Digital Commons Network^™

Automatically Determining Missing Product Attributes Using A Large Language Model, Gokul Pillai, Sandaleen Taqdees, Rishabh Agrawal, Alex Stillwell, Joey Chinoski, Sree Meruva

Defensive Publications Series

To surface products that are relevant to customers, search engines require online merchants to fill in certain attributes relating to their products. However, merchants often do not fill in the requested attributes, which can lead to search results not surfacing the most appropriate products and merchants. This disclosure describes techniques that leverage a large language model (LLM) to automatically determine and assign attributes to products in the inventory of a merchant. Per the techniques, available product information such as product title, category, image, brief description, etc. is used to predict with high probability the values of unfilled attributes. Attributes thus …

Using Baseline Feedback Rates To Dynamically Adjust Customer Service Alert Thresholds, Taliver Heath, Charles Zha, Arthur Wait

Defensive Publications Series

Clusters of similar customer queries can arise at certain times in the lifecycle of a product or service, e.g., during a software release, during a service outage, etc. At such times, customer support can experience challenges in efficiently escalating emerging issues reported by customers. This disclosure describes techniques that enable rapid and automated categorization of customer issues for a product or service. Using the described techniques, clusters of similar issues can be recognized and the issue escalated as necessary. Machine learning (e.g., large language models) or other categorization techniques are applied to categorize reported customer issues. Categories are ranked into …

Facilitating A Dynamic Relationship Understanding Of Information Technology Infrastructure Through Tomographic Sources For Cognitive Network Operation, Greg Galvan, Steve Holl, Jason Coleman

Defensive Publications Series

Information technology (IT) networks can experience dynamic changes due to a variety of events (e.g., hardware refreshes, configuration changes, feature updates, etc.). The dynamic nature of such networks can make it challenging to understand and document the relationships between devices, and also between different devices and business services that may be provided by the devices. In order to address such challenges, techniques are proposed herein through which a device's generated events (or logs) can be observed and, from content of the events, pertinent information can be extracted in order to build an understanding of how devices are related to understand …

Relevance Determination For Multimedia Posts Using A Multimodal Ranker, Saswati Ray, Huy Thong Nguyen

Defensive Publications Series

Many online platforms permit various entities to post multimedia content, e.g., text together with visual content. For example, merchants in online stores, digital maps, etc. can add merchant posts to their account to highlight their merchandise and special offers. Matching text queries provided by users against only the text portion of the multimedia content fails to take into account the visual portion (image) of the merchant post. This disclosure describes the use of dual encoders that allow matching user query embeddings against embeddings obtained from textual content and embeddings obtained from the visual content of the same multimedia post to …

Method For Approximate Content Discovery, Philippe Ombredanne

Defensive Publications Series

Efficient approximate digital content discovery and identification at scale is an important and complex topic.

This proposed method for matching approximate digital content and text fragments consists of these three elements:

1. Content-defined chunking algorithm to split content into fragments and select a subset of these fragments

2. Fingerprinting fragments with a locality sensitive hashing (LSH) function for approximate matching of these fragments. This fingerprint can embedmultiple precision in a single bit string to tune the search precision and organize the search in rounds of progressively increasing precision.

3. Indexing-time approximate matching for deduplication where each new content fragment is …

Artificial Intelligence Agent For Contextual Guidance In Form Filling, Wicksell Metellus, Sindhura Balireddi, Mike Maelzer, Marim Ganaba, Erika Shiroma

Defensive Publications Series

Online forms in enterprise settings enable collection of information relevant to complex tasks such as putting together sales quotes. Form complexity can lead to errors, confusion, and/or delays as users that attempt to fill out the forms struggle to interpret the field meanings and navigate intricate processes. This disclosure describes techniques that build an understanding of forms using semantic and structural analysis, build contextual awareness based on profile data for entities completing a form and a dynamic knowledge base, and incorporate adaptive learning for error detection in forms. Techniques such as natural language processing, named entity recognition, question answering engines, …

Shared Interest Representation And Recommendation

Defensive Publications Series

Methods, systems, and storage media for representing and recommending shared interests are disclosed. Exemplary implementations may: determine, for each user of a group of users of a message thread associated with a messaging platform, an interest associated with the user; aggregate the interests associated with each user of the group of users; determine, based on the aggregated interests, a shared interest associated with the group; determine, based on the shared interest associated with the group, a content recommendation for the group; and provide, to at least one user of the group of users, the content recommendation.

Programmable Functional Lattice Structures, Hp Inc

Defensive Publications Series

Lattice structures provide a unique opportunity to engineer material properties, even those not yet exists in solid materials. Through programmable designs, which can be realized using advanced manufacturing techniques such as 3D printing, customer and design specific material properties can be achieved. We present a method for programming lattice structures by adjusting the positions of nodes along beams (referred to as 'offsets'), while maintaining a fixed lattice topology, incorporating a method to accurately predict the programed lattice structures’ property using machine learning-based algorithms. The flexibility of programming lattice structures by node offsets enables the realization of a broad spectrum of …

Exhaust Gas Energy Recovery Via Stirling Engine To Increase The Efficiency Of Combustion Engines, Konrad Hirschbeck

Defensive Publications Series

Use of a stirling engine with generator in the exhaust tract of an internal combustion engine to recover heat energy from the exhaust gases.

Deep Content Understanding For Recipe Guidance Using A Large Language Model, Nagaraj Kota, Varsha Kukreja, Jatin Jindal

Defensive Publications Series

Online recipe searches performed by users are often highly personalized; users seek specific culinary guidance, cooking methods, styles, etc. A large language model (LLM) can be used to generate recipe insights. However, an LLM can generate ungrounded results, and/or miss specific nuances of recipes along dimensions specified by the user. This disclosure describes techniques that leverage a large language model (LLM) to mine, aggregate, and summarize recipe information across multiple pages/sources and to deliver personalized, contextual, and insightful responses to user queries for recipes via a deep understanding of the pages/sources of the recipe. Recipe tips and guidance snippets are …

Redirect-And-Sync For Distributed Stateful Network Applications, Na

Defensive Publications Series

Cloud-based networking uses various stateful networking functions. There is currently no mechanism to synchronize the states of the functions across different binaries. As a result, the functions cannot be used on stateless devices that provide higher throughput or scalability. This disclosure describes techniques to synchronize states in distributed networking systems. The network function state is defined and maintained in a distributed manner and is sharded in units of flows. A shard is assigned an owner, which handles the flows in the shard. State synchronization is achieved by having the shard owner write actions for flows it owns, and by having …

Automation Of Data Center Service Moves, Na

Defensive Publications Series

Service moves are carefully choreographed movements of cloud-based services across data centers (DC). Service moves can equalize DC loads and enable more efficient usage of existing DC capacity through better bin packing. Service moves are laborious, expensive, and time-consuming operations involving intricate coordination between several DC personnel. This disclosure describes techniques to automate and quicken service moves of a data center. By enabling rapid service moves, the techniques can obviate DC builds, reduce purchases of new machines (since services can be moved to grow in a location with available capacity), reduce the buffer stock of machines, etc. The entire computing …

Mixed Reality Environment Selection Using A Virtual Object, Samuel Wu

Defensive Publications Series

A computer-implemented method comprising providing, within a first mixed reality environment, an environment selection user interface element; receiving, using the environment selection user interface element, an environment selection; and generating, according to the environment selection, a second mixed reality environment.

Mixed Reality Environment Time Selection Using A Wrist Device, Samuel Wu

Defensive Publications Series

A computer-implemented method comprising: providing a time selection user interface element; receiving, using the time selection user interface element, a time selection; and generating, according to the time selection, a past state of a mixed reality environment, the past state depicting the mixed reality environment at a previous time corresponding to the time selection.

Automatic Tool Selection To Reduce Large Language Model Latency, Derik Clive Robert, Ankur Garg, Jitesh Vinod Punjabi, Pavan Kumar Reddy M, Nihal Sandeep Balani, Rushin Shah

Defensive Publications Series

Large language models (LLMs) can interact with external tools to retrieve real-time information, make reservations, etc. While complete tool specifications for the external tools can be included in the prompt of an LLM, doing so increases context size and latency. This disclosure describes techniques that use small tool-selector models alongside large language models to select tools that are appropriate to the task specified in the prompt. The techniques preserve prompt space by ensuring that the LLM accesses the full specifications of a relevant tool only when necessary. The techniques enable an LLM to handle complex requests, such as multi-step tasks …

Times Synchronization Of Media And Transcript, Dongeek Shin

Defensive Publications Series

A media item and a transcript of the media item is received. A plurality of media tokens is generated based on the media item. A plurality of transcript item tokens is generated based on the transcript. A subset of the plurality of media tokens and a subset of the plurality of transcript tokens is inputted on a rolling basis into a transformer model. A sequence of probability distribution for the subset of the plurality of transcript token is obtained from the transformer model. With each submission, a transcript token is selected from the subset of the plurality of transcript tokens …

A Large Scale, Natural Language-Based Universal Predictive Model, Cipta Herwana, Shivangi Tomar, Sandeep Subramanian, Utkarsh Lath

Defensive Publications Series

Task-specific predictive machine learning models are used in many contexts. Such models require selection of input features and training of the model for a particular task with specific parameters. Building and maintaining such models is expensive and not scalable. This disclosure describes a universal predictive model (UPM) that is implemented using a dual encoder large language model (LLM). Both input features and task specifications are represented as natural language text or multimodal fragments that can be converted into embeddings by the LLM. A prediction model generates an output based on the feature embeddings and a task embedding. Fragments can be …

Entity-Aware Joint Translation Of Query And Semantic Parse, Sarthak Jauhari, Saisuresh Krishnakumaran, Dinkar Vasudevan, Rahul Goel

Defensive Publications Series

Machine translation is widely utilized to translate text between language pairs. However, the accurate recognition and translation of the entities within input text remains a challenge. Current translation models often struggle to identify entities, leading to literal and incorrect translations in queries. This disclosure describes techniques, referred to as entity-aware (EaT) that enable machine translation models to recognize entities in source text, resulting in accurate and idiomatic translations. Per the techniques, joint translation of an input query and its associated semantic parse is performed. Semantically parsed information is used as supplementary data in a manner that enables the translation model …

Dynamic Pricing System For Recommerce Transaction, Madhusudan St

Defensive Publications Series

The present disclosure related to a technique for real-time dynamic pricing recommendations. The technique includes fetching historical data and transactional information related to the plurality of products and corresponding similar products based on queries. The technique further include calculating optimal pricing recommendations based on the fetched historical data and transactional information. Subsequently, the technique include distributing price recommendations for recommerce transaction to the plurality of merchants

Assessing Temporal Video Quality Of Video Conference, Björn Sundman, Alex Bau

Defensive Publications Series

A frame identification system and method that receives a source video and generates, for each frame of the source video, a graphical symbol. The aforementioned graphical symbol is resistant to video artifacting and minimally affects the quality of the transmitted media. frame identification system generates, using the graphical symbol for each frame of the source video, a watermarked video. The frame identification system receives, from an image capturing device, a plurality of provisional frame IDs of the watermarked video. Each frame ID of the plurality of provisional frame IDs corresponds to a frame of the watermarked video captured by the …

A Unified Multimodal Video Conference Annotator For Real-Time Llm Applications, Yu-Chung Hsiao

Defensive Publications Series

A multimodal annotator is proposed herein that provides multimodal signals including slide presentations, facial expressions, speech tones, sentiment analysis, and join/leave metadata in addition to ordinary text meeting transcripts. The multimodal annotator may enable real-time multimodal Large Language Model (LLM) applications and can be implemented as a drop-in replacement for generating text transcripts in existing LLM applications. Thus, the multimodal annotator as proposed herein can enable more advanced, multimodal artificial intelligence (AI) assistant features for that can be used across a variety of use cases, such as for video conference use cases, call center use cases, presentation video recordings, or …

System, Method, And Computer Program Product For Mitigating Generative Machine Learning Model Hallucination, Sunpreet Singh Arora, Minghua Xu, Anderson Clayton Alves Nascimento, Yan Zhai, Yizhen Wang

Defensive Publications Series

Systems, methods, and computer program products are provided for mitigating generative model such as a large language model (LLM) hallucination. The system includes a processor configured to receive a query and input the query to a first generative model. The processor is also configured to determine candidate responses based on an output of the first generative model, generate embeddings based on the query, and retrieve data from an embedding-indexed data store based on the embeddings. The processor is further configured to input the data to a second generative model, generate a summary based on an output of the second generative …

Efficient Use Of Error Statistics Of Polar Codes

Defensive Publications Series

Current Forward Error Correction (FEC) modulation schemes struggle to meet the burgeoning requirements of data center interconnects. As the demand for higher bandwidth and faster data speeds escalates, there is a critical need for innovative transmission methods to ensure reliable and efficient data communication.

Multi-Modal Contextual Search For Enterprise Content, Adam Sabra, Rekha Gopikrishnan, Altanai Bisht, Hazim Dahir, Karthik Babu Harichandra Babu, Savita Kini

Defensive Publications Series

A multi-modal contextual search system is proposed herein that may revolutionize multimedia content retrieval in collaborative environments. Unlike traditional keyword-based searches, the proposed system employs advanced artificial intelligence (AI) techniques to directly search within audio and video files in order to retrieve content, overcoming transcription limitations. By leveraging neural-encoder inspired embedding systems and AI-powered audio search without transcription, the proposed system ensures swift and accurate access to crucial information, enhancing productivity and knowledge dissemination. With enhanced features such as audio metadata extraction and semantic matching, the innovative solution proposed herein offers a comprehensive and contextually relevant search experience tailored for …

Transformer Based Sensitive Data Detecting Systems And Methods, Debdeep Bandyopadhyay, Alok Roy

Defensive Publications Series

The present disclosure relates to detection of sensitive data in digital platforms. More particularly, but not exclusively, the present disclosure relates to transformer based sensitive data detecting systems and methods for identifying and classifying the sensitive data in the digital platforms. The transformer based sensitive data detecting system may: extract text in a diverse dataset. The system may identify and classify sensitive data in the diverse dataset and generate an alert only if the identified data being sensitive. Thus, the transformer-based system may identify and classify the sensitive data to address security challenges associated with protecting the sensitive data across …

On-Device Sensitive Content Filtering For Social Media Upload

Defensive Publications Series

A social media platform refers to a website or software application that allows users to share content with each other. Users of a social media platform typically generate content to share, such as still images, audio, or video (often including audio), on a mobile device, because mobile devices typically include a suitable camera, a microphone, a connection to a communications network such as the internet, and can execute a software application that uploads content to the social media platform.

Automated Travel Sharing Feature For A Social Media Platform, Arun Manickam

Defensive Publications Series

A social media platform refers to a website or software application that allows users to share content with each other. A friend on a social media platform refers to a connection of the user to another user. Some presently available social media platforms include a manual check-in, in which a user can share a current location, a photo or video of where the user is or what the user is currently doing, or future travel plans with others (e.g., direct connections, or a larger group, at the user’s discretion). However, users often want to find travel companions, or friends to …

Graphql Fine-Grained Access Control And Query Optimization, Kumud Singh, Tanmay Batham, Amit Kumar Sinha, Mohit Kumar, Amit Rohatgi, Kavitesh Kamboj

Defensive Publications Series

The following description outlines fine-grained authorization along with optimization for both GraphQL and database queries.

Authorization:

• Node-level and field-level access control.

Query Optimization:

• Optimizing GraphQL queries with node-level and field-level authorization.
• Enhancing database query efficiency by resolving the N+1 problem.

Production System Outage Forecasting For Software Services, N/A

Defensive Publications Series

This disclosure describes techniques for forecasting production system outages for software services. Per techniques of this disclosure, a forecast of the total (aggregate) volume of outages across all types of outages is determined for different intervals. A forecast of a relative percentage of each type of outage (based on a root cause of the outage) is determined for the different intervals. A forecast of an outage caused by a particular root cause is determined by calculating the product of the forecasted total outage and the forecasted relative percentage for the particular root cause. Data characteristics analysis is performed to heuristically …

Modifying Segment Routing Flex Algorithm Definition (Fad) Without Causing Routing Loops In Traffic Engineering Networks, Anonymous

Defensive Publications Series

The present disclosure relates to facilitating loop-free Flex Algorithm Definition (FAD) modifications in Segment Routing networks. The proposed solution enables a seamless transition of traffic from a path computed using an old FAD to a path computed using a new FAD, while avoiding loops in the network. This is achieved by utilizing a set of Backup Flex Algorithm (FA) Prefix Segment Identifiers (SIDs) and a set of timers that control the transition between the old FA PFX SIDs and the Backup FA PFX SIDs. This approach allows FAD modifications without causing network loops or traffic loss.