Skip to content
toshish

Research

Papers & patents

A selection of published research and granted patents, with the story behind each. The full list is on Google Scholar.

Papers

Are Human Conversations Special? A Large Language Model Perspective

arXiv:2403.05045 · 2024 · with Chaitanya Animesh, Sekhar Vallath, Kartik Talamadupula & Larry Heck

The question behind most of my career, made measurable. We analyzed how LLM attention behaves on human dialogue versus web text, code, and math: attention distance, dispersion, and entropy across layers. Conversations demand longer-range, higher-complexity attention, and models that clearly specialize for code and math show a measurable gap on human dialogue.

arXiv → Plain-English explainer →

Reading with Intent (Emotional RAG LLMs: Reading Comprehension for the Open Internet)

arXiv:2408.11189 · 2024 · with Benjamin Reichman, Kartik Talamadupula & Larry Heck

RAG benchmarks retrieve tidy, neutral Wikipedia prose. The real internet is sarcastic, emotional, and unreliable. We built emotionally-inflected and sarcastic versions of retrieval corpora, showed how badly they distort reader models, and developed methods that make LLMs interpret tone instead of taking it literally.

arXiv → Code & dataset →

Patents

US 11,403,596 B2 · First filed October 2018 · Granted August 2022

Integrated framework for managing human interactions

Our first filing, from when Symbl was still called Rammer. It lays out the whole pipeline as one system: audio and video come in from wherever conversations happen, live or recorded, get turned into text and then a single unified dataset, and from there it works out the context of each conversation and flags the moments that don't fit it. Those get resolved into pending items, the action items, follow-ups, and questions someone would otherwise track by hand. The later patents each go deeper on one piece of this.

Google Patents →

US 11,093,718 B1 · Granted August 2021

Determining conversational structure from speech

Raw transcripts are a wall of words. This patent covers detecting the structure hiding inside live speech (who is addressing whom, what refers to what) so downstream systems have something to reason over.

Google Patents →

US 11,263,407 B1 · Granted March 2022

Determining topics and action items from conversations

The heart of conversation intelligence: surfacing what a meeting was actually about and what people committed to do, in real time, from messy, multi-party speech, years before LLMs made this look easy.

Google Patents →

US 11,302,314 B1 · Granted April 2022

Tracking specialized concepts, topics, and activities in conversations

Every customer cares about different concepts: their product names, their objections, their compliance phrases. This covers following those custom concepts through live conversations without retraining, by expanding a small tracker vocabulary through a universal model.

Google Patents →

US 11,599,713 B1 · Granted March 2023

Summarizing conversational speech

Summarizing dialogue is nothing like summarizing documents: meaning is co-constructed across turns, corrections, and interruptions. This covers our approach to compressing spoken conversations without losing what mattered.

Google Patents →

Also granted

  • US 11,562,149 B2 · Determining conversational structure from speech (continuation)
  • US 11,593,566 B2 · Determining topics and action items from conversations (continuation)
  • US 11,580,961 B1 · Tracking specialized concepts, topics, and activities in conversations (continuation)
  • US 11,842,144 B1 · Summarizing conversational speech (continuation)
  • US 11,019,004 B1 · Bot engine abstraction, from earlier work at Amdocs