Multi-source and multi-modal media intelligence analysis system and methodology.

Abstract

The invention is used in the technical fields of information and communication technologies, artificial intelligence, multimodal content processing, media monitoring, social network analysis, real-time event processing, and enterprise decision support systems. Specifically, the invention relates to a computer-implemented system and method that collects video, image, audio, speech, text, interaction, dissemination, citation, and reference data from social media, live broadcasts, digital press, and echo news sources within a common data model; analyzes this data using open-source artificial intelligence models; establishes semantic relationships between different modalities; identifies trends, anomalies, echoes, replications, and events; and presents the results in the form of real-time alerts, reports, networking, and query interfaces.

Description

1 TARIFF Multi-source and multi-modal media intelligence analysis system and methodology. Technical Area 5 The invention involves information and communication technologies, artificial intelligence, multimodal content processing, and media monitoring. social network analytics, real-time event processing, and enterprise decision support systems techniques It is used in various fields. The invention is particularly derived from social media, live streaming, digital press, and echo news sources. shared video, image, audio, speech, text, interaction, dissemination, citation and reference data It collects data within a data model; analyzes this data using open-source artificial intelligence models; Establishing semantic relationships between different modalities; tendency, anomaly, echo, duplication, and Event identifier; results real-time alert, report, networking and query interface 15 It relates to a system and method implemented by a computer that presents things in a certain format. State of the Art Today's available media monitoring and social listening solutions are mostly single-source, single-20 These are modular, batch processing-oriented, or closed-source structures. The basic techniques of these solutions... The shortcomings are explained below: Resource and Modality Silos Social media texts, live stream audio, video stills, digital press content, and citations / dissemination 25 The data is stored in different databases and in disconnected processes. The same event... Because image, speech, and text formats could not be matched on a common timeline The integrity of the event is being disrupted; variations of the same content on different platforms are presented as separate events. is being processed. Lack of Time Synchronization and Multimode Alignment The moment the image changes during live broadcasts, the moment the speech is transcribed into text, and There are delays and timestamp discrepancies between the moment the text on the screen is read and the actual text. Current systems combine audio segments, video scenes, OCR output, and news text in a common time. Because it does not relate them along the axis, it produces incorrect person-event, person-statement, and image-claim pairings. 35 2 Superficial Keyword Tracking Many solutions only use an exact match keyword or a simple Boolean logic query. It uses synonyms, inflectional suffixes, spelling errors, colloquial language, implicit reference, irony, context, and visuals. Because the logo, screen text, and semantic proximity are not taken into account, both the text is missed and the information is incorrect. The alarm rate is high. 5 Inadequate Echo, Amplification, and Source Origin Analysis Current systems often only provide the number of shares; original source, copy / repost. writing relationships, citation chains, degree of news format transformation, cross-platform jumping, technically 10: propagation rate, influencing nodes, and coordinated replication patterns. It does not remove it. Real-Time Processing and Scalability Issues Video, audio, and text streams generate high volume and speed. Batch-operating architectures handle events. It analyzes after minutes or hours. Fixed allocation of GPU resources, model 15. The sequential execution of calls and the repeated processing of repetitive content cause delays and It increases the cost. Lack of Model Reliability, Explainability, and Human Feedback Model outputs are presented with a confidence score, source of evidence, timestamp, and pieces of data used. When not viewed together, the user cannot understand why a particular result was produced. Incorrect. The model does not involve marking the results and incorporating them into an active learning and retraining cycle. This causes his performance to decline over time. Language and Field Adaptation Deficiency 25 General purpose models include Turkish news language, spoken language, abbreviations, institution / product names, It may not provide sufficient accuracy in regional pronunciations and media jargon. Field adaptation, Custom dictionaries, low-resource labeling, and continuous learning mechanisms are found in most commercial products. It is not under user control. Closed Source and Supplier Dependence Closed model and services; data privacy, cost predictability, offline operation, This creates limitations in terms of monitoring model behavior and internal customization. Different Changes to provider APIs negatively impact system integrity and sustainability. 35 Unified Data Family Tree and Audit Record Deficiency 3 Which version of the raw content passed through which preprocessing steps, which model which parameters the version uses for inference and which report the result is converted into If it cannot be traced, reproducibility, legal oversight, and error analysis become impossible. Dependence on Single Output Channels 5 Most systems only offer a dashboard. Alerts, reports, networking, programmatic search, Retrospective queries and event transmission to other enterprise systems all stem from the same core data model. Since it cannot be produced domestically, the integration cost increases. As a result of technical investigations, application number US9984427B2 was found to be 10 In summary; "Detecting events based on data streams from numerous sources and a system and method for summarizing. The resources in question are social, among other things. This could include media networks, text messages, and news feeds. The system can utilize these sources to potentially... It can obtain raw information containing data on events. The method for event detection is any recorded information. This may include preprocessing and normalizing data input from a source; this also includes 15 In order to validate / verify an event, the extraction of events and entities and eliminating uncertainties, linking events and entities, and presenting different data. matching events and entities obtained from data input from the source It may also include. Subsequently, the validated / verified event is stored in a local data repository and / or a It can be stored on the web server. 20 As can be seen, the invention is based on events derived from data streams from numerous sources. It relates to a system and method for identification and summarization, as well as the above. It does not mention a structure that could provide a solution to the aforementioned disadvantages. In conclusion, due to the negative aspects described above and the current solutions, the subject matter... Due to its shortcomings, it has become necessary to make improvements in the relevant technical field. Purpose of the Invention 30 The invention represents a new breakthrough in this field, unlike the structures used in existing technology. The aim is to create a structure with different technical specifications that bring these elements together. 4 The primary purpose of the invention is to facilitate the use of video, images, sound, speech, text, interaction, attribution, and dissemination. Data includes common identifier, timestamp, source ID, content summary, trust score, and link. The goal is to transform these areas into a single event envelope. Thus, different modalities become parts of the same event. It is processed as such. Another aim of the invention is event-driven data flow, micro-stacking, backpressure, The goal is to process critical content with low latency through prioritization and model guidance. Duplicate content is unified using perceptual hashing, text summarization, and vector similarity; Unnecessary GPU output is reduced. Another purpose of the invention is to integrate speech timestamps, video scene boundaries, OCR boxes, and matching text sentences within specific tolerance windows, thus ensuring that what a person says... the sentence being linked to a person, logo, object, or screen text displayed simultaneously to provide. Another purpose of the invention is to address content origin, redistribution, quotation, copying, summarization, and It is about jointly evaluating rewriting relationships while preserving meaning. Cross-platform The diffusion graph identifies the initial source and influential nodes; the news produced by the institution and the institution itself. The news reports about it are broken down. The system that is the subject of the invention is not only word matching, but also semantic similarity, visual... logo / object presence, specific person-event association, change in emotion or attitude, threshold crossing, and sequential It uses event patterns as trigger conditions. In the invention, each result includes: source link, processed media segment, time interval, model version, 25 The confidence score is stored along with the triggered rule and associated evidence. The user can see the results in the report. It can go back to the raw content on which the finding is based. Since the data, model, and infrastructure components of the invention are open source, there is no requirement for a closed service. It disappears. The system can operate offline or on an isolated network; models are based on institutional data in area 30. It can be adapted. User feedback, model response, and actual result labels in invention data / model versioning. It is kept in the infrastructure. Active learning, retraining, shadow deployment, A / B testing, pilot application. And accuracy is preserved through recovery mechanisms. 35 In this invention, services that use the processor, memory, and GPU scale independently. It is of low complexity. Content is run in small models, obscure examples in large models. Dynamic stacking, With hybrid precision and quantification, the computational cost per unit of content is reduced. The invention includes the same analysis results in real-time alerts, semantic results, trend / anomaly reports, relationship networks, and 5. Echo is offered as a news dissemination and search / query / reporting API; integrated with enterprise workflows. Integration becomes easier. The primary areas of application for the invention are telecommunications operators, public institutions, municipalities, companies, media organizations, broadcasters, brand and reputation management units, crisis 10 management centers, security operations centers, research institutions, and intensive media These are corporate structures that constantly monitor data. The system determines which news items the institution produces itself. the speed at which it was reproduced in the sources and with what formal / semantic changes; the institution the source, actors, subject matter, emotions, attitudes, and distribution of content produced by third parties about it their relationships; events and statements made during live broadcasts; sudden changes in social media interactions It can be used to technically identify changes. The invention also applies to elections, disasters, public health, financial markets, sports, culture and arts, and product launches. customer experience, advertising campaigns, disinformation monitoring, copyright and content monitoring, competition It can be adapted to analysis and early warning applications. The system can be used both in the on-premises data center and in 20 and can be run on a distributed platform in a private cloud, hybrid cloud, or edge computing infrastructure, It is designed as a scalable and open-source software architecture. The source platforms within the scope of the invention are external data sources; the system's data collection, flow processing, storage, AI development, model presentation, user management, observability 25 and reporting components are selected from open-source software. Access to platform APIs is provided. This is done within the framework of the platforms' authorization and usage rules. To achieve the purposes described above, the invention adapts multi-source media content. It is a system for monitoring and multimodal analysis, and its feature is; 30  at least one processor, memory, permanent storage unit, network interface and at least one graphics processor at least one unit that contains the component and enables all elements to perform their functions computer system,  via authorized APIs, feeds, RSS / Atom, or authorized connectors from social platforms Data field; source 35 that transforms content, interaction, and relationship events into a common schema. The social media module, which is a set of adapters, 6  Continuous video / audio reception from satellite, internet, or live streaming protocols, with timestamp generating, creating seamless segments and passing the stream to the real-time processing layer live streaming module,  Gathers news sites, blogs, press releases and similar web content; page body, The digital press module, which extracts the title, author, date, links, and content structure, is number 5.  comparing the news produced by the institution with the news published about the institution; Echo graph by calculating the relationships of replication, citation, reference, origin, and diffusion. The echo news module that creates  source connectors, tails, segmentation, retry, back pressure, The data collection and streaming layer, which manages prioritization and streaming routing, 10  Converts data from different sources into a common event envelope; character encoding, clock slice, language, identity, content mix, rights / access, quality, and data family tree domains standardizing data normalization and preprocessing unit,  codec decoding, resizing, scene change detection, keyframe selection, perceptual hash, OCR region extraction, blur / quality control, and image 15 The video / image preprocessing unit that performs normalization, • Audio channel separation, sample rate conversion, noise reduction, audio efficiency speaker detection, speaker segmentation, audio level normalization, and time slicing. audio preprocessing unit that does it,  Unicode normalization, language detection, sentence splitting, spell check, URL / tag 20 parsing, personal data masking, text cleaning, and fragmentation processes text preprocessing unit that executes  The STT model, which converts spoken speech into timestamped text,  Natural language context for person / object / event / screen text on a frame, scene or short video and the VLM model, which describes it in a structured output format, 25  Detects objects, logos, faces, scenes, text, and motion; handles VLM calls. a computer vision model that acts as a fast, narrowing pre-filter,  LLM that produces event, entity, claim, summary, emotion, attitude, topic, relationship and structured output model,  STT, VLM, CV, and LLM outputs, along with resource metadata, are sorted by time and identity. unifying elements; connecting events / entities, drawing conclusions, similarities, contradictions, themes, emotions, and A multimodal semantic analysis engine that performs trust aggregation operations.  An updatable system containing sources, content, individuals, institutions, topics, events, references, and interactions. Information / diffusion graph is established by considering centrality, community, influence, relational strength, origin, and echo. The insight and relationship analysis engine that generates the scores, 35 7  Sliding time windows, seasonal baseline, change point, z-score, Insulation The forest, with its online clustering and topic shifting methods, exhibits sudden increases, decreases, and identifying unusual patterns and basing measurements on sliding windows A trend and anomaly analysis engine that produces a combined anomaly score by comparison.  word, expression, coherent expression, entity-event combination, semantic proximity, visual object / logo, 5 Key words and events that evaluate emotion / attitude, threshold, and sequence event rules. trigger motor,  Combining repeated alerts linked to the same root event into a single event card The factors that provide this are importance, trust, resource diversity, repetition, user role, and time. Combining according to its window; suppressing recurring alarm, upgrade and delivery 10 Status monitoring notification and alarm engine,  Raw object repository, relational database, lookup index, vector index, and graph. a multi-layer system that keeps track of version, family tree, retention period, and audit logs. data storage / knowledge base, which is a storage structure  Resource management, query, real-time stream, relationship graph, map / timeline, 15 Management, visualization, and operations providing model confidence score and audit log. user interface It includes. The structural and characteristic features and all the advantages of the invention are given in the figures below and these 20 Thanks to the detailed explanation written with references to the figures, it becomes clearer. This will be understood, and therefore the evaluation should also take these figures and detailed explanations into account. It must be done by taking it. Figures to Help Understand the Invention 25 Figure 1 is a schematic representation of the system that is the subject of the invention. The drawings do not necessarily need to be scaled and are necessary for understanding the invention. Details that are not present may have been overlooked. Furthermore, at least to a large extent, 30 Elements that are identical or at least have substantially identical functions are numbered the same. It is shown. Explanation of Part References 35 1. User 8 2. User interface 101. Social Media Module 111. Source X 112. Instagram Feed 113. TikTok Source 5 114. Facebook Source 115. Social Media Video / Image Data 116. Social Media Conversation Data 117. Comment Text 118. Interaction Data 10 102. Live Streaming Module 123. Satellite Broadcasts 124th YouTube Live Stream 125. Live Broadcast Video / Image Data 126. Live Stream Audio Data 15 127. Text / Transcript Data 103. Digital Press Module 128. News Websites 129th Blog Post 130. Title / Text 20 131. Content Data 104. Echo News Module 132. News Reported by the Institution 133. News Reports About the Institution 134. Echo and Replication Data 25 135. Dissemination Data 136. Citation / Reference Data 201. Data Collection and Streaming Layer 202. Data Normalization and Preprocessing Unit 203. Video / Image Preprocessing Unit 30 204. Audio Preprocessing Unit 205. Text Preprocessing Unit 206. STT Model 207. VLM Model 215. Computer Vision Model 35 208. LLM Model 9 209. Multimodal Semantic Analysis Engine 210. Insight and Relationship Analysis Engine 211. Trend and Anomaly Analysis Engine 212. Keyword and Event Triggering Engine 213. Notification and Alarm Engine 5 214. Data Storage / Knowledge Base 301. Instant Alerts and Notifications 302. Results of Semantic Analysis 303. Trend and Anomaly Reports 304. Interaction and Relationship Networks 10 305. Echo News Dissemination Results 306. Search / Query / Reporting API Detailed Description of the Invention In this detailed explanation, the preferred configurations of the invention are not merely for better understanding the subject. in order to facilitate understanding and without imposing any limiting effects It is explained. The invention is particularly based on 20 news sources including social media, live streaming, digital press and echo. shared video, image, audio, speech, text, interaction, dissemination, citation and reference data It collects data within a data model; analyzes this data using open-source artificial intelligence models; Establishing semantic relationships between different modalities; tendency, anomaly, echo, duplication, and Event identifier; results real-time alert, report, networking and query interface. It relates to a system and method implemented by a computer that presents the information in a specific format. 25 The elements and functions used in the system and method that are the subject of the invention are as follows: User (1) can access resources, monitoring topics, key concepts, event rules, reports and on the system. Authorized personnel who identify alarms; review analysis results; provide feedback and labeling. 30 or service account. User interface (2), preferably created with open source web technologies; resource management, query, real-time stream, relationship graph, map / timeline, model confidence score and It is a management, visualization, and operations interface that provides audit logs. 35 Social media module (101), authorized API, feed, RSS / Atom or authorized from social platforms It receives data through connectors; transforms content, interaction, and relationship events into a common schema. It is a set of source adapters. Source X (111), Post, repost, reply, media and interaction meta 5 from X platform It is an external data source from which data can be retrieved. Instagram source (112) states that Instagram content, comment, image / video and interaction metadata It is an external data source from which data can be obtained. TikTok source (113), from which TikTok video, description, comment and interaction metadata can be retrieved. It is an external data source. Facebook source (114) states that Facebook post, comment, media and interaction metadata It is an external data source from which data can be obtained. 15 Social media video / visual data (115), video associated with the post, keyframe, thumbnail, Metadata includes the image and screenshot, along with their resolution, codec, and duration. Social media speech data (116), found in social media video or audio recording; audio 20 It is the speech signal given to the preprocessing and speech-to-text conversion flow. Comment text (117) includes natural language content such as post text, description, comment, reply and tag. Interaction data (118), likes, shares, retweets, comments, replies, followers, views, 25 Event records that include time and actor relationships are used in diffusion and correlation analysis. Live streaming module (102) provides continuous video / audio from satellite, internet or live streaming protocols. the field, generates timestamps, creates seamless segments, and processes the stream in real time. It is the module that transfers to the layer. 30 Satellite broadcasts (123), Linux DVB from DVB-based satellite or broadcast receiver, Live streams obtained using open-source driver and media tools such as FFmpeg / GStreamer. It is the source. 35 11 YouTube live streams (124), live internet broadcasts obtained through authorized access mechanisms It is the source. Live stream video / visual data (125), live stream frames, scene clips and screen components. 5 Live stream audio data (126), speech, music, ambient sound and channel data separated from the live stream. It is an audio stream containing information. Text / transcription data (127), broadcast subtitle, teletext, OCR output or STT result 10 It is a timestamped text that has been produced. Digital press module (103) news site, blog, press release and similar web content It is a module that collects and extracts the page body, title, author, date, links, and content structure. News sites (128), RSS, sitemap, permissioned web access or publisher integration These are digital news sources that are monitored through [platform / channel name]. Blog posts (129), text, images and links published in corporate or individual blogs. It is the source of the content. 20 Headline / text (130), consisting of news headline, lead, body, subheading, description and tags. It is natural language data. Content data (131), author, publication date, URL, chapter, media attachments, source links, content 25 These are hash data and other structured metadata. Echo news module (104) displays news produced by the institution and news published about the institution. Comparing; calculating echo plot by considering reproduction, citation, reference, origin and diffusion relationships. It is the module that creates it. 30 News produced by the institution (132), the institution’s official website, press room, social account or source content published through authorized broadcasting channels. News about the company (133), the company, brand, product, 35 from third party sources This refers to content published by managers about activities or events. 12 Echo and reproduction data (134), full copy, partial copy, summary, rewrite of content, whether it is a semantic derivative or a visual / auditory derivative; its original source and degree of reproduction This is the data that shows. Dissemination data (135) shows the transition of content between platforms and sources over time. This data represents speed, reach, wave magnitude, and cascade structure. Citation / reference data (136), source URL, news name, personal mention, document, post or media. It is data that includes explicit or implicit reference links to the part. 10 Data collection and streaming layer (201), an event streaming platform and streaming processing engine (e.g. Resource linking, queuing, and partitioning on Apache Kafka and Apache Flink. It manages retrying, back pressure, prioritization, and flow routing. Data normalization and preprocessing unit (202), combines data from different sources into a common event envelope. It converts; character encoding, time zone, language, identity, content mix, rights / access, quality and It standardizes data family tree fields. Video / image preprocessing unit (203), codec decoding, resizing, scene change 20 detection, keyframe selection, perceptual hash, OCR region extraction, blur / quality control, and It performs image normalization. Audio preprocessing unit (204), audio channel decomposition, sample rate conversion, noise reduction, voice activity detection, speaker segmentation, volume normalization, and time 25 It slices. Text preprocessing unit (205), Unicode normalization, language detection, sentence splitting, Turkish morphology, spell checking, URL / tag parsing, personal data masking, text cleaning and performs the disassembly operations. 30 STT model (206) with a speech recognition model (e.g. Whisper or open licensed (with its equivalent) converting speech into timestamped text; using Turkish media data. The model supports adaptation, custom dictionary redirection, and trust score generation. 35 13 VLM model (207), frame, scene with a visual-language model (e.g. Qwen2.5-VL-7B-Instruct) or the context of people / objects / events / screen text on short videos using natural language and structured language. It explains it in output format. Computer vision model (215), based on an object detection and OCR library (e.g. 5 (MMDetection, OpenCV, PaddleOCR) objects, logos, faces, scenes, text, and motion. It is a model that detects and also acts as a fast pre-filter that narrows down VLM calls. LLM model (208) uses a large language model (e.g. Qwen2.5-7B-Instruct) to identify events, entities, claims, The system produces a summary, sentiment, attitude, topic, relationship, and structured JSON output, using RAG and field 10. It is a model that supports adaptation. Multimodal semantic analysis engine (209), source metadata with STT, VLM, CV and LLM outputs. It combines data based on time and identity; performs event / entity linking, claim inference, and similarity analysis. It performs operations of combining contradiction, subject, emotion, and trust. 15 Insight and relationship analysis engine (210), source, content, person, institution, subject, event, citation and It establishes an updatable information / diffusion graph from interactions; centrality, community, influence, relationship. It is the unit that generates power, origin, and echo points. Trend and anomaly analysis engine (211), sliding time windows, seasonal baseline, Change point, z-score, Isolation Forest, online clustering, and topic shift. methods that identify sudden increases, decreases, and unusual patterns in sliding windows It is a unit that produces a composite anomaly score by comparing measurements to the baseline. Keyword and event triggering engine (212), word, phrase, regular phrase, entity-event composition, semantic proximity, visual object / logo, emotion / attitude, threshold, and sequence of events rules It is evaluated by a complex event handling component (for example, Apache Flink CEP or Drools). Notification and alarm engine (213), 30 repeated alerts related to the same root event in a single event card The triggers that enable this integration are importance, trust, resource diversity, repetition, user role, and Combining according to time window; suppressing recurring alarm, upgrading and delivering. It is the unit that monitors the situation. 14 Data storage / knowledge base (214), raw object store, relational database, search index, Consisting of a vector index and a graphic layer; version, family tree, retention period, and control. It is a multi-storage structure that keeps track of the records. Instant alerts and notifications (301), interface for specified events, email, messaging, web 5 These are low-latency notifications generated via a trigger or enterprise event bus. The results of the semantic analysis (302) include summary, topic, emotion, attitude, entity, claim, event, similarity, These are results that present contradictions and evidentiary connections in a structured manner. Trend and anomaly reports (303), time series, subject set, source, region, actor and These are reports that include trend, turning point, and anomaly information on a platform-by-platform basis. Interaction and relationship networks (304) are between resources, content, actors, institutions and events. The outputs visualize the relationship graph along with centrality, community, and influence metrics. 15 Echo news dissemination results (305), original source, copy / derivative content, citation chain, time The chart shows the results indicating platform migration, deployment speed, and effective nodes. Search / query / reporting API (306), REST / gRPC / GraphQL or OpenSearch compatible 20 Filtered search, vector similarity, graphic query, report generation, and via the query interface. It provides external system integration. The steps involved in the process carried out with the system that is the subject of the invention are listed below:  Institution, person, topic, event to be monitored by user (1) via user interface (2), 25 Defining the source, language, geography, key concepts, and alarm rules.  Social media module (101), live broadcast module (102), digital press module (103) and The echo news module (104) connects to the defined sources and provides the most for each piece of content received. event ID, source ID, source time, acquisition time, media type, media An event envelope containing the location, parent / child content link, access, and quality fields is 30 creation,  Data collection and flow layer (201) by classifying events by source and importance class Segmentation, retry, backpressure, prioritization, and delivery guarantee. implementation,  Data normalization and preprocessing unit (202) by events common event envelope conversion and time zone, language, character encoding, URL, media type, rights information, Standardization of data family trees using hash values,  Text hash, minimal hash / similarity hash, visual perceptual hash, and rapid embedding The process of uniquely combining identical or highly similar content by utilizing similarity, same 5 If it is determined that the content has been processed previously, the existing model outputs will be affected. referencing and only new interaction data (118) and spread data (135) updating,  Normalized event video / visual preprocessing unit (203), audio preprocessing unit according to media type redirection to the processing unit (204) or text preprocessing unit (205) and content from one to 10 If it involves multiple modalities, the sub-streams in question should be executed in parallel.  Social media video / visual data (115) by the video / visual preprocessing unit (203) and Finding scene changes on live stream video / visual data (125), key Selecting squares and eliminating distorted / blurry squares,  The selected squares are first fed into a computer vision model (215) to determine the person / face region, 15 Detection of company logos, objects, screen text, scenes and motion events, boxes, generating a confidence score for the class and only showing the frames or clips containing the relevant finding. Transmission to the VLM model (207),  Textual and image or short video segment transmitted by the VLM model (207) a structured explanation and contextual analysis of the event, person, object, and text in the visual. removal of the relationship  Social media speech data (116) and live stream by the voice preprocessing unit (204) Separation of voice data (126) into speech segments, noise reduction, sampling standardizing the rate and determining speaker responses,  The STT model (206) converts speech into word or segment timestamped text 25 Transformation and generation of alternative textual knowledge with language, confidence,  Title / text (130), comment text (117) and by the text preprocessing unit (205). Turkish morphology, sentence division, spelling on text / transcription data (127) normalization, name protection, personal data masking, and text fragmentation,  The LLM model (208) extracts person / institution / product / location / date, event, claim, subject, from the text. Extraction of emotions, attitudes, intentions, cause-and-effect relationships, and summaries from data storage / knowledge base. (214) adding the inferred document fragments to the context of inference and the schema of the output Generating a JSON structure that depends on its validation,  Timestamp and semantic embedding by multimodal semantic analysis engine (209) The video scene, STT segment, OCR text, and text paragraph are 35 by using the space together. matching; time with cosine similarities between text, image and sound embeddings, 16 The combined similarity is the weighted sum of the source and parent / child content metadata match. calculation of the score,  Calibration of modality confidence scores by multimodal semantic analysis engine (209) By calculating the combined event confidence, low confidence results can be re-evaluated. directing to processing or human examination and the results of semantic analysis (302) 5 production,  News about the institution (132) made by the Echo news module (104) through the headline / text, visual, speech and source metadata of the news (133) comparison,  Comparison of the calculated combined similarity score with the threshold and source-time relationship 10 Taking into consideration the echo and reproduction data (134) full copy, partial copy, quote, summary, with one of the semantic rewriting or visual / audio derivative classes and a confidence score. labeling,  Insight and relationship analysis engine (210) by diffusion data (135) and citation / reference Converting the data (136) into a timestamped directional graph; content A from content B 15 first publication, exceeding the threshold of semantic similarity between them, explicit reference the occurrence of at least one of the following conditions: being present or sharing a common identity. In this case, increasing the weight of the origin / propagation edge in the A→B direction,  The resulting directional graph shows the initial source, cross-platform jump, and diffusion cascade. and obtaining echo news spread results (305) by determining the content family tree, 20  Content, actor, source, institution, subject, event and insight and relationship analysis engine (210) Establishing weighted and timestamped graphic edges between citation nodes,  The relationship graph created includes degree, eigenvector / page order centrality, ensemble, and bridge. Interaction and relationship networks are determined by calculating node, common behavior, influence, and echo scores. (304) generation, (An echo score e.g. E = α·R + β·C + γ·V + δ·D + ε·Q with 25 It can be calculated. R represents the number of reshares / copies, C represents the number of unique resources, and V represents... D represents the rate of deployment, D represents cross-platform diversity, and Q represents the source quality / reliability. (Coefficients can be learned based on the usage scenario or determined by the administrator.)  Sliding track for each topic, entity and event by the trend and anomaly analysis engine (211). Number of content items in windows, unique source, interaction, emotion, topic, spread rate, and 30 Comparison of echo score measurements with baseline and trend and anomaly reports. (303) creation,  Seasonal average, z-score, online exchange point, and Isolation Forest scores Generating a combined anomaly score by combining them and sudden increases, decreases, and topic shifts, Coordinated sharing, unusual resource concentration, or model confidence decline detected 35 producing an anomaly event when done, 17  Keyword and event triggering engine (212) by word / phrase, semantic similarity, entity-event combination, visual logo / object, emotion / attitude, threshold or event Evaluation of the order of rules,  The notification and alarm engine (213) detects recurring alerts related to the same root event in a single system. Combining the factors in the event card: impact, trust, rate of spread, source diversity, and user role. Priority upgrade implementation and generation of instant alerts and notifications (301),  Raw media, normalized event, model output, embedding, graphic edge, report, and audit. the record is stored in the data storage / information database in version (214),  Real-time flow, timeline, resource allocation, map, relationship via user interface (2) the presentation of the graph, echo chain and evidence media pieces to the user (1), 10  Search / query / report API (306) via keyword, filter, full text, vector Meeting the queries for similarity, graph neighborhood, and time interval,  User's (1) correct / incorrect, relevant / irrelevant, same event / different event or alarm importance feedback recording of the notification in the model training data repository,  Low safety, high impact, or model 15 by active learning mechanism Selection of samples with discrepancies into the labeling queue and STT model (206), The VLM model (207), the computer vision model (215) and the LLM model (208) are the areas Updating adaptation datasets,  Monitoring model quality, latency, resource utilization, data distribution, and error rate, In case of data / model shift, the shadow model, A / B testing, or retraining process is 20 after initiation and testing of the new model version on shadow traffic phased implementation  Implementation of role-based access, encryption, audit logging, retention, and deletion policies. and relating model and data versions to the results. Below are the hardware and components used in a preferred implementation of the system described in the invention. Software layers are described. Product and project names mentioned in this section are for informational purposes only. This is for illustrative purposes only and is not limiting; each layer provides the described functionality. Equivalent open-source or commercial components can be used. Physical Infrastructure Layer The system must include at least one general-purpose processor, memory, a permanent storage unit, a network interface, and one or more components. on at least one computer system that includes more graphics processing units / hardware accelerators is executed. Graphics processing unit, VLM model (207), STT model (206), computer vision model (215) and for the LLM model (208) to work in real time or near real time 35 It meets the required computational intensity. These components are located on a single server. 18 as it can be located in a server cluster, private cloud, hybrid cloud or edge nodes. They can also be deployed in a distributed manner. Container and Orchestration Layer The system components are containerized operating systems (preferably Linux) running on a single operating system. time (e.g., containerd) and container orchestration platform (e.g., Kubernetes) This is done through microservices. Processor and graphics processing unit resource requests, horizontal scaling, node selection, health check, reboot in case of failure, and High availability is managed at this layer. This allows resource modules (101, 102, 103, 104) and the load of the analysis engines (209, 210, 211, 212, 213) depends on the number of sources monitored and the data 10 It can scale independently depending on its size. Event Flow Layer Data collection and streaming layer (201), on an event streaming platform (e.g. Apache Kafka) This is implemented; events from source modules are displayed in persistent and segmented topic streams. It is retained. Real-time conversion, windowing, merging, complex event processing, status holding and processing semantics, either fully once or effectively at least once, a flow processing engine (provided by Apache Flink, for example); keyword and event triggering engine (212) sliding window of the trend and anomaly analysis engine (211) with rule evaluation Calculations can be performed on this layer. Periodic literature review, model training, and 20 Scheduled workflows, such as report generation, can be done with a workflow scheduler (e.g., Apache Airflow). It is managed. Data and Information Layer Data storage / knowledge base (214), multiple storage serving different access patterns 25 It consists of the following components: Raw media in an object repository (e.g., Ceph); processing and... configuration data in a relational database (e.g., PostgreSQL); semantic embedding vectors, in a vector index (e.g., pgvector); full-text content and log records, in a search and log indexing (e.g., OpenSearch); insight and relationship analysis engine (210) The relationship graph generated by this is then displayed in a graph database layer (e.g., Apache AGE) 30 Each record is stored with a version, data family tree, retention period, and audit log. Equivalent components that provide the same functionality can be used, depending on the needs. Artificial Intelligence Development Layer A deep learning framework for training, fine-tuning, and domain adaptation of models (e.g., 35 PyTorch), model libraries, and parameter-efficient fine-tuning tools (e.g., Hugging) 19 Face Transformers and PEFT), classic machine learning libraries (e.g., scikit-learn), object detection and optical character recognition tools (e.g., MMDetection and PaddleOCR), Turkish morphology and natural language processing with speech processing tools (e.g., SpeechBrain) Libraries (e.g., spaCy and Zemberek) are used. Training experiments and model records are created. In a model lifecycle platform (e.g., MLflow), dataset versions are categorized as data 5. It is monitored with a versioning tool (e.g., DVC); thus, each analysis result is tracked by the model that produced it. and can be associated with the data version. Model Presentation Layer The STT model (206), VLM model (207), LLM model (208) and computer vision model (215), a 10 model serving infrastructure (e.g., KServe) and large language model inference engine (e.g., vLLM) It is offered as a service via REST or gRPC interfaces. At this layer dynamic stacking, graphics processing unit sharing, model parallelism, automatic scaling, pilot Application and health monitoring are performed. Computerized vision requiring low latency and Speech-to-text services operate independently as separate microservices. 15 Scalable. Application and API Layer Search / query / reporting API (306), implemented with a web application framework (e.g., FastAPI) It offers REST, gRPC or GraphQL interfaces through services. The user interface (2) is a 20 It is built with a web interface library (e.g., React) and searches indexes query endpoints. It runs filtered search, vector similarity, and graph neighborhood queries. Instantaneous Warnings and notifications (301), WebSocket or server-sent event stream to user is transmitted to the interface (2); also via web trigger or enterprise event bus to the internal organization. can be transferred to systems. 25 Security and Observability Layer Authentication and authorization are done through an identity management component (e.g., Keycloak). This is provided by OAuth2 / OIDC protocols; role-based access control and multi-tenant isolation. It is implemented at the layer. Transmission security is with TLS, and persistent data security is with storage encryption. Metric, trace and log records are provided, along with an observability standard and a set of tools. (e.g., OpenTelemetry, Prometheus, Grafana, and Loki) are used to collect data on model quality, latency, Data or model drift can be detected by monitoring resource usage, data distribution, and error rate. Shadow modeling, A / B testing, or retraining processes are triggered. 35 Distribution Flexibility In deployment, some elements can be combined within the same software service, or a single element can be combined from multiple sources. It can be devoted to additional physical services. These changes affect the function of the element as defined in the specification. It is within the scope of the invention as long as it is protected. Similarly, the software mentioned as an example above. Replacing the components with equivalent components that perform the same technical function, This does not mean going beyond the scope of the invention. 5 Below are alternative implementations that can be applied without altering the basic operation of the invention. The forms are described. Implementation with any of these alternatives constitutes a breach of the scope of the invention. It does not mean going outside of it. Alternatives to Data Sources The data sources to which the resource modules (101, 102, 103, 104) are connected are exemplified in the specification. This is not limited to social media platforms, live broadcasts, and digital press sources. It also... operations, call center voice recordings, podcast broadcasts, radio broadcasts, corporate messaging channels, forums, e-commerce product reviews, customer complaint platforms, app store 15 This can also be applied to evaluations or sensor event streams. From these sources The received content is also converted into the same event envelope structure and fed into the same processing pipeline. Live broadcast module (102), satellite broadcasts (123) as well as IPTV, OTT, RTSP, SRT, HLS, DVB-T / T2 can also be used over cable television, radio broadcasting or conference broadcasting protocols. It can receive live video and audio streams. Alternatives to the Implementation of Artificial Intelligence Models STT model (206), VLM model (207), computer vision model (215) and LLM model (208), They can be operated as separate, independent models, or as a single integrated multi-mode 25 The model can also be implemented as different service ends. The model architecture, its provider, or Changing the number of parameters does not alter the function of the relevant element as defined in the specification, and It does not affect the scope of the invention. Alternatives to Multimodal Fusion Methods 30 The intermodal matching and merging function of the multimodal semantic analysis engine (209); shared embedding space, late fusion, early fusion, cross-attention mechanism, graph-based through the hybrid use of fusion or rule-based learning methods realizable. 35 Alternatives to the Application Object of Echo and Replication Analysis 21 The Echo news module (104) replication, origin and dissemination analysis is only for news content. No; product descriptions, campaign messages, video clips, speeches, visuals, logo usage, technical documents, regulatory announcements or academic This can also be applied to publications. In this case, the news made by the institution (132) and about the institution The elements of the news reports (133) correspond to the source and derivative pools of the relevant content type. 5 Alternatives for Storing and Processing Relationship Graphs The graph produced by the insight and relationship analysis engine (210) is an information graph, feature graph, RDF. It can be stored in a triple repository or time-series edge list format. Centrality, community, and influence. Graphical algorithms that include calculations are streamed online or in periodic batches. They can be employed in the workplace. Alternatives to Trend and Anomaly Analysis Methods Function of the trend and anomaly analysis engine (211); statistical threshold methods, Bayesian variation point detection, deep autoencoder-based error reconstruction, online learning, 15 The Hawking process can be implemented using bounce detection methods or a combination thereof. Alternatives to Event Triggering Methods Keyword and event triggering engine (212), defined only by user (1) It does not have to rely on rules; a trained classifier can classify 20 samples with zero or few samples. LLM classification is a pattern identified on a semantic search result or relationship graph. Matching can also trigger an event. User-Related Alternatives The user (1) does not have to be a real person. User interface (2) and 25 Another software that calls the system's functions via the search / query / reporting API (306), An autonomous agent or a process orchestrator can also fulfill this function. Alternatives for Output Integration Instant alerts and notifications (301), semantic analysis results (302), trend and anomaly reports 30 (303), interaction and relationship networks (304) and echo news spread results (305), user interface In addition to being offered via (2); to the corporate data lake, SIEM / SOAR platforms, customer relationship management systems, call center applications, content management to their systems, newsroom systems, workflow engines, or decision support systems It can be transmitted. 35 22 Areas of Application The system and methods covered in the invention are: brand reputation monitoring, crisis early warning, and disinformation. Detection and verification, election and disaster monitoring, investor relations, competitive intelligence, copyright infringement detection, content moderation, security intelligence, ad verification, and public opinion analysis. It can be used in applications. Changes in the area of ​​use may affect the technical operation of the system and 5 It does not change the function of its elements.

Claims

23 REQUESTS 1. It is a system for monitoring and multimodal analysis of multi-source media content, and its feature is;  at least one processor, memory, permanent storage unit, network interface and at least one graphics processor a unit containing at least one 5 elements that enables all elements to perform their functions computer system,  via authorized APIs, feeds, RSS / Atom, or authorized connectors from social platforms Data field; a source that transforms content, interaction, and relationship events into a common schema. Social media module (101), which is a set of adapters,  Continuous video / audio reception from satellite, internet, or live streaming protocols, timestamp 10 generating, creating seamless segments and passing the stream to the real-time processing layer live broadcast module (102),  Gathers news sites, blogs, press releases and similar web content; page body, Digital press module that extracts title, author, date, link and content structure (103),  Comparing the news produced by the institution with the news published about the institution; 15 Echo graph by calculating the relationships of replication, citation, reference, origin, and diffusion. Echo News Module (104), which creates  source connectors, tails, segmentation, retry, back pressure, Data collection and flow layer that manages prioritization and flow routing (201),  Converts data from different sources into a common event envelope; character encoding, 20 o'clock slice, language, identity, content mix, rights / access, quality, and data family tree domains standardizing data normalization and preprocessing unit (202),  codec decoding, resizing, scene change detection, keyframe selection, perceptual hash, OCR region extraction, blur / quality control, and image video / image preprocessing unit that performs normalization (203), 25 • Audio channel separation, sample rate conversion, noise reduction, audio efficiency speaker detection, speaker segmentation, audio level normalization, and time slicing. sound preprocessing unit (204),  Unicode normalization, language detection, sentence splitting, spell check, URL / tag parsing, personal data masking, text cleaning and fragmentation processes 30 text preprocessing unit (205),  STT model which converts speech into timestamped text (206),  Natural language context for person / object / event / screen text on a frame, scene or short video VLM model (207), which describes the structured output form, 24  Detects objects, logos, faces, scenes, text, and motion; handles VLM calls. Computer vision model that acts as a fast pre-filter that narrows (215),  LLM that produces event, entity, claim, summary, emotion, attitude, topic, relationship and structured output model (208),  STT, VLM, CV, and LLM outputs, along with resource metadata, are sorted by time and identity. unifying elements; connecting events / entities, drawing conclusions, similarities, contradictions, themes, emotions, and Multimodal semantic analysis engine that performs trust merging operations (209),  An updatable system containing sources, content, individuals, institutions, topics, events, references, and interactions. Information / diffusion graph is established by considering centrality, community, influence, relational strength, origin, and echo. insight and relationship analysis engine that generates scores (210), 10  Sliding time windows, seasonal baseline, change point, z-score, Insulation The forest, with its online clustering and topic shifting methods, exhibits sudden increases, decreases, and identifying unusual patterns and basing measurements on sliding windows Trend and anomaly analysis engine that produces a combined anomaly score by comparing (211),  word, expression, coherent expression, entity-event combination, semantic proximity, visual object / logo, 15 Key words and events that evaluate emotion / attitude, threshold, and sequence event rules. trigger motor (212),  Combining repeated alerts linked to the same root event into a single event card The factors that provide this are importance, trust, resource diversity, repetition, user role, and time. Combining according to its window; suppressing recurring alarm, upgrade and delivery 20 Notification and alarm engine (213) that monitors the status,  Raw object repository, relational database, lookup index, vector index, and graph. a multi-layer system that keeps track of version, family tree, retention period, and audit logs. data storage / knowledge base which is storage structure (214),  Resource management, query, real-time stream, relationship graph, map / timeline, 25 Management, visualization, and operations providing model confidence score and audit log. user interface (2) It includes.

2. The system is compliant with Request 1 and its feature is; 30 obtained from the social media module (101). data, including video, keyframe, thumbnail, photo, and screenshot associated with the post. social media video / image with its resolution, codec and duration metadata. data (115); audio preprocessing and found in social media video or audio recording. Social media is a speech signal given to the speech-to-text conversion process. Speech data (116); natural language such as post text, description, comment, reply and tag 35 comment text (117) with content and likes, shares, reposts, comments, replies, In distribution and relationship analysis, which includes follower, view, time, and actor relationships. It includes interaction data (118) which are event logs used.

3. The system complies with Request 1 and its feature is that the data obtained from the live broadcast module (102), Live Stream 5 is the visual frames, scene clips, and screen components of a live broadcast stream. video / visual data (125); speech, music, ambient sound and channel data separated from live broadcast. live broadcast audio data (126) and broadcast subtitles, teletext, which are audio streams containing information. Text / script that is a timestamped text produced as an OCR output or STT result. It contains data (127).

4. The system complies with Claim 1, and its feature is that the sources monitored by the digital press module (103), Digital tracking via RSS, sitemap, authorized web access, or publisher integration. news sources such as news websites (128) and corporate or individual blogs blog posts that are the source of published text, visual and link content (129) It includes. 15 5. The system complies with Claim 1 and its feature is that the data obtained from the digital press module (103), A news item is a natural language data consisting of headline, lead paragraph, body, subheading, description, and tags. title / text (130) with author, publication date, URL, chapter, media attachments, source links, It includes content data (131), which is the content mix and other structured metadata. 20 6. The system complies with Request 1 and its features include: the institution's official website, press room, and social media. The content produced by the institution is source content published through its account or authorized broadcasting channels. news (132) and third-party sources about the institution, brand, product, manager, activity or news about the institution which is the content published about the incident (133) 25 It includes the comparing echo news module (104).

7. The system compliant with Claim 1 is characterized by being a VLM model implemented with a visual-language model. (207) is included.

8. The system compliant with Claim 1 is characterized by being based on an object detection and OCR library. It includes a computer vision model (215).

9. It is a system that complies with Claim 1, and its feature is that it is implemented with a large language model, RAG and field. It includes the LLM model (208) which supports its adaptation. 35 26 10. The system complies with Claim 1, and its characteristics include: source, monitoring subject, and key on the system. Defines concepts, event rules, reports, and alarms; reviews analysis results; provides feedback. and includes the user (1) who is the authorized human or service account providing the label.

11. The system compliant with Claim 1 is characterized by its interface, email, and messaging for specified events. Low latency generated via web trigger or enterprise event bus It includes instant alerts and notifications (301) which are notifications.

12. The system conforming to Claim 1 is characterized by its summary, subject, emotion, attitude, entity, claim, event, 10 results that present similarities, contradictions, and evidence links in a structured way The results of the semantic analysis (302) include time series, subject set, source, region, actor and Trends are reports that include platform-based trend, change point, and anomaly information. and includes anomaly reports (303).

13. A system that complies with Claim 1, and whose characteristics are; resources, content, actors, institutions and events 15 visualizing the relationship graph between them, along with centrality, community, and influence metrics. the outputs are interaction and relationship networks (304) with the original source, copy / derivative content, citation chain, The results show the timeline, platform transition, deployment rate, and effective nodes. The echo includes news spread results (305).

14. It is a method for monitoring and multimodal analysis of multi-source media content, feature;  social media module (101), live broadcast module (102), digital press module (103) and The echo news module (104) connects to the defined sources and for each piece of content received. at least event ID, source ID, source time, acquisition time, media type, media 25 an event envelope that includes location, parent / child content linking, access, and quality fields creation,  data normalization and preprocessing unit (202) by events common event envelope conversion and time zone, language, character encoding, URL, media type, rights information, Standardization of data family trees using hash values, 30  video / visual preprocessing unit (203), audio preprocessing according to media type of normalized event redirection to the processing unit (204) or text preprocessing unit (205) and the content If it includes more than one modalities, these sub-streams can be run in parallel. execution, 27  Social media video / visual data (115) by video / visual preprocessing unit (203) and scene changes on live stream video / visual data (125), Selecting the keyframe and eliminating distorted / blurry frames,  The selected squares are first fed into the computer vision model (215) to determine the person / face region, Identifying company logos, objects, screen text, scenes, and motion events, boxes, 5 generating a confidence score for the class and only showing the frames or clips containing the relevant finding. Transmission to the VLM model (207),  Textual and image or short video segment transmitted by the VLM model (207) a structured explanation and the event, person, object, and text in the visual Extraction of contextual relationship, 10  Social media speech data (116) and live by the voice preprocessing unit (204) Separation of broadcast audio data (126) into speech segments, noise reduction, standardizing the sampling rate and determining speaker returns,  Speech time stamped into word or segment text by the STT model (206) transformation and generation of alternative textual knowledge with language, confidence, 15  Title / text (130), comment text (117) and by the text preprocessing unit (205). Sentence splitting, spelling normalization, custom on text / transcription data (127) Name protection, personal data masking, and text fragmentation are performed.  Person / institution / product / location / date, event, claim, subject, from the text by the LLM model (208), extracting emotion, attitude, intention, cause and effect, and summary, 20  timestamp and semantics by multimodal semantic analysis engine (209) The embedding space is used together to create a video scene, STT segment, OCR text, and text. Paragraph matching and cosine between text, image, and audio embeddings Similarities and weighted averages for time, source, and upper / lower content metadata consistency. Calculation of the combined similarity score from the total, 25  News made by the institution (132) by the echo news module (104) and the institution Headline / text, visual, speech and source metadata of news reports (133) about it Comparison based on data,  Comparison of the calculated combined similarity score with the threshold and source-time Taking into account the relationship, the echo and reproduction data (134) full copy, partial copy, 30 with one of the following classes: quote, summary, semantic rewriting, or visual / audio derivative, and labeling with a trust score,  Insight and relationship analysis engine (210) by diffusion data (135) and citation / reference Converting the data (136) into a time-stamped directional graph,  Trend and anomaly analysis engine (211) by seasonal mean, z-score, 35 The combined anomaly is the result of combining online exchange point and Isolation Forest points. 28 the generation of points and sudden increases, decreases, topic shifts, coordinated sharing, Anomaly when unusual resource concentration or model confidence decline is detected. the production of the event,  keyword and event triggering engine (212) by word / phrase, semantic Similarity, entity-event combination, visual logo / object, emotion / attitude, threshold or event 5 Evaluation of the order of rules,  The notification and alarm engine (213) detects recurring alerts related to the same root event. an event card, combining factors such as impact, trust, rate of spread, and source diversity. Priority and upgrades are applied according to user role. It includes the steps of the process. 10 15. The method is compliant with claim 14 and its feature is that the user interface (2) is used by the user (1). the institutions, individuals, topics, events, sources, languages, geographies, key concepts to be tracked through This includes the step of defining alarm rules.

16. The method is compliant with claim 14 and its feature is; data collection and flow layer (201) retrying, revisiting by classifying events according to their source and importance. This includes the step of implementing pressure, priority, and delivery guarantee.

17. This method complies with Claim 14 and its characteristics are: text hash, minimum hash / similarity hash, 20 Using visual perceptual complexity and rapid embedding similarity, the same or higher The process of singularizing similar content, identifying that the same content has been processed before. referencing existing model outputs and only new interaction data (118) This includes the step of updating the spread data (135).

18. The method is compliant with claim 14, and its feature is data storage / information by the LLM model (208). the step of adding the document fragments brought from the base (214) to the context of the inference It includes.

19. The method is compliant with claim 14 and its feature is; multimodal semantic analysis engine (209) 30 by calibrating modality confidence scores to determine combined event confidence. calculation, reprocessing of low confidence results or human review It includes the steps of directing and producing the results of semantic analysis (302). 29 20. This method complies with Claim 14 and its characteristic is that the initial source is retrieved via the resulting directional graph. Cross-platform leaps, diffusion cascades, and content lineage identification are used to create echo news. This includes the step of obtaining the spread results (305).

21. The method is compliant with claim 14 and its feature is that it is performed by the insight and relationship analysis engine (210) 5 weighted and temporal relationships between content, actor, source, institution, subject, event, and reference nodes. This involves the step of establishing stamped graphic edges.

22. This method complies with Claim 14 and its characteristic is that the generated relationship graph shows the degree, eigenvector / page. The ranking is based on 10 scores for centrality, community, bridge node, common behavior, influence, and echo. It includes the step of generating interaction and relationship networks (304) by calculation.

23. The method is compliant with claim 14 and its feature is that it is performed by a trend and anomaly analysis engine (211). The number of content items in sliding windows for each topic, entity, and event, unique source, The baseline for measures of interaction, emotion, subject, dissemination speed, and echo score is 15. This includes the step of comparing and generating trend and anomaly reports (303).

24. The method is in accordance with claim 14, and its feature is that it is instantaneous by the notification and alarm engine (213). It includes the step of generating warnings and notifications (301).

25. This method complies with Claim 14 and its characteristics include: raw media, normalized event, model output, embedded, graphic edge, report and audit record data storage / information database (214) It includes the step of storing it in a versioned manner.

26. The method is compliant with claim 14 and its feature is; real-time flow via user interface (2), time 25 chart, source allocation, map, correlation graph, echo chain, and evidence media pieces It includes the step of presenting it to the user (1).

27. The method compliant with Claim 14 is characterized by its search / query / reporting via API (306). keyword, filter, full text, vector similarity, graphic neighborhood, and time interval 30 It includes the step of answering their questions.

28. The method is in accordance with claim 14 and its feature is that the user (1) can choose right / wrong, relevant / irrelevant, same Event / different event or alarm severity feedback to the model training data pool. It includes the recording step. 35 29. This method complies with claim 14 and is characterized by its low efficiency due to the active learning mechanism. Safe, high-impact, or pattern-mismatched samples are added to the labeling queue. selection and STT model (206), VLM model (207), computer vision model (215) and LLM It includes the step of updating the field adaptation datasets of the model (208).

30. This method complies with Claim 14 and is characterized by its model quality, latency, resource utilization, and data usage. Monitoring the distribution and error rate, shadow model in case of data / model shift, A / B testing or initiating the retraining process and shadow traffic of the new model release It involves a phased rollout process after testing.

31. This method complies with Claim 14 and features role-based access, encryption, and audit logging. The implementation of retention and deletion policies and model and data versions with results It includes the step of establishing the relationship. 20

Images