09-15-2021-jam.md

• Rationales for Sequential Predictions, Keyon Vafa et al. emnlp2021 interpretability nmt combinatorial optimization

  • Rationales: subset of context that ...;
  • combinatorial optimization formulation of rationale finding: the best rationale is the smallest subset of input tokens that could predict the same prediction as the orignal ones;
  • how to measure faithfulness?

• Type of Out-of-Distribution Texts and How to Detect Them, Udit Arora et al. emnlp2021 ood issue analysis

  • Motivation is "there is little consensus on formal def. of OOD examples";
  • Propose a categorization of OOD instances according to background shift or semantic shift
  • Methods like calibration and density estimation for OOD detection are evaluated over 14 datasets

• Uncertainty-Aware Machine Translation Evaluation, Taisiya Glushkova et al. emnlp2021 nmt evaluation

  • Motivation: old works of evaluation metrics are point estimation, "which provides limitd knowledge at segment level"
  • Monte Carlo dropout and deep ensembles are used for collect groups of predictions for uncertainty estimation
  • (no reference) uncertainty-aware quality estimation can locate translation errors

• Compression, Transduction and Creation: A Unified Framework for Evaluating NLG, Mingkai Deng et al. emnlp2021 evaluation nlg

  • information alignment as a technique is used for different NLG tasks, namely, summarization, style transfer and dialog
  • a good point for writing a paper, great as a survey paper

• The Perils of Using Mechanical Turk to Evaluate Open-Ended Text Generation, Marzena Karpinska et al. emnlp2021 evaluation nlg

  • Meta-research on other (45) papers' evaluation part of open-ended text generation
  • Rethinking the faithfulness of text generation by GPT-2 and other big LMs

• Phrase-BERT Improved Phrase Embeddings from BERT with an Application to Corpus Exploration, Shufan Wang et al. emnlp2021 sentence embedding compositionality dataset analysis

  • Motivation: old phrase-embedding-from-BERT methods lacks compositionality
  • They design a contrastive fine-tuning objective function
  • Phrase-BERT embeddings can be integrated into a neural topic model

• A Massively Multilingual Analysis of Cross-linguality in Shared Embedding Space, Alex Jones et al. multilinguality representation analysis cross-lingual alignment

  • What is cross-linguality?
  • What are the for intrinsic measures of cross-linguality?

• Multi-Sentence Resampling: A Simple Approach to Alleviate Dataset Length Bias and Beam-Search Degradation, Ivan Provikov et al. arXiv nmt inference

  • Identification of dataset length bias as a key factor for large-beam issue
  • They propose a Data Augmentation methods aka. multi-sentence resampling to mitigate such problem

• Improving Zero-shot Cross-lingual Transfer between Closely Related Languages by injecting Character-level Noise, Noemi Aepli et al. emnlp2021 zero-shot cross-lingual transfer

• STraTA: Self-Training with Task Augmentation for Better Few-shot Learning, Tu Vu et al. emnlp2021 few-shot fune-tuning

  • This paper drives my interest due to this sentence in its abstract "STraTA with only 8 training examples per class achieves comparable results to standard fine-tuning with 67K training examples"

• LM-Critic: Language Models for Unsupervised Grammatical Error Correction, Michihiro Yasunaga et al. emnlp2021 Grammar Error Correction unsupervised learning

  • Use a pre-trained LM as critic to judge and estabilish the grammaticality of a sentence