Papers by Armand Joulin
Training Hybrid Language Models by Marginalizing over Segmentations (P19-1)
Copied to clipboard
| Challenge: | Statistical language modeling is the problem of estimating a probability distribution over text data. |
| Approach: | They propose to marginalize over the segmentations efficiently to compute the true probability of a sequence. |
| Outcome: | The proposed model marginalizes over the segmentations to compute the true probability of a sequence on three datasets comprising seven languages. |
Loss in Translation: Learning Bilingual Word Mapping with a Retrieval Criterion (D18-1)
Copied to clipboard
| Challenge: | Existing approaches to learn orthogonal matrix aligning bilingual lexicons are suboptimal . resulting models suffer from "hubness problem" because word vectors tend to be nearest neighbors of abnormally high number of other words. |
| Approach: | They propose a unified formulation that directly optimizes a retrieval criterion in an end-to-end fashion. |
| Outcome: | The proposed approach outperforms the state-of-the-art on word translation on standard benchmarks. |
Adaptive Attention Span in Transformers (P19-1)
Copied to clipboard
| Challenge: | We extend the maximum context size of a neural network called Transformer to 8k characters. |
| Approach: | They propose a self-attention mechanism that can learn its optimal attention span . this allows for models with longer context and the capability to catch longer dependencies. |
| Outcome: | The proposed model achieves state-of-the-art performance on text8 and enwiki8 using 8k characters with no loss of performance, and maintains control over memory footprint and computational time. |
CCMatrix: Mining Billions of High-Quality Parallel Sentences on the Web (2021.acl-long)
Copied to clipboard
| Challenge: | Using a curated common crawl corpus, we were able to mine 10.8 billion parallel sentences out of which only 2.9 billions are aligned with English. |
| Approach: | They use 32 snapshots of a curated common crawl corpus totaling 71 billion unique sentences to mine 10.8 billion parallel sentences out of which only 2.9 billions are aligned with English. |
| Outcome: | The proposed system outperforms the best single systems on the WMT’19 test set for English-German/Russian/Chinese and outperformed the best submission at the 2020 WAT workshop. |
Time Sensitive Knowledge Editing through Efficient Finetuning (2024.acl-short)
Copied to clipboard
Xiou Ge, Ali Mousavi, Edouard Grave, Armand Joulin, Kun Qian, Benjamin Han, Mostafa Arefiyan, Yunyao Li
| Challenge: | Existing locate-and-edit knowledge editing methods suffer from two limitations: they are infeasible for large scale KE in practice and require long run-time. |
| Approach: | They propose to use parametric fine-tuning techniques to update obsolete knowledge and induce new knowledge into LLMs. |
| Outcome: | The proposed methods improve the performance of KE and knowledge update in a temporal dataset with knowledge update and knowledge injection examples. |
Learning Word Vectors for 157 Languages (L18-1)
Copied to clipboard
| Challenge: | Distributed word representations, or word vectors, have been used in natural language processing for many tasks. |
| Approach: | They propose to use the encyclopedia Wikipedia and the common crawl corpus to train distributed word representations on large corpora and use them in downstream tasks. |
| Outcome: | The proposed model performs very well on 10 languages for which evaluation dataset exists. |
CCNet: Extracting High Quality Monolingual Datasets from Web Crawl Data (2020.lrec-1)
Copied to clipboard
Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzmán, Armand Joulin, Edouard Grave
| Challenge: | Pre-training text representations have led to significant improvements in many areas of natural language processing. |
| Approach: | They propose a pipeline to extract monolingual datasets from Common Crawl . pipeline follows data processing introduced in fastText that deduplicates documents . |
| Outcome: | The proposed pipeline performs standard document deduplication and language identification similar to the pipeline introduced in fastText and a filtering step to select documents close to high quality corpora like Wikipedia. |
Advances in Pre-Training Distributed Word Representations (L18-1)
Copied to clipboard
| Challenge: | Pre-trained word representations are a building block of many Natural Language Processing and Machine Learning applications. |
| Approach: | They propose to combine known tricks and a set of publicly available pre-trained word vector representations to train high-quality representations. |
| Outcome: | The proposed models outperform the current state of the art on a number of tasks while maintaining a high training speed to scale to massive amount of data. |
Target Conditioning for One-to-Many Generation (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Neural Machine Translation models lack diversity in their generated translations, even when paired with search algorithm, like beam search. |
| Approach: | They propose to model one-to-many mapping by conditioning a decoder on a latent variable that represents the domain of target sentences. |
| Outcome: | The proposed method can scale to any number of domains without affecting performance or training time. |
Cooperative Learning of Disjoint Syntax and Semantics (N19-1)
Copied to clipboard
| Challenge: | Existing models that learn to jointly infer an expression’s syntactic structure and its semantics fail to learn the correct parsing strategy on mathematical expressions generated from a simple context-free grammar. |
| Approach: | They propose a recursive model that learns to jointly infer an expression’s syntactic structure and its semantics without requiring a formal supervision. |
| Outcome: | The proposed model performs competitively on several natural language tasks, such as Natural Language Inference and Sentiment Analysis. |