilingual translation, core viewpoint summary, intelligent Q&A for research reports, and timestamped video navigation service, forming a complete closed loop for investment research information processing.
[0107] The connection relationship between the modules is as follows: the output end of the data storage module is respectively connected to the input ends of the audio and video preprocessing module and the text chunking module; the output end of the audio and video preprocessing module is connected to the input end of the text chunking module; the output end of the text chunking module is connected to the input end of the vectorization and storage module; the output end of the vectorization and storage module is connected to the input end of the dual-path recall module; the output end of the dual-path recall module is connected to the input end of the multi-model pluggable inference module; the output end of the multi-model pluggable inference module is connected to the input end of the output and service module.
[0108] Embodiment 3
[0109] This embodiment is based on the method described in Embodiment 1, and is applied to the following specific scenarios:
[0110] Scenario 1: Quick review of multilingual roadshow videos
[0111] A Hong Kong-based fund manager needs to review a fully English roadshow of an overseas company, but has limited time. The system automatically completes ASR transcription of the video, generates English subtitles, and translates it into a Chinese abstract through a large model. Through the video navigation service, the fund manager directly jumps to the time period discussing "2025 profit guidance", and completes the comprehension of video content that originally takes 1 hour to watch within 3 minutes.
[0112] Scenario 2: Cross-question answering of multiple research reports
[0113] A researcher needs to compare three research reports on the AI chip industry from different securities firms. After the system vectorizes and stores all three research reports, the user asks "What are the respective forecasts of each firm for the HBM market size in 2026?" The system performs dual-path recall on the relevant paragraphs of the three research reports, generates an answer in the form of a comparison table through a large model, and marks the source page number of each data, which significantly improves the efficiency of research report analysis.
[0114] Scenario 3: Joint analysis of roadshows and research reports
[0115] When an investment manager is watching a roadshow video of a company, the speaker mentions that "the logic of gross margin improvement is explained in detail in the Q3 quarterly report". The user enters "reasons for gross margin improvement" through the platform, and the system not only recalls the speaker's content from the video transcription text, but also recalls relevant paragraphs from the associated quarterly report research report, and integrates them to generate a comprehensive answer, realizing joint analysis of multi-source information.
[0116] Although the present invention has been described above with reference to preferred embodiments, it is not intended to limit the present invention. Those of ordinary skill in the technical field to which the present invention pertains can make various modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the definition of the claims.
Claims
1. A method for intelligent processing of multilingual investment research content based on a large overseas model, characterized in that, Includes the following steps: S1. Synchronize publicly and compliantly obtained investment research data to AWS S3 storage. The investment research data includes research reports and roadshow videos. S2. Separate audio and video from the roadshow video, use ASR technology to convert speech to text, and simultaneously record timestamps to establish a time-series mapping; use overlapping block technology to cut the research report text and transcribed text, retaining the N character / word overlap areas of adjacent blocks to avoid semantic breaks; S3. Vectorize the preprocessed text from step S2 and store it in VectorDB. Combine vector similarity retrieval and keyword retrieval to recall contextual information in a dual-pathway manner. S4. Through a multi-model pluggable architecture, the differences in calling different models are shielded by a unified API interface layer, and the optimal model is dynamically switched according to the task type to achieve accurate and efficient model inference. S5. Based on the above processing results, output timestamped subtitle transcription, multilingual translation, core viewpoint summary, intelligent Q&A for research reports, and timestamped video navigation services to form a complete closed loop for investment research information processing.
2. The intelligent processing method for multilingual investment research content based on an overseas large-scale model according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.1 Data Screening: Select research reports and roadshow video data that can be legally obtained by Xingzhi Overseas Edition clients, and exclude sensitive information such as client data and core company data to ensure data compliance; S1.2 Data Transmission: Data originates from the Xingzhi system and is transmitted to AWS S3 storage; S1.3 Data Classification and Storage: The research report text and roadshow video are classified and stored separately, and the data source and format information are recorded to provide data support for subsequent preprocessing.
3. The intelligent processing method for multilingual investment research content based on an overseas large-scale model according to claim 1, characterized in that, Step 2 includes the following sub-steps: S2.1 Audio and video separation: The roadshow video is split into audio and video files. Pure audio files are extracted for speech-to-text processing, while video files are retained for subsequent navigation association. S2.2 Time-Synchronized Transcription: ASR technology is used to convert audio to text, and the precise timestamp of each text segment / sentence / paragraph is recorded in real time to establish the temporal alignment relationship between text and audio / video, ensuring that the time mapping accuracy reaches the millisecond level; S2.3 Overlapping text segmentation: The research report text and video transcription text are segmented into blocks, and an overlapping area design is introduced. N characters / words are retained between adjacent text blocks to avoid semantic information being broken at the block boundaries and to maintain contextual coherence. S2.4 Text Vectorization Storage: The segmented text is vectorized and stored in VectorDB. At the same time, the source, timestamp, and segment number of each text block are recorded to support retrieval and source tracing.
4. The intelligent processing method for multilingual investment research content based on an overseas large-scale model according to claim 1, characterized in that, Step 3 includes the following sub-steps: S3.1 Vector Retrieval Recall: After vectorizing user queries or task requirements, the Nearest Neighbor Search (NNS) vector similarity retrieval algorithm is used to calculate and store the similarity of text blocks, and recall the content with the highest semantic relevance. S3.2 Keyword Retrieval: To address the potential issues of insufficient accuracy or information loss in vector retrieval, a keyword retrieval strategy is added to extract core keywords from the query, match key information in text blocks, and supplement and optimize the retrieval results to improve the comprehensiveness of the retrieval.
5. The intelligent processing method for multilingual investment research content based on an overseas large-scale model according to claim 1, characterized in that, Step 4 includes the following sub-steps: S4.1 Build a multi-model pluggable architecture: Design a unified API interface layer to shield the differences in calling different large language models and support dynamic switching of the optimal model for different tasks such as transcription, translation, summarization, and question answering; S4.2 Task Type Identification: Upon receiving user requests or system-triggered tasks, automatically identify the task type and determine the corresponding model selection strategy; S4.3 Model Invocation and Inference: Based on the task type, the appropriate model is invoked, and combined with the context information of the retrieval and recall, model inference processing is performed to generate preliminary results.
6. The intelligent processing method for multilingual investment research content based on an overseas large-scale model according to claim 1, characterized in that, Step S5 includes the following sub-steps: S5.1 Subtitle Transcription Output: Outputs video subtitle text with millisecond-level timestamps to ensure that the subtitles are aligned with the audio and video timing, and supports users to view or edit them directly. S5.2 Multilingual Translation Output: Enables multilingual translation between research report texts and video transcription texts, supports accurate multilingual translation, optimizes professional terminology recognition, and improves translation accuracy; S5.3 Summary of core viewpoints: Based on the transcribed text or research report content, the core viewpoints and recommended questions are automatically extracted to generate concise summary content, improving the efficiency of investment research information acquisition; S5.4 Intelligent Q&A for Research Reports: For research questions raised by users, professional and accurate answers are generated by combining the retrieved research report paragraphs, and the answers can be traced back to the corresponding research report or video clip; S5.5 Video Navigation Service: Based on text search results, the corresponding timestamps are automatically extracted to construct a three-dimensional index of "text-time-video", enabling direct jump from text query to video clip, and supporting users to quickly locate the core content of the video.
7. The intelligent processing method for multilingual investment research content based on an overseas large-scale model according to claim 3, characterized in that, The dual-path recall in step S3 also includes result fusion and reordering optimization: S3.3 Dual-path result fusion: The result set retrieved by vector retrieval and the result set retrieved by keyword retrieval are merged, and the highest similarity score is used as the final score for the text blocks that are repeatedly retrieved. S3.4, Time Proximity Weighting: For recall results derived from video-to-text transcription, a time proximity bonus is added to the original similarity score. That is, when a user query involves a continuous time interval, the score of consecutive text blocks within the query time interval is multiplied by a preset time weighting coefficient. To enhance the continuity of time sequence; S3.5 Reordering: The fused result set is reordered using a lightweight cross-encoder, and the Top-K text blocks with the highest semantic matching degree with the user query are output as the final context.
8. The intelligent processing method for multilingual investment research content based on an overseas large-scale model according to claim 4, characterized in that, In step S4, the multi-model pluggable architecture also includes model quality assessment and automatic degradation mechanisms, which further include: S4.4 Model Call Monitoring: Real-time recording of response time, inference result integrity, and abnormal status for each model call; S4.5 Model Health Score: Based on three indicators—response success rate, average inference latency, and output format standardization—a weighted scoring method is used to calculate the current health of each available model. S4.6 Automatic Degradation Switching: When the health score of the currently invoked optimal model falls below the preset threshold for three consecutive times, the system automatically switches subsequent inference tasks of the same type to the backup model with the highest health score and records the switching event in the log for operation and maintenance analysis.
9. The intelligent processing method for multilingual investment research content based on an overseas large-scale model according to claim 1, characterized in that, It also includes step S6: Constructing a quality closed-loop optimization process based on feedback learning. S6.1 User Feedback Collection: Embed an explicit feedback control in the output results interface to collect user ratings for three indicators: translation accuracy, summary completeness, and question-and-answer usefulness, as well as the corrected text after user corrections. S6.2 Feedback Labeling: The collected feedback is labeled according to the task type. Feedback data with outliers below the preset threshold is marked as "low-quality sample" and enters the manual review queue. S6.3 Model Fine-tuning Trigger: When the number of low-quality samples accumulated under a certain task type exceeds the preset sample size threshold, the system automatically triggers the model fine-tuning task, using the accumulated feedback samples to perform LoRA or full parameter fine-tuning on the currently deployed model. S6.4 Fine-tuning Model Gray Release: The fine-tuned model is first deployed to the shadow environment and inferred in parallel with the production model in a bypass manner. The consistency between the output of the two models and the historical feedback samples is compared. After passing the consistency assessment, the model is gradually rolled out in a gray-scale manner with a traffic ratio of 5%→20%→50%→100%.
10. A multilingual investment research content intelligent processing system based on an overseas large-scale model, used in accordance with the multilingual investment research content intelligent processing method based on an overseas large-scale model as described in any one of claims 1-9, characterized in that, include: The data storage module is configured to synchronize publicly and compliantly obtained investment research data to AWS S3 storage, including research reports and roadshow videos. The audio and video preprocessing module is configured to: separate audio and video in the roadshow video, convert speech to text using ASR technology, and simultaneously record timestamps to establish a time-series mapping. The text segmentation module is configured to use overlapping segmentation technology to cut the research report text and transcribed text, retaining N overlapping characters / words in adjacent segments to avoid semantic breaks; the vectorization and storage module is configured to vectorize the segmented text and store it in VectorDB. The dual-path recall module is configured to combine vector similarity retrieval and keyword retrieval, and to recall contextual information in a dual-path manner. The multi-model pluggable inference module is configured to: shield the differences in calling different models through a unified API interface layer, dynamically switch the optimal model according to the task type, and achieve accurate and efficient model inference; The output and service module is configured to output, based on the processing results of the above modules, timestamped subtitle transcription, multilingual translation, core viewpoint summary, intelligent Q&A for research reports, and timestamped video navigation services, forming a complete closed loop for investment research information processing.