GumGum Contextual Description of Methodology
- 1 Media Ratings Council (MRC) Content-Level Accreditation
- 2 Primary Users and Use Cases
- 3 GumGum Contextual Functions
- 4 GumGum Contextual Machine Learning Technology
- 5 Architecture and Flow
- 6 Brand Safety
- 7 Content Classification
- 8 GumGum Contextual Classification and Brand Safety Report
- 9 Classification Approaches
- 10 GumGum Contextual and the GARM Brand Safety Floor
- 11 Integration Methods
- 11.1 API Integration
- 11.2 Page Tags
- 12 Processing Time
- 13 Machine Learning Model Development
- 14 Page Minimum Reporting Requirements
- 15 Image Formats Analyzed
- 16 Video Data Analyzed
- 17 User Information is Not Analyzed
This Description of Methodology (DoM) describes the processes that deliver GumGum Contextual – GumGum’s content-level contextual analysis and brand safety solution.
Powered by GumGum’s AI technology, GumGum Contextual applies sophisticated machine learning techniques to analyze digital content, including web pages, images, and videos (plus audio).
GumGum Contextual returns a detailed report featuring brand safety scores for the content, along with contextual targeting categories, prominent keywords, and sentiment categories.
GumGum Contextual supports the contextual targeting categories defined in the Interactive Advertising Bureau (IAB) Content Taxonomy v1.0, 2.0, and 3.0.
Media Ratings Council (MRC) Content-Level Accreditation
GumGum Contextual is the first independent third-party solution to achieve MRC accreditation for content-level brand safety.
This recognition by the MRC validates that GumGum’s proprietary contextual intelligence solution is able to consider all available signals (text, image, audio, and video) needed to give a true contextual reading.
GumGum Contextual is officially accredited for content-level Contextual Analysis, Brand Safety and Brand Suitability for English-language text/image, video image, and audio classification (Desktop, Mobile Web, CTV).
Primary Users and Use Cases
GumGum Contextual serves agencies, advertisers, DSPs, and publishers as a third-party content-level contextual analysis and brand safety data solution.
Operating as a fee-based third-party service in the cloud, publishers can integrate GumGum Contextual into content management systems (CMS) or data management platforms (DMP) to analyze and optimize media content.
Supply-side and demand-side platforms (SSPs and DSPs) can implement the GumGum Contextual service on their own technology platforms, ad exchanges, and ad servers. There are two primary product use cases:
Increased Brand Safety — Advertisers can deploy GumGum Contextual to detect objectionable content and avoid serving their advertising messaging adjacent to or embedded within that content. Publishers can use GumGum Contextual to identify and assess potentially objectionable digital content prior to publication.
Optimum Contextual Targeting — Advertisers and Publishers can access the GumGum Contextual service to locate content that is highly relevant, enabling contextually aligned advertising to be served.
GumGum Contextual’s core technology remains unchanged for each implementation. Integrations are accomplished via the GumGum Contextual API.
As of September 2024, GumGum's GumGum Contextual service processes ~2.5 billion unique monthly requests for content and brand safety classification globally.
GumGum Contextual Functions
GumGum Contextual’s function is to provide data to clients who explicitly request and pay for analysis information about specific digital content. The clients are interested in establishing brand suitability and contextual classification for specific content, to drive their own content creation or ad serving.
GumGum Contextual applies natural language processing (NLP) and computer vision (CV) based machine learning techniques to analyze digital content. Multiple kinds of content can be analyzed, such as desktop and mobile web pages, images, and Online Video platforms (OLV) and connected TV (CTV) videos (including audio).
Web Page Analysis Functions
Going beyond simple strategies like identifying keywords on the page or in the URL string or metadata, GumGum Contextual works by scanning the full text and prominent imagery of a web page. GumGum Contextual’s NLP processes analyze the core page content, while CV processes analyze the imagery.
GumGum Contextual provides what the Media Ratings Council (MRC) refers to as content-level reporting defined as “more granular context and brand safety measurement and reporting for video and display content within a domain, site, platform, mobile application or URL”.
Note the following details about GumGum Contextual web page content-level processing:
GumGum Contextual does not apply content-level analysis to code or objects (including third-party code or objects) that appear outside, adjacent to, or embedded within the core text on a page.
GumGum Contextual does not download or analyze the CSS, JavaScript, navigation, footer, sidebars, and other areas extraneous to the core textual content on the page. For example, on a typical Blog page GumGum Contextual extracts and analyzes the central content of the page, but not the surrounding elements such as third-party advertising or related content.
GumGum Contextual also does not provide analysis of continually changing dynamically loaded user-generated content within publisher pages (e.g., reviews sections, comments sections, social media plug-ins) or social media environments.
GumGum Contextual applies logic to identify the prominent image on a web page for analysis. Additional images on the page may be subject to image extraction limitations.
GumGum informs clients that GumGum Contextual analyzes the web page (not the surrounding material) specifying that the analysis includes the core textual content and prominent imagery but nothing else – not graphics, sidebar content, or third-party insertions such as paid advertising.
GumGum Contextual acknowledges that surrounding, adjacent, or embedded content on a web page (which may be provided by JavaScript executions or non-textual content) can affect the context of a page as presented to users and may be a consideration for advertisers.
Other key platform functions such as ad serving, detection of ad fraud, identification of invalid traffic (IVT/SIVT), measurement of viewability, measurement of audiences, and other cookie implementations are not handled by GumGum Contextual or its technology.
Video Analysis Functions
GumGum Contextual analyzes video content by applying powerful classifiers to the video’s transcribed audio track and image data from sampled video frames.
Video analysis leverages GumGum’s industry-leading NLP text analysis and CV image analysis processes, plus fast and accurate audio transcription services.
GumGum Contextual Machine Learning Technology
GumGum Contextual is the only solution that applies machine learning techniques to provide content-level brand safety and contextual analysis. Alternative solutions may only leverage keyword methodologies that consider the text and are limited to page-level analysis, use of Allow or Blocklists, or URL-level analysis. These cruder contextual approaches often eliminate safe and relevant inventory. They also miss relevant content (e.g., keywords that are spelled differently), overlook related content, and mistakenly target irrelevant content (e.g., keywords with multiple meanings).
GumGum Contextual’s supervised machine learning works by first training a machine learning model with training data that comprises thousands of pieces of example content (i.e. pages, images, and videos) for each category paired with the correctly labeled outputs. For example, to learn how to classify a GumGum threat category on “Drugs and alcohol”, first a human has to hand-annotate thousands of pieces of content that have something to do with drugs or alcohol.
The supervised learning algorithm searches for patterns in the data that correlate with the desired outputs. After training, the supervised learning algorithm can process new unseen pages and label them with a classification based on the prior training data. For example, the model could predict whether digital content references drugs or alcohol and classify it accordingly for the purposes of brand safety.
Architecture and Flow
Customers use GumGum Contextual to analyze specific digital content and determine the eligibility of the content for ads. GumGum Contextual does not crawl the internet for content; instead, a client application calls GumGum Contextual (via their integration with the GumGum Contextual API) specifying the URLs of specific content they’d like to analyze.
GumGum's GumGum Contextual service exists entirely within a secure Cloud infrastructure. GumGum Contextual’s Cloud-based architecture is massively scalable and currently processes approximately 2.5 billion unique requests per month for content and brand safety classification.
Access for GumGum Contextual User Agents
If a requested URL blocks a GumGum Contextual browser, GumGum Contextual cannot process the content and returns an error. GumGum Contextual customers are therefore requested to configure their domain access permissions to enable GumGum Contextual to access their site in order to extract and process content.
Page Analysis Process
The GumGum Contextual page analysis process involves the following core components:
GumGum Contextual API Gateway: The GumGum Contextual API Gateway receives a page URL request, authenticates the client request and passes the URL to the GumGum Contextual API.
GumGum Contextual API: The GumGum Contextual API initiates the request and then orchestrates the Content Extractor, Text and Image analyses systems to extract the page data and perform the analyses.
Content Extractor: The Content Extractor accepts page requests sent by the GumGum Contextual API from a queue. The Content Extractor loads the page URL, downloads the page title, metadata, and HTML and saves it as a text string in the database. If a prominent image is identified for the page, the Content Extractor downloads and saves the image to the database with identification information for the associated page. The Content Extractor passes the Page URL and image information on for text and image analysis.
Text Analysis: The Text Analysis engine applies Natural Language Processing (NLP) for text classification (e.g. IAB and Threat categories) and information extraction (e.g. Keywords).
Image analysis: The Image Analysis engine houses GumGum’s core Computer Vision capabilities in a modular architecture. The Image Analysis component passes images through multiple data models to determine their classification information.
GumGum Contextual Report: The GumGum Contextual API retrieves the text and image classification results, applies weighting and merging logic to the results, and returns the final GumGum Contextual page report to the client.
Video Analysis
GumGum Contextual analyzes videos for the purposes of content-level contextual targeting and brand safety.
GumGum Contextual works by applying machine learning techniques to the video audio track, sampled video frames, and video metadata (where available) and assigning contextual categories, detecting keywords, and calculating a brand safety score.
GumGum Contextual Video Analysis leverages the following systems:
Transcribe Service – Applies automatic speech recognition (ASR) to convert speech to text.
OCR Service – Performs Optical Character Recognition (OCR) to detect text in video and convert the detected text into machine-readable text.
GumGum Contextual Text Processing – Applies machine learning models to the video metadata, title, transcription text, and OCR text and provides a brand safety and contextual classification report.
GumGum Contextual Image Processing – Applies machine learning models to sampled video frames and provides a brand safety report.
Video Analysis Process
The GumGum Contextual video analysis process involves the following core components:
GumGum Contextual API Gateway: The GumGum Contextual API Gateway receives a video URL request, authenticates the client request and passes the URL to the GumGum Contextual API.
GumGum Contextual API: The GumGum Contextual API passes the request to the Video Service to orchestrate video analysis.
Video Service: the Video Service downloads video and audio into separate files.
Audio Transcribe: The audio file is sent for transcription.
Optical Character Recognition (OCR): GumGum Contextual API verifies if the audio transcription results contain a sufficient sample of at least 50 words. If not, GumGum Contextual API initiates an OCR job to detect text in the video file and convert the detected text into machine-readable text.
Prism Video Frame Threat Classifier: Video is sent to the Video Threat Classifier for brand safety analysis of video frames.
GumGum Contextual Text Processing: GumGum Contextual API passes concatenated text results (comprising transcription, OCR if available, Client metadata title and description) to GumGum Contextual Tapas Text Processing. The Text Processing engine processes the video transcription, OCR, client metadata title and description by applying Natural Language Processing (NLP) for text classification (e.g. IAB Content Categories v2.0 and Threat categories) and information extraction (e.g. Keywords).
GumGum Contextual Report: The GumGum Contextual API accepts the text analysis results, applies result weighting and merging logic, then returns the final video analysis GumGum Contextual Report to the client.
Brand Safety
GumGum Contextual Machine learning predicts threat categories by applying data models trained on collections of various kinds of threatening content. GumGum Contextual’s sophisticated Computer Vision machine learning can identify threatening scenes, such as natural disasters or accidents. Object detection picks out potentially threatening objects within an image, such as weapons, exposed skin or drinks.
GumGum Contextual detects brand safety threats for each of the following categories.
Violence and gore
Criminal
Drugs and alcohol
Sexually charged
Profanity and vulgarity
Hate speech, harassment, and cyberbullying
Disasters
Malware and phishing
Medical
These categories align with GARM’s Brand Safety Floor and Brand Suitability Framework.
Clients can set a unique threshold or risk-tolerance level for each threat category. For example, a healthcare provider may choose to set no threshold for the “Medical” threat category, yet higher thresholds for categories that are less suitable for ad placement (e.g., “Hate”, “Violence”, or “Obscene”).
Content Classification
GumGum Contextual works by applying machine learning techniques to relevant content to assign contextual categories.
IAB Categories
The Interactive Advertising Bureau (IAB) defines a Content Taxonomy to provide publishers with a consistent and easy way to organize their website content, and enable advertisers to target standard content categories. GumGum Contextual returns all IAB hierarchy tiers for versions 1.0, 2.0 and 3.0 of the taxonomy:
IAB V1 – 2 tiers - 372 categories
IAB V2 – 4 tiers - 698 categories
IAB V3 - 4 tiers - 709 categories
For example, GumGum Contextual analysis of an article on “The Rise of Alternative Venture Capital” identifies IAB v1.0 categories in 2 tiers, and IAB v2.0 and v3.0 categories in 4 tiers.
Keywords
Keywords are derived from content, metadata, and headlines. GumGum Contextual ranks keywords according to frequency of use and prominence. Objects detected in an image may be included in the list of keywords.
GumGum Contextual Classification and Brand Safety Report
The GumGum Contextual report includes complete brand safety, keyword, and categorization analysis data for the requested content. Each report contains the following analysis results:
dataAvailable | States whether the classification request has already been processed. If processed data exists, GumGum Contextual returns the results from the database. If not GumGum Contextual starts a new processing request. |
---|---|
status | The current processing status of the analysis request. |
pageUrl | The URL of the page, video, image, or text analyzed by GumGum Contextual, as applicable. |
uuid | A unique identifier generated for the classification request. |
languageCode | The standard ISO 639-1 code for the language of the content. Refer to the Language Support Grid for the latest supported languages. Note: If GumGum Contextual detects an unsupported language, a status of NOT_SUPPORTED is returned. |
iab | IAB contextual categories are defined in the IAB Content Taxonomy and are widely adopted in programmatic and Real-Time-Bidding (RTB) ad marketplaces. GumGum Contextual supports current versions of the IAB Content Taxonomy. The GumGum Contextual team keeps track of new taxonomy releases and implements updates in a timely fashion. Refer to the GumGum Contextual Taxonomy document for a listing of IAB contextual categories. |
keywords | The top Keywords identified for the content, listed in order of prominence. |
safe | The final aggregated Brand Safety summary result for the content. If any threat classifications are identified with a risk level of HIGH, the safe value is false and the content is considered unsafe. If no (or low-risk) threat classifications are identified, the safe value is true, and the content is considered safe. |
threats | Threat categories are part of GumGum’s brand safety taxonomy. GumGum classifies content into nine threat categories. For a complete list of Threat category IDs and Names, refer to Threat Categories in the GumGum Contextual Taxonomy document. To detect possible threats, GumGum Contextual analyzes and scores all the extracted content. GumGum Contextual then correlates the scores to determine a per-category threat risk-level for the content. Possible threat category risk-levels are:
|
processedAt | The date and time of the classification. |
Classification Approaches
GumGum Contextual analyses threat, contextual categories, keywords and sentiment results in different ways. The data models GumGum Contextual implements vary for different purposes and are fine-tuned and optimized on an ongoing basis.
Partners should be aware that, as with any machine learning technology, performance is highly dependent on the specific data set being analyzed, consequently no single error rate nor range exists. GumGum Contextual handles proprietary data sets and cannot disclose proprietary partner result data.
GumGum Contextual calculates and measures error rates in the form of Precision, Recall, F1, and F2 for each machine learning model. As part of this process, GumGum:
Engages data annotation leveraging human-annotators to establish Ground Truth for various data sets.
Works with third-party vendors and research consultants to conduct relevancy testing.
Note: If a GumGum Contextual data set that has been delivered to a partner is deemed erroneous or incomplete, GumGum will follow the GumGum Contextual Data Reissuance Policy.
The following sections outline the data models and scoring used for Brand Safety and Contextual Classification in GumGum Contextual, and points to a relevant third-party study.
Brand Safety Classification and Scoring
GumGum Contextual’s brand safety classification relies on GumGum’s threat data model. The threat model is trained on collections of various kinds of threatening content.
As brand safety and content classification serve different purposes, GumGum Contextual considers different approaches for scoring brand safety versus content classification models. Both approaches use Recall scoring (e.g. out of all the images of weapons in a dataset, how many weapons were identified) and Precision scoring (e.g. the number of times an image identified as a weapon was actually a weapon).
Brand safety is a threat detection algorithm, so in this case GumGum Contextual favors Recall over Precision. Data Scientists use Precision-Recall curves to maximize Recall with minimum loss in Precision, thereby maximizing the number of potential threats classified.
GumGum Contextual results comprise risk and confidence levels for each Threat category.
The risk level represents the risk potential of unsafe content within a page, video, image, or text string. Possible risk levels are LOW, MEDIUM and HIGH.
In traditional statistical measures, confidence in observed results may be assessed according to the number of samples involved in a test. Larger scale sampling leads to a higher confidence score. However, GumGum Contextual confidence levels are not related to the quantity of sample data. For example:
A threat category result “confidence”: “VERY_LOW” should be interpreted as GumGum Contextual identifying a very low risk for that category within the content, with a high level of confidence.
A threat category result “confidence”: “VERY_HIGH” should be interpreted as GumGum Contextual identifying a very high level risk for that category within the content, with a high level of confidence.
Contextual Classification and Scoring
GumGum Contextual analyses contextual categories, keywords and sentiment results using various methods and data models, outlined in the following table:
IAB Content Categories | Content classifier predicts the likelihood that the given content belongs to one or more IAB categories. |
---|---|
Keywords | A set of rules derives, scores, and ranks the most important keywords. |
Sentiments | Machine learning predicts the sentiment of each sentence within content by applying models trained on content with varying tones of voice. GumGum Contextual returns an aggregated breakdown of the proportion of sentences in the content that are positive, neutral or negative (referred to as Document Level Sentiment Analysis). There are inherent accuracy limitations for sentiment reporting, as this varies by data set, largely due to the subjective nature of the classification task. Our studies have shown that Neutral is typically the highest scoring sentiment value for documents analyzed. |
Content classification is used for targeting purposes so GumGum Contextual favors Precision over Recall. Data Scientists use Precision-Recall curves to maximize Precision with minimum loss in Recall, thereby maximizing the accuracy of the classified targets.
Contextual Intelligence Relevancy Study
GumGum participates in publicly available third-party media studies, such as the Comparison of Contextual Intelligence Vendors and Behavioral Targeting undertaken with the Dentsu Aegis Network in 2020. The study report found that:
GumGum GumGum Contextual™ had the highest percentage of relevant pages across all four Contextual Intelligence vendors.
Partners may review the complete report, available from this link Understanding Contextual Relevance and Efficiency.
GumGum Contextual and the GARM Brand Safety Floor
Integration Methods
GumGum Contextual integration clients include publishers who can sell ad space directly to advertisers, using GumGum Contextual data to place ads with contextually targeted content, or to avoid brand-unsafe content.
GumGum Contextual client integrations also include video implementations, such as a Contextual Video Marketplace where brands and advertisers can access GumGum Contextual’s contextual and brand-safety data for the marketplace publishers’ video inventory.
Clients leverage GumGum Contextual data via RESTful API or Page Tag integration. In both cases, GumGum Contextual analysis results are returned in a JSON response body.
API Integration
GumGum Contextual offers separate APIs for Page and Video Analysis via server-to-server (S2S) connections. In either case a user or client application calls the GumGum Contextual API, specifying the URL of content to be analyzed. Clients implement webhooks to listen for the JSON response body results on a GumGum Contextual callback URL.
Page Tags
In this case, publishers implement a page tag that automatically calls GumGum Contextual to analyze a page whenever a user visits the page.
For example, a publisher could set up a page tag to fetch new ads for the page based on the keywords identified by GumGum Contextual. Initial ad loading is disabled until GumGum Contextual returns the keyword data. A callback publishes targeting keywords using the GumGum Contextual data, then fetches new ads via Google publisher Tag refresh functionality.
Processing Time
Once a request is sent, GumGum Contextual takes less than a second to return an initial response, indicating whether or not data is already available for the URL.
If data is available (i.e. the content has been processed recently and results are in the database) the GumGum Contextual response is returned immediately.
If the request is for new digital content, GumGum Contextual initiates an asynchronous process to analyze the content and correlate the results into a GumGum Contextual response. It may take a few minutes to complete processing for new media.
Machine Learning Model Development
The GumGum Contextual team carefully selects and trains machine learning models for each contextual and brand-safety classification. As part of the normal GumGum Contextual lifecycle, existing models are continually enhanced or seamlessly replaced with higher-performing models.
GumGum develops machine learning models and also works with technology partners in various ways. GumGum: