Learning human beliefs with language models
Measuring explicitly expressed beliefs, whether stated on online platforms or captured in public opinion polls, can shed insight into present and future societal patterns of behavior. Mass media and social media both reflect and help form public opinion, which can ultimately lead to positive outcome...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2022
|
| Online Access: | https://hdl.handle.net/1721.1/143376 |
| _version_ | 1826193149549084672 |
|---|---|
| author | Chu, Eric |
| author2 | Roy, Deb K. |
| author_facet | Roy, Deb K. Chu, Eric |
| author_sort | Chu, Eric |
| collection | MIT |
| description | Measuring explicitly expressed beliefs, whether stated on online platforms or captured in public opinion polls, can shed insight into present and future societal patterns of behavior. Mass media and social media both reflect and help form public opinion, which can ultimately lead to positive outcomes such as civil engagement across divides, but also negative outcomes such as non-adherence to health-beneficial social guidelines. Understanding viewpoints expressed online and the relationship between media content and beliefs is increasingly pertinent today, in a world of constant connectivity and shrinking common ground.
This dissertation introduces new deep, neural language model -based approaches for capturing beliefs reflected in and formed by media. In part one of this dissertation, we introduce a model for automatically summarizing multiple documents about the same subject, which we apply to opinionated posts found on popular review websites. Summaries can help organize large amounts of often siloed information, and help people understand the most salient viewpoints from people in different communities. In contrast to typical approaches that require large, labeled datasets, our method is the first unsupervised model for abstractive multi-document summarization.
In part two of the dissertation, motivated by the effects of information in the COVID-19 pandemic, we introduce an approach for using “media diet models”, which can act as proxies for human media consumption. By probing these models, we can predict public opinion as measured by nationally representative surveys. We validate our method in two domains: attitudes towards COVID-19 and consumer confidence, and show that the approach is valid and intuitive in a number of ways. We find that it is robust and has predictive power across mediums and outlets, has increased predictive power when people are paying more attention to news, and may capture duration effects of media consumption. These results both provide insight into a driving force of human belief formation and suggest practical implications for pollsters, public health officials, and policymakers moving forward. |
| first_indexed | 2024-09-23T09:34:29Z |
| format | Thesis |
| id | mit-1721.1/143376 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2024-09-23T09:34:29Z |
| publishDate | 2022 |
| publisher | Massachusetts Institute of Technology |
| record_format | dspace |
| spelling | mit-1721.1/1433762022-06-16T03:48:57Z Learning human beliefs with language models Chu, Eric Roy, Deb K. Program in Media Arts and Sciences (Massachusetts Institute of Technology) Measuring explicitly expressed beliefs, whether stated on online platforms or captured in public opinion polls, can shed insight into present and future societal patterns of behavior. Mass media and social media both reflect and help form public opinion, which can ultimately lead to positive outcomes such as civil engagement across divides, but also negative outcomes such as non-adherence to health-beneficial social guidelines. Understanding viewpoints expressed online and the relationship between media content and beliefs is increasingly pertinent today, in a world of constant connectivity and shrinking common ground. This dissertation introduces new deep, neural language model -based approaches for capturing beliefs reflected in and formed by media. In part one of this dissertation, we introduce a model for automatically summarizing multiple documents about the same subject, which we apply to opinionated posts found on popular review websites. Summaries can help organize large amounts of often siloed information, and help people understand the most salient viewpoints from people in different communities. In contrast to typical approaches that require large, labeled datasets, our method is the first unsupervised model for abstractive multi-document summarization. In part two of the dissertation, motivated by the effects of information in the COVID-19 pandemic, we introduce an approach for using “media diet models”, which can act as proxies for human media consumption. By probing these models, we can predict public opinion as measured by nationally representative surveys. We validate our method in two domains: attitudes towards COVID-19 and consumer confidence, and show that the approach is valid and intuitive in a number of ways. We find that it is robust and has predictive power across mediums and outlets, has increased predictive power when people are paying more attention to news, and may capture duration effects of media consumption. These results both provide insight into a driving force of human belief formation and suggest practical implications for pollsters, public health officials, and policymakers moving forward. Ph.D. 2022-06-15T13:16:23Z 2022-06-15T13:16:23Z 2022-02 2022-02-27T16:55:33.867Z Thesis https://hdl.handle.net/1721.1/143376 In Copyright - Educational Use Permitted Copyright MIT http://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology |
| spellingShingle | Chu, Eric Learning human beliefs with language models |
| title | Learning human beliefs with language models |
| title_full | Learning human beliefs with language models |
| title_fullStr | Learning human beliefs with language models |
| title_full_unstemmed | Learning human beliefs with language models |
| title_short | Learning human beliefs with language models |
| title_sort | learning human beliefs with language models |
| url | https://hdl.handle.net/1721.1/143376 |
| work_keys_str_mv | AT chueric learninghumanbeliefswithlanguagemodels |