Neural systems for informative conversations
Abstract/Contents
- Abstract
- Humans, through deep and expressive conversations, have perfected the art of exchanging information about the world around them seamlessly. But even with the latest NLP methods, chatbots struggle in being informative. In this dissertation, I describe my work on building neural systems for informative conversations. First, I describe Chirpy Cardinal, our Alexa Prize 2020 Socialbot, that was deployed to tens of thousands of users across the US, and served as a test-bed for an initial system for informative conversations. While we used state-of-the-art models that improved over prior work, they fell short of expectations when deployed in the real-world setting. In particular, our system had two components: a retriever to find conversationally relevant passages from a large corpus (like Wikipedia) and a language generator to weave it into the dialogue with conversational-sounding utterances, and these two components were unable to cohesively work together. Second, inspired by linguistics literature on conversations, I analyze human-human informative conversations and identify various strategies for acknowledgement, presentation, transition and detail-selection. I also present a case study, where I improve acknowledgements by using conditional mutual information to select better chatbot utterances. Third, I explore the possibility of learning these strategies from data by jointly training a neural retriever and a neural generator such that they work together cohesively. To train them, we need to know which passages are relevant to the conversation, but the abundant conversational data available for training is not annotated for relevant passages! Our method, HINDSIGHT, uses a posterior retriever to find relevant passages during training. The posterior retriever is jointly trained alongside the original retriever and the generator using the evidence lower bound (ELBo). We find that HINDSIGHT has better inductive biases than existing methods - at inference, the retriever finds more relevant passages and the generator is more grounded in the retrieved passages, resulting in better end-to-end performance. Together, these projects provide a strong practical motivation, rich linguistic guidance and an effective training method for our aim of building neural systems to have deep and topically broad conversations.
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 | 2022; ©2022 |
| Publication date | 2022; 2022 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Paranjape, Ashwin Pradeep |
|---|---|
| Degree supervisor | Manning, Christopher D |
| Thesis advisor | Manning, Christopher D |
| Thesis advisor | Hashimoto, Tatsunori |
| Thesis advisor | Jurafsky, Dan, 1962- |
| Degree committee member | Hashimoto, Tatsunori |
| Degree committee member | Jurafsky, Dan, 1962- |
| Associated with | Stanford University, Computer Science Department |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Ashwin Paranjape. |
|---|---|
| Note | Submitted to the Computer Science Department. |
| Thesis | Thesis Ph.D. Stanford University 2022. |
| Location | https://purl.stanford.edu/rz437xd9519 |
Access conditions
- Copyright
- © 2022 by Ashwin Pradeep Paranjape
- License
- This work is licensed under a Creative Commons Attribution 3.0 Unported license (CC BY).
Also listed in
Loading usage metrics...