Microsoft Corp v. Dialect LLC

1. Case Identification

• Case #: IPR2025-01352
• Patent #: 8,620,659
• Filed: August 5, 2025
• Petitioner(s): Microsoft Corporation
• Patent Owner(s): Dialect, LLC
• Challenged Claims: 42

2. Patent Overview

• Title: Speech-Based Information Query, Processing, and Presentation System
• Brief Description: The ’659 patent discloses a system for processing natural language utterances that uses predictive models to enhance responses. The system employs a "personalized cognitive model" based on an individual user's interaction history and a "general cognitive model" based on data from multiple users to predict a user's next action.

3. Grounds for Unpatentability

Ground 1: Obviousness over 2007 Grandparent Pub. in view of Lee - Claim 42

• Prior Art Relied Upon: 2007 Grandparent Pub. (Application # 2007/0038436) and Lee (Application # 2005/0054381).
• Core Argument for this Ground:
  • Prior Art Mapping: Petitioner argued that claim 42 is not entitled to the filing date of its ancestor applications, making the 2007 Grandparent Pub. prior art under 35 U.S.C. §102(b). Petitioner asserted the 2007 Grandparent Pub. teaches every element of claim 42 except for "determining whether a personalized cognitive model...includes sufficient information for predicting one or more subsequent actions." Lee allegedly teaches this missing element by disclosing a system that predicts a user's next telephone call based on a history window. Lee’s system determines if its prediction model has sufficient information by assessing both the robustness of the data (e.g., if the history window is full) and whether a prediction probability exceeds a threshold (e.g., 50%).
  • Motivation to Combine: A POSITA would combine Lee's teachings with the system of the 2007 Grandparent Pub. because the latter provides only general disclosures about using a personalized cognitive model and switching to a general model when an interaction is "not handled." Lee provides specific, concrete guidance on how to determine when a personalized model contains sufficient data to make a reliable prediction. Implementing Lee's sufficiency determination would improve the usability and performance of the Grandparent Pub.'s system, particularly for its disclosed application of automatic telephone dialing.
  • Expectation of Success: A POSITA would have a reasonable expectation of success in combining the references because both relate to natural language processing on telephonic devices. Lee’s method for determining prediction sufficiency is a well-defined, rule-based algorithm that could be readily incorporated into the model-based framework of the 2007 Grandparent Pub.

Ground 2: Obviousness over Ramaswamy - Claim 42

• Prior Art Relied Upon: Ramaswamy (Patent 6,622,119).

• Core Argument for this Ground:

  • Prior Art Mapping: Petitioner argued that Ramaswamy teaches every element of claim 42. Ramaswamy discloses a command prediction system that uses both a general prediction model (based on a large corpus of user data) and a personal prediction model (based on an individual user's command history). Ramaswamy explicitly teaches determining data sufficiency for the personal model. When a user is new, or the personal model has accumulated insufficient data, Ramaswamy's system assigns the personal model a very low weight (e.g., a coefficient of 0.01) and relies almost exclusively on the general model for predictions. This assignment of weight based on the amount of data constitutes a determination of whether the personal model "includes sufficient information." In response to this determination (i.e., when information is insufficient), Ramaswamy predicts subsequent actions using the heavily weighted general model, which is generated from tracking interactions of a plurality of users.
  • Key Aspects: This ground asserted that Ramaswamy's sliding-scale weighting system inherently performs the claimed "determining" step. Even if the claim requires a binary determination, Petitioner argued it would have been an obvious design choice to use a threshold to switch between models, rather than Ramaswamy's continuous weighting scheme.

• Additional Grounds: Petitioner asserted additional obviousness challenges based on Ramaswamy in view of Ramaswamy NPL (a 1999 journal article) and Ramaswamy in view of Hartono (WO 2000/011571). These grounds were presented as alternatives in case the Patent Owner argued Ramaswamy alone fails to teach the "domain agent" limitation, with Ramaswamy NPL and Hartono allegedly providing supplemental teachings on domain-specific software agents.

4. Key Claim Construction Positions

• "Domain Agent": Petitioner noted that in related litigation, this term was construed as "software with domain-specific behavior and information." Petitioner argued that its invalidity grounds render claim 42 obvious even under this construction, contending that the "command executor" in Ramaswamy, particularly when considered with its incorporated references, meets this definition.

5. Key Technical Contentions (Beyond Claim Construction)

• Lack of Priority Date Entitlement: A central contention of the petition, foundational to Ground 1, was that claim 42 is not entitled to the priority date of its ancestor applications. Petitioner argued the ancestor specifications fail to provide adequate written description or enablement for the limitation of "determining whether a personalized cognitive model...includes sufficient information." The specifications allegedly only describe switching to a general model when an interaction is "not handled" by the personal model, without disclosing what constitutes "sufficient information," how to measure it, or any process for making such a determination.

6. Relief Requested

• Petitioner requests institution of an inter partes review and cancellation of claim 42 of the ’659 patent as unpatentable.