Papers by Benjamin Towle
End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Reply suggestion systems are poorly suited for out-of-the-box retrieval architectures, which only consider individual message-reply similarity. |
| Approach: | They propose an autoregressive text-to-text retrieval model that learns the smart reply task end-to end from a dataset of (message, reply set) pairs obtained via bootstrapping. |
| Outcome: | The proposed approach outperforms state-of-the-art methods on three datasets and shows that it is more diverse and relevant to the user. |
Model-Based Simulation for Optimising Smart Reply (2023.acl-long)
Copied to clipboard
| Challenge: | Existing methods to learn to predict responses to messages are based on post-hoc diversification rather than learning to predict sets of responses. |
| Approach: | They propose a method that employs model-based simulation to discover high-value response sets by simulating possible user responses with a learned world model. |
| Outcome: | Empirically, the proposed method improves ROUGE score and Self-ROUGE scores on two public datasets compared to baselines. |
Enhancing AI Assisted Writing with One-Shot Implicit Negative Feedback (2024.emnlp-main)
Copied to clipboard
| Challenge: | Various systems have been proposed to draft and automate replies for users . yet, the heterogeneity of the inputs and architectures often renders it difficult to combine insights from user behaviour in one system to improve performance in another. |
| Approach: | They propose an approach that uses classifier guidance to controllably integrate implicit user feedback into the text generation process. |
| Outcome: | The proposed approach improves Rouge-L, generating the correct intent and generating an 86% win-rate on the multiWOZ and Schema-Guided Dialog datasets. |
Learn What Is Possible, Then Choose What Is Best: Disentangling One-To-Many Relations in Language Through Text-based Games (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Language models pre-trained on large self-supervised corpora, followed by task-specific fine-tuning has become the dominant paradigm in NLP. |
| Approach: | They propose to train language models pre-trained on large self-supervised corpora, followed by task-specific fine-tuning on the target domain. |
| Outcome: | The proposed model improves on the previous state-of-the-art model on the Jericho Walkthroughs dataset by 49%. |