Issued System and methods for neural topic modeling using topic attention networksUS Patent App 17/644,856 (granted in 2021-12-17). Assigned to Adobe Inc.
Filed Hyperbolic and Interpretable Hierarchical Topic Models Accepted Internally (Adobe Inc.) in December 2022. Filed Enable marketers increase visitor engagement on B2B websites via GenAI powered conversational systems. Accepted Internally (Adobe Inc.) in June 2023.
Filed A Novel Perspective on Content Personalisation: Tailoring Examples within Content to Resonate. Accepted Internally (Adobe Inc.) in August 2023.
In-Filing A Novel method and apparatus for text-guided style transfer. Accepted Internally (Adobe Inc.) in July 2023.
Submitted Novel Framework for Fair Text Generation through Introspective Prompting. Waiting for Internal Approval (Adobe Inc.)
Submitted Mind the Gap: A Novel Framework to Identify Content Gaps. Waiting for Internal Approval (Adobe Inc.)
Submitted Accurate Summaries Matter: Detecting Factual Inconsistencies generated by LLMs. Waiting for Internal Approval (Adobe Inc.)
Submitted Mask-CLIPstyler: Localized text-based style transfer in images. Waiting for Internal Approval (Adobe Inc.)
Document Recommendation
Language Model Patent Adobe Product
Product 1: Identify similar user queries on the Adobe Support Community platform (similar to Stack Exchange), with the goal of reducing the number of new queries posted for an existing one. We observed a 22% increase in user’s engagement and a 20% decrease in new posts.
Product 2: We designed a workflow to recommend blogs using search signals and different types of user interactions with posts on the Adobe Marketo Nation Community Platform. In the first 8 weeks of launch, we observed a 40% increase in engagement.
Product 3: With this new cutting-edge approach to surfacing content recommendations on Marketing Nation Community, users receive recommended content based on their interactions with Marketo's email programs with the ultimate goal of increasing adoption and Community engagement. We observed a 30% increase in new users’ engagement with the Marketo offerings.
Conversational Systems
Language Models Patent Adobe Product
[1] Sales summary generation: A concise summary for the sales agent with call to action for their leads. A fact-checking algorithm with Natural Language Inference is added to ensure that summaries and action items generated are not hallucinated. We created a golden dataset for benchmarking models on this task.
[2] Answering visitor questions: In this we use Retrieval Augmented Answer Generation (RAG) to find answer for the visitor question.
Routing user queries to experts.
Topic Models Patent Adobe Product
Adobe Support Community is a trusted online platform for Adobe customers to discuss queries about products, similar to Stack Exchange. These forums have millions of queries and it becomes difficult for the experts to find posts that they can answer. This calls for efficient and quick routing of user queries to interested experts. We proposed a routing framework, where we develop expert profiles using topic models and match it with the post topics.
Automatic topic labeling with Hyperbolic Geometry
Topic Models [CODE]
The topics generated by topic models often have words inconsistent with the theme a particular topic is representing, thus, making it difficult to interpret such topics. In this project, I worked on a solution for the automatic labeling of topics using concept hierarchies.
Content Gap with Session History
Session Transformer