Papers by Yoav Artzi

22 papers
Executing Instructions in Situated Collaborative Interactions (D19-1)

Copied to clipboard

Challenge: a collaborative game with natural language instruction allows users to adapt to the system abilities by changing their language or deciding to accomplish tasks themselves.
Approach: They propose a collaborative game where a user instructs a system to complete tasks, but acts alongside it.
Outcome: The proposed game allows users to adapt to the system abilities by changing their language or deciding to accomplish tasks themselves.
Simple Recurrent Units for Highly Parallelizable Recurrence (D18-1)

Copied to clipboard

Challenge: recurrent neural networks scale poorly due to the intrinsic difficulty in parallelizing their state computations.
Approach: They propose a simple recurrent unit that provides expressive recurrence and allows highly parallel implementation.
Outcome: The proposed model achieves 5—9x speed-up over cuDNN-optimized LSTM on classification and question answering datasets and delivers stronger results than LS and convolutional models.
lilGym: Natural Language Visual Reasoning with Reinforcement Learning (2023.acl-long)

Copied to clipboard

Challenge: Existing approaches to language-conditioned reinforcement learning in visual environments are limited by language semantics.
Approach: They propose a new benchmark for language-conditioned reinforcement learning in visual environments . they annotate 2,661 highly-compositional human-written natural language statements .
Outcome: The proposed approach is based on 2,661 highly-compositional human-written natural language statements grounded in an interactive visual environment.
Continual Learning for Grounded Instruction Generation by Observing Human Following Behavior (2021.tacl-1)

Copied to clipboard

Challenge: Despite its potential and prevalence, this signal is understudied for learning to generate natural language.
Approach: They propose to use this signal to improve the system's ability to generate instructions via contextual bandit learning.
Outcome: The proposed system improves its ability to generate natural language through interaction with users, and the results are shown.
CoGen: Learning from Feedback with Coupled Comprehension and Generation (2024.emnlp-main)

Copied to clipboard

Challenge: Existing studies on coupling comprehension and generation in computational systems show that the ability to finish incomplete partner utterances in dialogue is closely related to comprehension and vice versa.
Approach: They propose techniques to tightly integrate comprehension and generation capabilities with focus on continually learning from interaction with users.
Outcome: The proposed models improve performance by 26% and 17% over time, while the non-coupled system is more human-like.
Crowdsourcing Beyond Annotation: Case Studies in Benchmark Data Collection (2021.emnlp-tutorials)

Copied to clipboard

Challenge: Developing a theory of crowdsourcing for practical language problems remains an open challenge .
Approach: This tutorial exposes NLP researchers to data collection crowdsourcing methods and principles through case studies.
Outcome: This tutorial exposes NLP researchers to various data collection crowdsourcing methods and practices through case studies.
Situated Mapping of Sequential Instructions to Actions with Single-step Reward Observation (P18-1)

Copied to clipboard

Challenge: Existing approaches to map context-dependent sequential instructions to actions are based on discourse and state dependencies . we evaluate on SCONE domains and show absolute accuracy improvements of 9.8%-25.3% .
Approach: They propose a model that considers previous utterances and the state of the world to map sequential instructions to actions.
Outcome: The proposed model improves on the SCONE domains and on the target domains.
Newsroom: A Dataset of 1.3 Million Summaries with Diverse Extractive Strategies (N18-1)

Copied to clipboard

Challenge: a dataset of 1.3 million articles and summaries written by authors and editors in newsrooms of 38 major news publications . identifying large, high-quality resources for summarization has called for creative solutions in the past.
Approach: They present a summarization dataset of 1.3 million articles and summaries written by newsrooms of 38 major news publications.
Outcome: The summarization dataset shows high diversity of summarizing styles . authors train existing methods on the data to evaluate its utility and challenges.
Abstract Visual Reasoning with Tangram Shapes (2022.emnlp-main)

Copied to clipboard

Challenge: We use tangrams as stimuli in cognitive science to study abstract visual reasoning . pre-trained weights demonstrate limited abstract reasoning, we observe .
Approach: They propose a resource for studying abstract visual reasoning in humans and machines . they use tangram puzzles as stimuli to create an annotated dataset with >1k distinct stimuli .
Outcome: The proposed resource is visually and linguistically richer than previous resources . pre-trained weights demonstrate limited abstract reasoning, the authors note .
Simulating Bandit Learning from User Feedback for Extractive Question Answering (2022.acl-long)

Copied to clipboard

Challenge: Explicit feedback from users can be used to continually improve system performance.
Approach: They study the potential of learning from user feedback for extractive question answering by simulating feedback using supervised data.
Outcome: The proposed model improves on a few examples and can be deployed in new domains without any data annotation effort.
Success and Cost Elicit Convention Formation for Efficient Communication (2026.acl-long)

Copied to clipboard

Challenge: ad hoc conventions allow people to coordinate on short, less costly utterances that are understood using shared conversational context.
Approach: They propose a method to train large multimodal models to form conventions . they use simulated reference games to produce training data .
Outcome: The proposed method reduces message length by up to 41% while increasing success by 15% over the course of the interaction.
Continually Improving Extractive QA via Human Feedback (2023.emnlp-main)

Copied to clipboard

Challenge: a study of extractive question answering systems using human feedback shows promising potential for continual learning.
Approach: They study extractive question answering system by using user feedback to improve it . they design and deploy an iterative approach where users ask questions and provide feedback .
Outcome: The proposed model improves over time across different data regimes and domains . human user feedback is more affordable and abundant than annotations provided by trained experts .
Interactive Classification by Asking Informative Questions (2020.acl-main)

Copied to clipboard

Challenge: Existing methods for intent classification rely on a single user input and do not interact with the user to reduce ambiguity and improve the final prediction.
Approach: They propose a limited form of interaction to natural language intent classification . they add binary or multi-choice questions to the system to ask missing information .
Outcome: The proposed method can be bootstrapped without interaction data and is scalable to two domains.
Evaluating Models’ Local Decision Boundaries via Contrast Sets (2020.findings-emnlp)

Copied to clipboard

Challenge: Standard test sets for supervised learning evaluate in-distribution generalization but are misleading when a dataset has systematic gaps.
Approach: They propose a more rigorous annotation paradigm for NLP that helps to close systematic gaps in the test data.
Outcome: The proposed model performs significantly lower on contrast sets than on the original test sets—up to 25% in some cases.
Mapping Instructions to Actions in 3D Environments with Visual Goal Prediction (D18-1)

Copied to clipboard

Challenge: Existing models that map from inputs to actions are inefficient and require hand-crafted meaning representations.
Approach: They propose to decompose instruction execution to goal prediction and action generation . they introduce two benchmarks for instruction following: LANI and CHAI .
Outcome: The proposed model decomposes instruction execution to goal prediction and action generation.
Analysis of Language Change in Collaborative Instruction Following (2021.findings-emnlp)

Copied to clipboard

Challenge: Prior work has found that language complexity is reduced along multiple dimensions as conventions are formed.
Approach: They analyze language change over time in a collaborative task where utility-maximizing participants form conventions and increase their expertise.
Outcome: The study shows that instructors increase language complexity along dimensions to collaborate with skill followers.
CB2: Collaborative Natural Language Interaction Research Platform (2023.acl-demo)

Copied to clipboard

Challenge: CB2 is a platform to study collaborative grounded natural language interactions in task-oriented scenarios . it includes a 3D game environment, backend server, tools and processes to enable scalable studies.
Approach: They propose a multi-agent platform to study collaborative natural language interactions in a task-oriented scenario.
Outcome: The proposed model is a demonstration of a collaborative natural language agent . it is based on the CEREALBAR scenario, but is scalable to emphasize accessibility .
Learning to Map Context-Dependent Sentences to Executable Formal Queries (N18-1)

Copied to clipboard

Challenge: Existing models that map utterances to executable queries are context-dependent and can incorporate interaction history.
Approach: They propose a context-dependent model that maps utterances to executable queries . their approach combines implicit and explicit modeling of references between utterations .
Outcome: The proposed model can map utterances to executable queries based on interaction history . key to mapping utterrances to queries is resolving references .
A Corpus for Reasoning about Natural Language Grounded in Photographs (P19-1)

Copied to clipboard

Challenge: a dataset for visual reasoning with natural language and images is available.
Approach: They propose a dataset for joint reasoning about natural language and images . they crowdsource 107,292 examples of English sentences paired with web photographs .
Outcome: The proposed dataset combines 107,292 examples of English sentences with web photographs . Qualitative analysis shows the data requires compositional joint reasoning .
When in Doubt: Improving Classification Performance with Alternating Normalization (2021.findings-emnlp)

Copied to clipboard

Challenge: a classifier that uses a nonparametric post-processing step for classification suffers when given examples that are close to its decision boundary.
Approach: They propose a nonparametric post-processing step that re-adjusts predicted class probability distributions using high-confidence validation examples.
Outcome: The proposed method improves classifier accuracy on difficult examples.
What is Learned in Visually Grounded Neural Syntax Acquisition (2020.acl-main)

Copied to clipboard

Challenge: Visual features are promising for learning bootstrap textual models, but blackbox learning models make it difficult to isolate the specific contribution of visual components.
Approach: They propose to use alignments between phrases and images as a learning signal for syntax acquisition.
Outcome: The proposed model performs better than the previous model, but it is significantly less expressive.
Retrospective Learning from Interactions (2025.acl-long)

Copied to clipboard

Challenge: Multi-turn interactions between large language models and users naturally include implicit feedback signals.
Approach: They propose a method to learn from feedback signals in past interactions without annotations . they use a multimodal LLM to solve a reasoning task with a combinatorial solution space .
Outcome: The proposed method improves task completion rate from 31% to 82% without annotations.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations