Papers by Daniel Smith
How Much Does Attention Actually Attend? Questioning the Importance of Attention in Pretrained Transformers (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Pretrained language models use the attention mechanism to contextualize input inputs . but, we find that it is not as important as thought for pretrained models . |
| Approach: | They propose a probing method that replaces input-dependent attention matrices with constant ones. |
| Outcome: | The proposed method improves performance of pretrained language models without input-dependent attention. |
NarrowBERT: Accelerating Masked Language Model Pretraining and Inference (2023.acl-short)
Copied to clipboard
| Challenge: | Large-scale language model pretraining is expensive as the models and pretraining corpora have become larger over time. |
| Approach: | They propose a modified transformer encoder that increases throughput for masked language model pretraining by more than 2x. |
| Outcome: | The proposed model increases throughput on IMDB and Amazon reviews classification and CoNLL NER tasks by 3.5x with minimal performance degradation. |
The Engage Corpus: A Social Media Dataset for Text-Based Recommender Systems (2022.lrec-1)
Copied to clipboard
| Challenge: | Existing studies have examined the impact of recommendation algorithms on how users discover and join online groups, but there are few standardized datasets for generating such models. |
| Approach: | They propose to use Reddit to build a dataset that can be used to build models of user engagement with online groups. |
| Outcome: | The proposed model is based on the behavior of subreddits banned in June 2020 as part of Reddit's efforts to stop the dissemination of hate speech. |
Time Waits for No One! Analysis and Challenges of Temporal Misalignment (2022.naacl-main)
Copied to clipboard
| Challenge: | a pretrained model is optionally adapted through domain-specific pretraining, followed by task-specific finetuning. |
| Approach: | They establish a suite of eight tasks across different domains to quantify the effects of temporal misalignment in modern NLP systems. |
| Outcome: | The proposed tasks are based on eight domains and periods of time spanning five years or more and show that they have stronger effects than previous studies. |
Self-Instruct: Aligning Language Models with Self-Generated Instructions (2023.acl-long)
Copied to clipboard
Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, Hannaneh Hajishirzi
| Challenge: | Large “instruction-tuned” language models depend heavily on human-written instruction data . this limited quantity, diversity, and creativity hinders the generality of the tuned model . |
| Approach: | They propose a framework for improving instruction-following capabilities of pretrained language models by bootstrapping off their own generations. |
| Outcome: | The proposed framework outperforms existing public instruction datasets by 5% . it generates instructions, input, and output samples, then filters invalid or similar ones . |
Domain Mismatch Doesn’t Always Prevent Cross-lingual Transfer Learning (2022.lrec-1)
Copied to clipboard
| Challenge: | Recent studies have reported that domain mismatch prevents cross-lingual transfer . UBLI and UNMT do not work well when underlying monolingual corpora come from different domains . |
| Approach: | They show that a simple initialization regimen can overcome domain mismatch in cross-lingual transfer . they pre-train word embeddings on concatenated domain-mismatched corpora and use them as initializations . |
| Outcome: | The initialization regimen can overcome the domain mismatch effect in cross-lingual transfer learning . the initializations were used for MUSE UBLI, UN Parallel UNMT, and the SemEval 2017 task . |
On the Impact of Cross-Domain Data on German Language Models (2023.findings-emnlp)
Copied to clipboard
Amin Dada, Aokun Chen, Cheng Peng, Kaleb Smith, Ahmad Idrissi-Yaghir, Constantin Seibold, Jianning Li, Lars Heiliger, Christoph Friedrich, Daniel Truhn, Jan Egger, Jiang Bian, Jens Kleesiek, Yonghui Wu
| Challenge: | Traditionally, large language models have been trained on general web crawls or domain-specific data. |
| Approach: | They present a German dataset and a dataset aimed at containing high-quality data to examine the importance of data diversity over quality. |
| Outcome: | The proposed model outperforms models trained on quality data on multiple downstream tasks. |
Multilingual Holistic Bias: Extending Descriptors and Patterns to Unveil Demographic Biases in Languages at Scale (2023.emnlp-main)
Copied to clipboard
Marta Costa-jussà, Pierre Andrews, Eric Smith, Prangthip Hansanti, Christophe Ropers, Elahe Kalbassi, Cynthia Gao, Daniel Licht, Carleigh Wood
| Challenge: | Multilingual HolisticBias dataset includes 20,459 sentences in 50 languages . dataset is intended to uncover demographic imbalances and quantify mitigations . |
| Approach: | They propose a multilingual extension of the HolisticBias dataset . they use 118 demographic descriptors and three patterns to build multilingual sentences . |
| Outcome: | The proposed model improves translation quality when the source input only differs in gender . it also improves when the masculine human reference is used in the model . |
Evaluating Models’ Local Decision Boundaries via Contrast Sets (2020.findings-emnlp)
Copied to clipboard
Matt Gardner, Yoav Artzi, Victoria Basmov, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala, Nitish Gupta, Hannaneh Hajishirzi, Gabriel Ilharco, Daniel Khashabi, Kevin Lin, Jiangming Liu, Nelson F. Liu, Phoebe Mulcaire, Qiang Ning, Sameer Singh, Noah A. Smith, Sanjay Subramanian, Reut Tsarfaty, Eric Wallace, Ally Zhang, Ben Zhou
| 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. |
GENIE: Toward Reproducible and Standardized Human Evaluation for Text Generation (2022.emnlp-main)
Copied to clipboard
Daniel Khashabi, Gabriel Stanovsky, Jonathan Bragg, Nicholas Lourie, Jungo Kasai, Yejin Choi, Noah A. Smith, Daniel Weld
| Challenge: | Effective human evaluation of text generation tasks remains an important, open area for research. |
| Approach: | They propose a system for running standardized human evaluations across different generation tasks. |
| Outcome: | The proposed system produces standardized human evaluations across tasks . it crowdsources predictions and ranks systems on leaderboards . the proposed system is not reproducible over time and different annotator populations . |
Toxicity in Multilingual Machine Translation at Scale (2023.findings-emnlp)
Copied to clipboard
Marta Costa-jussà, Eric Smith, Christophe Ropers, Daniel Licht, Jean Maillard, Javier Ferrando, Carlos Escolano
| Challenge: | In this paper, we evaluate and analyze added toxicity when translating a large dataset from English into 164 languages. |
| Approach: | They evaluate added toxicity when translating a large dataset from English into 164 languages. |
| Outcome: | The results show that added toxicity is more prevalent in low-resource languages than in high-resolution translations. |
Causal Matching with Text Embeddings: A Case Study in Estimating the Causal Effects of Peer Review Policies (2023.findings-acl)
Copied to clipboard
| Challenge: | Using a method that uses text embeddings, we find that human judges prefer manuscript matches from our method in 70% of cases. |
| Approach: | They propose to use text embeddings to integrate a method to estimate the causal effects of peer review policies in publication venues that shift policies from single-blind to double-blinded. |
| Outcome: | The proposed method is compared with a widely-used method of stratified propensity score matching and a baseline of randomly selected matches. |
NeuroLogic A*esque Decoding: Constrained Text Generation with Lookahead Heuristics (2022.naacl-main)
Copied to clipboard
Ximing Lu, Sean Welleck, Peter West, Liwei Jiang, Jungo Kasai, Daniel Khashabi, Ronan Le Bras, Lianhui Qin, Youngjae Yu, Rowan Zellers, Noah A. Smith, Yejin Choi
| Challenge: | Existing paradigms for text generation are left-to-right decoding from autoregressive language models. |
| Approach: | They propose a decoding algorithm that incorporates heuristic estimates of future cost that are efficient for large-scale language models. |
| Outcome: | The proposed method outperforms baselines on five generation tasks and achieves new state-of-the-art performance on table-to-text generation, constrained machine translation, and keyword-constrained generation. |