Positional Artefacts Propagate Through Masked Language Model Embeddings (2021.acl-long)
Copied to clipboard
| Challenge: | Existing word embedding models have a uniform pitfall in assigning a static vector to a word type. |
| Approach: | They propose a neuron-level analysis method to investigate the source of this information by comparing outlier neurons within BERT and RoBERTa’s hidden state vectors. |
| Outcome: | The proposed method pre-trains the RoBERTa-based models and shows that the outliers disappear without positional embeddings. |
Similar Papers
Dwell in the Beginning: How Language Models Embed Long Documents for Dense Retrieval (2024.acl-short)
Copied to clipboard
| Challenge: | Existing studies have shown that Transformer-based language models lose information in the middle of input sequences, especially in the context of web document retrieval. |
| Approach: | They examine position biases at multiple stages of the training pipeline for an encoder-decoder neural retrieval model, namely language model pre-training, contrastive pre- training, and contrastive fine-tuning. |
| Outcome: | The proposed model generates embeddings that better capture the beginning of the input content, with fine-tuning further aggravating this effect. |
Exploring Anisotropy and Outliers in Multilingual Language Models for Cross-Lingual Semantic Sentence Similarity (2023.findings-acl)
Copied to clipboard
| Challenge: | Recent studies have shown that contextual language models display outlier dimensions . this is true for monolingual and multilingual models, but little work has been done on multilingual contexts . |
| Approach: | They investigate outlier dimensions and their relationship to anisotropy in multilingual contexts . they focus on cross-lingual semantic similarity tasks . |
| Outcome: | The proposed model improves on cross-lingual semantic similarity tasks. |
Better Word Embeddings by Disentangling Contextual n-Gram Information (N19-1)
Copied to clipboard
| Challenge: | Pre-trained word vectors are ubiquitous in Natural Language Processing applications. |
| Approach: | They show that word embeddings with bigram and trigram embedds improve unigram embeds . they claim this removes contextual information from unigrammes, resulting in better unigraph embedders . |
| Outcome: | The proposed model outperforms competing models on a wide variety of tasks. |
On the Language Encoder of Contrastive Cross-modal Models (2024.findings-acl)
Copied to clipboard
Mengjie Zhao, Junya Ono, Zhi Zhong, Chieh-Hsin Lai, Yuhta Takida, Naoki Murata, Wei-Hsiang Liao, Takashi Shibuya, Hiromi Wakaki, Yuki Mitsufuji
| Challenge: | Pretrained audio-language models such as AudioCLIP and AudioCLAP have shown promising results on vision-language (VL) tasks. |
| Approach: | They extensively evaluate how unsupervised and supervised sentence embedding training affect language encoder quality and cross-modal task performance. |
| Outcome: | The proposed model improves on visual-language (VL) and audio-language tasks when the amount of training data is large. |
Do Transformers Parse while Predicting the Masked Word? (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing studies show that pre-trained language models encode linguistic structures like parse trees while being trained unsupervised. |
| Approach: | They propose to train pre-trained language models to encode linguistic structures like parse trees while unsupervised. |
| Outcome: | The proposed model performs optimally for masked language modeling loss on the English PCFG. |
Contextual Embeddings: When Are They Worth It? (2020.acl-main)
Copied to clipboard
| Challenge: | In recent years, rich contextual embeddings have enabled rapid progress on benchmarks like GLUE, but require significant computational resources during pretraining and during downstream task training and inference. |
| Approach: | They empirically compare contextual embeddings with classic pretrained embedders and a random word embeddable with a simple baseline. |
| Outcome: | The proposed models perform within 5 to 10% accuracy on industry-scale data. |
On the Sentence Embeddings from Pre-trained Language Models (2020.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained contextual representations like BERT have been widely used for NLP tasks. |
| Approach: | They propose to transform anisotropic sentence embedding distribution to smooth and isotropic Gaussian distribution by normalizing flows that are learned with an unsupervised objective. |
| Outcome: | The proposed method achieves significant performance gains over state-of-the-art embeddings on a variety of semantic textual similarity tasks. |
General Purpose Text Embeddings from Pre-trained Language Models for Scalable Inference (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Large pre-trained language models are currently used for many NLP tasks . however, inference for these models requires significant computational resources . |
| Approach: | They propose to use a shared text encoder to amortize the computational cost of inference over multiple tasks. |
| Outcome: | The proposed method reduces the size of the extracted representations by a factor of 16 to store them for later use. |
How to represent a word and predict it, too: Improving tied architectures for language modelling (D18-1)
Copied to clipboard
| Challenge: | Recent state-of-the-art models use word embeddings as input and output mappings instead of tied models. |
| Approach: | They propose to decouple hidden state from word embedding prediction . they extend their proposed modification to word2vec models . |
| Outcome: | The proposed architectures achieve comparable or better results compared to previous models without tying . the proposed architecture reduces parameters, enabling more compact models and faster learning. |
Contrastive Visual Semantic Pretraining Magnifies the Semantics of Natural Language Representations (2022.acl-long)
Copied to clipboard
| Challenge: | Large-scale "natural language supervision" using image captions has enabled the first "zero-shot" AI image classifiers, which allow users to create their own image classes using natural language, yet outperform supervised models on common language-and-image tasks. |
| Approach: | They compare the geometry and semantic properties of contextualized English language representations formed by GPT-2 and CLIP, a zero-shot multimodal image classifier which adapts the GPT2 architecture to encode image captions. |
| Outcome: | The proposed classifier outperforms GPT-2 on word-level semantic intrinsic evaluation tasks and achieves a new corpus-based state of the art for the RG65 evaluation. |