| Challenge: | Current text generators require sampling from a modified softmax to avoid degenerate text . entmax sampling creates a mismatch between training and testing conditions . |
| Approach: | They propose to use entmax transformation to train and sample from a sparse language model to avoid degenerate text. |
| Outcome: | The proposed model improves fluency and consistency, fewer repetitions, and n-gram diversity closer to human text. |
Similar Papers
Speeding Up Entmax (2022.findings-naacl)
Copied to clipboard
| Challenge: | Recent studies suggest that sparsity is a problem when the trained model is used for inference. |
| Approach: | They propose an alternative to softmax that produces a dense probability distribution but is slower than softmax. |
| Outcome: | The proposed method keeps its virtuous characteristics but is slower than softmax and achieves on par or better performance in machine translation task. |
For Generated Text, Is NLI-Neutral Text the Best Text? (2023.findings-emnlp)
Copied to clipboard
| Challenge: | a perfectly informative agent would eschew utterances that are redundant or contradict that which they have already said. |
| Approach: | They propose to use a pre-trained NLI model to assess whether a sentence entails, contradicts, or is neutral to prompt and preceding text. |
| Outcome: | The proposed model is predictive of generation errors made by GPT-3. |
Smoothing and Shrinking the Sparse Seq2Seq Search Space (2021.naacl-main)
Copied to clipboard
| Challenge: | entmax-based sparse sequence-to-sequence models give high scores to short hypotheses . ent max models can shrink the search space by assigning zero probability to bad hypothese . |
| Approach: | They propose entmax-based sparse sequence-to-sequence models that minimize cross-entropy and use softmax to compute local normalized probabilities over target sequences. |
| Outcome: | The proposed models remove a major source of model error for word-level tasks . the proposed models improve cross-lingual morphological inflection and machine translation . |
RankGen: Improving Text Generation with Large Ranking Models (2022.emnlp-main)
Copied to clipboard
| Challenge: | Modern language models assign high probabilities to output sequences that are repetitive, incoherent, or irrelevant to the prefix. |
| Approach: | They propose a 1.2B parameter encoder model for English that scores model generations given a prefix. |
| Outcome: | The proposed model outperforms decoding algorithms on automatic metrics and human evaluations with English writers. |
Sparse Sequence-to-Sequence Models (P19-1)
Copied to clipboard
| Challenge: | Sequence-to-sequence models are dense and assigning nonzero probability to implausible outputs. |
| Approach: | They propose a new family of -entmax transformations that includes softmax and sparsemax as particular cases and is sparser for any > 1 . they provide fast algorithms to evaluate these transformations and their gradients, which scale well for large vocabulary sizes. |
| Outcome: | The proposed models are able to produce sparse alignments and assign nonzero probability to short list of plausible outputs, sometimes rendering beam search exact. |
Generating Text from Language Models (2023.acl-tutorials)
Copied to clipboard
| Challenge: | a growing percentage of natural language processing tasks focus on the generation of text from probabilistic language models. |
| Approach: | They will provide a centralized discussion of critical considerations when choosing how to generate from a language model. |
| Outcome: | This tutorial will provide a centralized discussion of critical considerations when choosing how to generate from a language model. |
On the Blind Spots of Model-Based Evaluation Metrics for Text Generation (2023.acl-long)
Copied to clipboard
| Challenge: | Existing methods for text generation evaluation metrics are lacking in robustness analysis. |
| Approach: | They propose to use stress tests to test for errors in text generation evaluation metrics . they find that BERTScore is confused by truncation errors in summarization . |
| Outcome: | The proposed stress tests show that they are insensitive to errors in open-ended generation, translation, and summarization. |
Can Diffusion Model Achieve Better Performance in Text Generation ? Bridging the Gap between Training and Inference ! (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing models for text generation use a discrete data embedding module to map the data into the continuous space. |
| Approach: | They propose two methods to bridge the gap between training and inference by mapping the discrete text into the continuous space. |
| Outcome: | The proposed methods can achieve 100 200 speedup with better performance on 6 generation tasks. |
The Curious Decline of Linguistic Diversity: Training Language Models on Synthetic Text (2024.findings-naacl)
Copied to clipboard
| Challenge: | a new study examines the effects of training language models on synthetic data generated by their predecessors. |
| Approach: | They propose to use recursive finetuning techniques to assess linguistic diversity of models. |
| Outcome: | The proposed metrics show a decrease in diversity of model outputs through successive iterations, especially for tasks demanding high levels of creativity. |
Semantic Evaluation of Multilingual Data-to-Text Generation via NLI Fine-Tuning: Precision, Recall and F1 scores (2025.findings-acl)
Copied to clipboard
| Challenge: | KG-to-Text models are prone to errors like Additions and Omissions, and few languages are taken into account since both train and test data are not readily available. |
| Approach: | They propose a multilingual evaluation framework that is reference-less . it allows estimating how much a KG-to-Text Model under- (omission) or over- (addition) generates. |
| Outcome: | The proposed evaluation framework outperforms prior reference-less metrics in correlation with human judgments and provides scores for precision and recall. |