Papers by Javier Ferrando
Information Flow Routes: Automatically Interpreting Language Models at Scale (2024.emnlp-main)
Copied to clipboard
| Challenge: | Current state-of-the-art language models (LMs) are built on top of the Transformer architecture. |
| Approach: | They propose to build graphs where nodes correspond to token representations and edges to computations . they show that attention heads and subword merging heads are important . |
| Outcome: | The proposed model can analyze behavior for specific types of predictions, or different domains. |
On the Similarity of Circuits across Languages: a Case Study on the Subject-verb Agreement Task (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Several algorithms implemented by language models have been successfully reverse-engineered, leaving it unclear how universal circuits are across different settings. |
| Approach: | They propose to use Gemma 2B to solve the subject-verb agreement task across two different languages, English and Spanish. |
| Outcome: | The proposed circuits solve the subject-verb agreement task across two different languages, and are language-independent and language-dependent. |
Measuring the Mixing of Contextual Information in the Transformer (2022.emnlp-main)
Copied to clipboard
| Challenge: | Experimentally, we show that ALTI provides more faithful explanations and increased robustness than gradient-based methods. |
| Approach: | They propose to measure token-to-token interactions within each layer and then use them to aggregate model predictions. |
| Outcome: | The proposed method provides more faithful explanations and increased robustness than gradient-based methods. |
Attention Weights in Transformer NMT Fail Aligning Words Between Sequences but Largely Explain Model Predictions (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Using attention weights, we show that NMT models make alignment errors by relying on uninformative tokens from the source sequence. |
| Approach: | They propose to use attention weights to regulate alignment errors in NMT models . they propose methods that largely reduce the word alignment error rate compared to standard induced alignments from attention weighted tokens. |
| Outcome: | The proposed methods reduce the word alignment error rate compared to standard induced alignments from attention weights. |
LM Transparency Tool: Interactive Tool for Analyzing Transformer Language Models (2024.acl-demos)
Copied to clipboard
| Challenge: | Existing tools focus on isolated parts of the decision-making process, but LM-TT makes the entire prediction process transparent. |
| Approach: | They present an open-source toolkit for analyzing the internal workings of Transformer-based language models. |
| Outcome: | The LM Transparency Tool makes the entire prediction process transparent . it shows the importance of specific component at each step . |
Towards Opening the Black Box of Neural Machine Translation: Source and Target Interpretations of the Transformer (2022.emnlp-main)
Copied to clipboard
| Challenge: | Neural Machine Translation (NMT) relies on source sentence and target prefix attributions for each input token. |
| Approach: | They propose an interpretability method that tracks input tokens’ attributions for both contexts and extends it to any encoder-decoder Transformer-based model. |
| Outcome: | The proposed method can be extended to any encoder-decoder Transformer-based model and provides insights into their behaviour. |
On the Locality of Attention in Direct Speech Translation (2022.acl-srw)
Copied to clipboard
| Challenge: | Recent advances in NLP have created problems with the complexity of the self-attention layer. |
| Approach: | They propose to substitute standard self-attention with a local efficient one to avoid the computation of attention weights. |
| Outcome: | The proposed model matches the baseline performance and improves efficiency by skipping the computation of weights that standard attention discards. |
Neurons in Large Language Models: Dead, N-gram, Positional (2024.findings-acl)
Copied to clipboard
| Challenge: | We analyze a family of large language models in such a lightweight manner that can be done on a single GPU. |
| Approach: | They analyze a family of large language models that rely on whether an FFN neuron is activated or not . they use a single GPU to analyze the OPT network and use it to analyze data . |
| Outcome: | The proposed model is lightweight and can be run on a single GPU. |
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. |
Explaining How Transformers Use Context to Build Predictions (2023.acl-long)
Copied to clipboard
| Challenge: | Existing methods for analyzing input attributions for a model's prediction are unclear how prior words affect the model' s decision throughout the layers. |
| Approach: | They propose a procedure to analyze models for language generation using the Transformer and a comparison of their results with evidence of the linguistic phenomena. |
| Outcome: | The proposed method consistently aligns better than gradient-based and perturbation-based baselines and generates human-like source-target alignments for building predictions. |