Papers by Marta R. Costa-jussà

23 papers
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.
Multilingual Machine Translation: Closing the Gap between Shared and Language-specific Encoder-Decoders (2021.eacl-main)

Copied to clipboard

Challenge: State-of-the-art multilingual machine translation relies on a universal encoder-decoder, which requires retraining the entire system to add new languages.
Approach: They propose an encoder-decoder approach that can be extended to new languages by learning their corresponding modules.
Outcome: The proposed approach outperforms the universal encoder-decoder by 3.28 BLEU points on average while allowing to add new languages without retraining the rest of the modules.
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.
Abusive language in Spanish children and young teenager’s conversations: data preparation and short text classification with contextual word embeddings (2020.lrec-1)

Copied to clipboard

Challenge: Existing studies on how to automatically detect abusive short texts are gaining interest in the natural language processing community.
Approach: They propose to use a contextual word embedding model to automatically detect abusive short texts for Spanish language.
Outcome: The proposed model outperforms classical methods in the detection of abusive short texts for the spanish language.
Automatic Spanish Translation of SQuAD Dataset for Multi-lingual Question Answering (2020.lrec-1)

Copied to clipboard

Challenge: Existing methods to train multilingual QA systems are limited for other languages . cross-lingual learning is a technique that transfers knowledge from source to target language with fewer training data.
Approach: They propose a translation method to translate the Stanford Question Answering Dataset to Spanish and a multilingual-BERT model to train Spanish QA systems.
Outcome: The proposed method outperforms the previous benchmarks for cross-lingual extractive QA.
Continual Lifelong Learning in Natural Language Processing: A Survey (2020.coling-main)

Copied to clipboard

Challenge: Existing approaches to continual learning (CL) are costly and time-consuming.
Approach: They propose to examine the problem of continual learning in NLP through the lens of various NLP tasks and provide a critical review of existing methods.
Outcome: The proposed methods are critical to the development of CL models and provide a critical review of existing methods and datasets.
Improving Language and Modality Transfer in Translation by Character-level Modeling (2025.acl-long)

Copied to clipboard

Challenge: Current translation systems cover only 5% of the world's languages . expanding to the long-tail of low-resource languages requires data-efficient methods that rely on cross-lingual and cross-modal knowledge transfer.
Approach: They propose a character-based approach to improve adaptability to new languages and modalities by using a teacher-student approach and parallel translation data to obtain a SONAR character-level encoder.
Outcome: The proposed model outperforms subword-based models in speech-to-text translation on the FLEURS benchmark on 33 languages and achieves state-of-the-art generalizability to unseen languages.
SpeechAlign: A Framework for Speech Translation Alignment Evaluation (2024.lrec-main)

Copied to clipboard

Challenge: Speech-to-Speech and Speech- to-Text translation are currently dynamic areas of research.
Approach: They propose a framework to evaluate source-target alignment in speech models . they introduce a speech gold alignment dataset and introduce two new metrics .
Outcome: The proposed framework evaluates source-target alignment quality within speech models.
From Bilingual to Multilingual Neural Machine Translation by Incremental Training (P19-2)

Copied to clipboard

Challenge: Existing approaches to multilingual neural machine translation are based on task specific models and the addition of one more language is only possible by retraining the whole system.
Approach: They propose a training schedule that scales to more languages without modification of previous components.
Outcome: The proposed training schedule shows close results to state-of-the-art in the WMT task.
GeBioToolkit: Automatic Extraction of Gender-Balanced Multilingual Corpus of Wikipedia Biographies (2020.lrec-1)

Copied to clipboard

Challenge: a tool for extracting multilingual parallel corpora at sentence level with document and gender information from Wikipedia biographies.
Approach: They propose a tool for extracting multilingual parallel corpora at sentence level with document and gender information from Wikipedia biographies.
Outcome: The proposed tool extracts a corpus of 2,000 sentences in English, Spanish and Catalan.
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.
BOUQuET : dataset, Benchmark and Open initiative for Universal Quality Evaluation in Translation (2025.emnlp-main)

Copied to clipboard

Challenge: BOUQUET is a multi-way, multicentric and multi-register/domain dataset and benchmark . the dataset is handcrafted in 8 non-English languages .
Approach: They propose to use BOUQuET to collect a multi-way, multicentric and multi-register/domain dataset and benchmark in 8 non-English languages.
Outcome: The proposed dataset is available at https://huggingface.co/datasets/facebook/bouquet.
Evaluating Gender Bias in Speech Translation (2022.lrec-1)

Copied to clipboard

Challenge: Existing evaluation techniques for gender biases are lacking in the field of machine translation.
Approach: They propose to use a free evaluation set to evaluate gender bias in speech translation.
Outcome: The proposed set is the speech version of WinoMT, an MT challenge set.
Detecting and Mitigating Hallucinations in Machine Translation: Model Internal Workings Alone Do Well, Sentence Similarity Even Better (2023.acl-long)

Copied to clipboard

Challenge: a recent study shows that without artificially encouraging models to hallucinate, existing methods fall short . hallucinations are cases when the model generates output that is partially or fully unrelated to the source sentence.
Approach: They propose a method that evaluates the percentage of the source contribution to a generated translation.
Outcome: The proposed method improves detection accuracy for the most severe hallucinations by a factor of 2.
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.
Multilingual, Multi-scale and Multi-layer Visualization of Intermediate Representations (D19-3)

Copied to clipboard

Challenge: Currently, the main alternatives to deal with sequences are Recurrent Neural Networks (RNN) architectures and the Transformer.
Approach: They propose a web-based tool that visualizes the sentence and token representations of RNNs and Transformer architectures at the sentence level.
Outcome: The proposed visualization tool analyses gender inequalities in contextual word embeddings and the common language representation in a multilingual machine translation system.
LCFO: Long Context and Long Form Output Dataset and Benchmarking (2025.findings-acl)

Copied to clipboard

Challenge: Using long text outputs to evaluate progress in summarization and summary expansion tasks is challenging.
Approach: They propose a framework for assessing gradual summarization and summary expansion capabilities across diverse domains.
Outcome: The proposed framework provides alignments between specific QA pairs and corresponding summaries in 7 domains.
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.
Combining Subword Representations into Word-level Representations in the Transformer Architecture (2020.acl-srw)

Copied to clipboard

Challenge: Currently dominant approaches use word-level tokens, but this increases the length of the sequences and makes it difficult to profit from word-based information.
Approach: They propose to combine subword-level representations into word-level ones in the first layers of the encoder, reducing the effective length of the sequences in the following layers.
Outcome: The proposed model maintains translation quality with no extra word-level information . it is superior to the current dominant method for incorporating word- level source language information a priori .
BLASER: A Text-Free Speech-to-Speech Translation Evaluation Metric (2023.acl-long)

Copied to clipboard

Challenge: End-to-End speech-to speech translation is generally evaluated with text-based metrics . this means generated speech has to be automatically transcribed, making the evaluation dependent on ASR systems.
Approach: They propose a text-free evaluation metric for end-to-end speech-tospeech translation, named BLASER, to avoid the dependency on automatic speech recognition systems.
Outcome: The proposed metric avoids the dependency on automatic speech recognition systems by encoding generated speech segments into a shared embedding space.
On the Role of Speech Data in Reducing Toxicity Detection Bias (2025.naacl-long)

Copied to clipboard

Challenge: Text toxicity detection systems produce disproportionate rates of false positives on demographic groups . toxicity classification systems often misinterpret benign group mentions as toxic .
Approach: They use group annotations to compare text-based and speech-based toxicity detection systems.
Outcome: The results show that access to speech data supports reduced bias against group mentions . the authors recommend improving classifiers, rather than transcription pipelines if possible .
2M-BELEBELE: Highly Multilingual Speech and American Sign Language Comprehension Dataset Download PDF (2025.findings-acl)

Copied to clipboard

Challenge: We extend the BELEBELE dataset to speech and sign, and extend the Automatic Speech Recognition Benchmark, FLEURS, by 20%.
Approach: They extend the BELEBELE and FLEURS speech comprehension datasets to speech and sign . they evaluate the datasets for 5-shot and zero-shot settings and find that the accuracy is 10% lower than reading comprehension.
Outcome: The proposed dataset covers 91 spoken languages and one sign language (ASL) it also extends the Automatic Speech Recognition Benchmark, FLEURS, by 20% across languages.
Enhancing Word Embeddings with Knowledge Extracted from Lexical Resources (2020.acl-srw)

Copied to clipboard

Challenge: In this paper, we present an effective method for semantic specialization of word vector representations.
Approach: They propose a method for semantic specialization of word vector representations using BabelNet.
Outcome: The proposed method improves on word similarity and dialog state tracking tasks.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations