Papers by Yash Kankanampati
Machine-Assisted Script Curation (2021.naacl-demos)
Copied to clipboard
Manuel Ciosici, Joseph Cummings, Mitchell DeHaven, Alex Hedges, Yash Kankanampati, Dong-Ho Lee, Ralph Weischedel, Marjorie Freedman
| Challenge: | Scripts have been of interest for encoding procedural knowledge and understanding stories for over 40 years . narrative descriptions often omit common knowledge . |
| Approach: | They propose a machine-aided script creator that automates script creation with suggestions for event types, links to Wikidata, and sub-events that may have been forgotten. |
| Outcome: | The proposed system automates portions of the script creation process with suggestions for event types, links to Wikidata, and sub-events that may have been forgotten. |
Multitask Easy-First Dependency Parsing: Exploiting Complementarities of Different Dependency Representations (2020.coling-main)
Copied to clipboard
| Challenge: | Existing dependency parsing models for Arabic use complementary annotations, CATiB and UD treebanks, and partially created trees for one annotation are also available to the other as features for the score function. |
| Approach: | They propose to use Arabic dependency annotations to parse projective dependency trees using CATiB and UD treebanks. |
| Outcome: | The proposed model gives 9.9% error reduction on CATiB and 6.1% on UD compared to a strong baseline and ablation tests show that the main contribution is given by sharing tree representation between tasks, and not simply sharing biLSTM layers as is often performed in NLP multitask systems. |
Remember what you did so you know what to do next (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies have shown large language models (LLMs) to be poor fit for a simulated robot to achieve 30 classes of goals. |
| Approach: | They use the 6B parameter GPT-J language model to create a plan for a simulated robot to achieve 30 classes of goals in ScienceWorld. |
| Outcome: | The proposed model outperforms the state-of-the-art by a factor of 1.4 when training on as many prior steps as will fit, and the results are 2.3x better than the state of the-art. |