Contextual grounding in human-computer dialogue
Abstract/Contents
- Abstract
- Context is fundamental to human communication, shaping the meanings of our expressions and enabling the efficient and effective transmission of information. In the sphere of human-machine interaction, contextual grounding is paramount for facilitating interactions that feel as natural and intuitive as human dialogue. This research delves into the mechanisms through which machines can extract contextual information from a variety of sources: a conversational partner, a history of interactions, external world knowledge, and cooperative repair. We begin by introducing a model that optimizes speech with respect to an internal model of their conversational partner. This approach serves as an amortization of recursive social reasoning, where dialogue is shaped by considering a conversational partner's contextual understanding. This approach captures the nuances of pragmatic language use and provides good generalization to new contexts, all while maintaining reasonable processing costs. We extend the exploration of past interactions by shifting focus towards explicit recall of specific information gleaned in earlier dialogue. We introduce the Long Range Common Ground benchmark, a dataset comprised of GPT-generated dialogues, which aids in evaluating methods of maintaining common ground over extended interactions. This work elucidates that existing knowledge retrieval models are not yet proficient at the nuanced task of common ground retrieval. Next, we detail a framework for harnessing external world knowledge for dialogue tasks. Our framework produces a dialogue system augmented with an information retrieval module, capable of extracting pertinent documents from company policies. Our results demonstrate how such knowledge augmentation can enhance performance on action-oriented dialogue tasks by utilizing procedural instructions available in external knowledge databases, particularly in settings with minimal supporting data. Finally, we present a method of enabling dialogue systems to engage in a lightweight, data-efficient method of cooperative repair. We showcase how pre-trained goal-directed agents can generate insightful questions for visually-grounded tasks by using a pre-trained image captioner in conjunction with expected information gain. This approach enables a model to seek further clarification in cases of uncertainty, ultimately contributing to the broader goal of enhancing human-machine communication. In sum, this comprehensive exploration of contextual grounding strategies contributes to the evolving sphere of human-machine communication, aiming to foster more natural and effective interactions between people and technology.
Description
Type of resource | text |
---|---|
Form | electronic resource; remote; computer; online resource |
Extent | 1 online resource. |
Place | California |
Place | [Stanford, California] |
Publisher | [Stanford University] |
Copyright date | 2023; ©2023 |
Publication date | 2023; 2023 |
Issuance | monographic |
Language | English |
Creators/Contributors
Author | White, Julia Isabel |
---|---|
Degree supervisor | Goodman, Noah (Noah D.) |
Thesis advisor | Goodman, Noah (Noah D.) |
Thesis advisor | Finn, Chelsea |
Thesis advisor | Sadigh, Dorsa |
Degree committee member | Finn, Chelsea |
Degree committee member | Sadigh, Dorsa |
Associated with | Stanford University, School of Engineering |
Associated with | Stanford University, Department of Electrical Engineering |
Subjects
Genre | Theses |
---|---|
Genre | Text |
Bibliographic information
Statement of responsibility | Julia White. |
---|---|
Note | Submitted to the Department of Electrical Engineering. |
Thesis | Thesis Ph.D. Stanford University 2023. |
Location | https://purl.stanford.edu/bv657xn0996 |
Access conditions
- Copyright
- © 2023 by Julia Isabel White
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...