Translation systems, including foundation models capable of translation, can produce errors that result in gender mistranslation, and such errors can be especially harmful. To measure the extent of such potential harms when translating into and out of English, we introduce a dataset, MiTTenS, covering 26 languages from a variety of language families and scripts, including several traditionally under-represented in digital resources. The dataset is constructed with handcrafted passages that target known failure patterns, longer synthetically generated passages, and natural passages sourced from multiple domains. We demonstrate the usefulness of the dataset by evaluating both neural machine translation systems and foundation models, and show that all systems exhibit gender mistranslation and potential harm, even in high resource languages.

Although the advancements of pre-trained Large Language Models have significantly accelerated recent progress in NLP, their ever-increasing size poses significant challenges for conventional fine-tuning, especially in memory-intensive tasks. We investigate the potential of Parameter-Efficient Fine-Tuning, focusing on Low-Rank Adaptation (LoRA), in the domain of multilingual summarization, a task that is both challenging (due to typically long inputs), and relatively unexplored. We conduct an extensive study across different data availability scenarios, including high- and low-data settings, and cross-lingual transfer, leveraging models of different sizes. Our findings reveal that LoRA is competitive with full fine-tuning when trained with high quantities of data, and excels in low-data scenarios and cross-lingual transfer. We also study different strategies for few-shot cross-lingual transfer, finding that continued LoRA tuning outperforms full fine-tuning and the dynamic composition of language-specific LoRA modules.

Transformer-based language models (LMs) are known to capture factual knowledge in their parameters. While previous work looked into where factual associations are stored, only little is known about how they are retrieved internally during inference. We investigate this question through the lens of information flow. Given a subject-relation query, we study how the model aggregates information about the subject and relation to predict the correct attribute. With interventions on attention edges, we first identify two critical points where information propagates to the prediction: one from the relation positions followed by another from the subject positions. Next, by analyzing the information at these points, we unveil a three-step internal mechanism for attribute extraction. First, the representation at the last-subject position goes through an enrichment process, driven by the early MLP sublayers, to encode many subject-related attributes. Second, information from the relation propagates to the prediction. Third, the prediction representation “queries” the enriched subject to extract the attribute. Perhaps surprisingly, this extraction is typically done via attention heads, which often encode subject-attribute mappings in their parameters. Overall, our findings introduce a comprehensive view of how factual associations are stored and extracted internally in LMs, facilitating future research on knowledge localization and editing.

Feature attribution a.k.a. input salience methods which assign an importance score to a feature are abundant but may produce surprisingly different results for the same model on the same input. While differences are expected if disparate definitions of importance are assumed, most methods claim to provide faithful attributions and point at the features most relevant for a model’s prediction. Existing work on faithfulness evaluation is not conclusive and does not provide a clear answer as to how different methods are to be compared.Focusing on text classification and the model debugging scenario, our main contribution is a protocol for faithfulness evaluation that makes use of partially synthetic data to obtain ground truth for feature importance ranking. Following the protocol, we do an in-depth analysis of four standard salience method classes on a range of datasets and lexical shortcuts for BERT and LSTM models. We demonstrate that some of the most popular method configurations provide poor results even for simple shortcuts while a method judged to be too simplistic works remarkably well for BERT.

Recent advances in NLP often stem from large transformer-based pre-trained models, which rapidly grow in size and use more and more training data. Such models are often released to the public so that end users can fine-tune them on a task dataset. While it is common to treat pre-training data as public, it may still contain personally identifiable information (PII), such as names, phone numbers, and copyrighted material. Recent findings show that the capacity of these models allows them to memorize parts of the training data, and suggest differentially private (DP) training as a potential mitigation. While there is recent work on DP fine-tuning of NLP models, the effects of DP pre-training are less well understood: it is not clear how downstream performance is affected by DP pre-training, and whether DP pre-training mitigates some of the memorization concerns. We focus on T5 and show that by using recent advances in JAX and XLA we can train models with DP that do not suffer a large drop in pre-training utility, nor in training speed, and can still be fine-tuned to high accuracies on downstream tasks (e.g. GLUE). Moreover, we show that T5’s span corruption is a good defense against data memorization.

Recent advances in NLP often stem from large transformer-based pre-trained models, which rapidly grow in size and use more and more training data. Such models are often released to the public so that end users can fine-tune them on a task dataset. While it is common to treat pre-training data as public, it may still contain personally identifiable information (PII), such as names, phone numbers, and copyrighted material. Recent findings show that the capacity of these models allows them to memorize parts of the training data, and suggest differentially private (DP) training as a potential mitigation. While there is recent work on DP fine-tuning of NLP models, the effects of DP pre-training are less well understood it is not clear how downstream performance is affected by DP pre-training, and whether DP pre-training mitigates some of the memorization concerns. We focus on T5 and show that by using recent advances in JAX and XLA we can train models with DP that do not suffer a large drop in pre-training utility, nor in training speed, and can still be fine-tuned to high accuracy on downstream tasks (e.g. GLUE). Moreover, we show that T5s span corruption is a good defense against data memorization.

(CITATION) argued for using random splits rather than standard splits in NLP experiments. We argue that random splits, like standard splits, lead to overly optimistic performance estimates. We can also split data in biased or adversarial ways, e.g., training on short sentences and evaluating on long ones. Biased sampling has been used in domain adaptation to simulate real-world drift; this is known as the covariate shift assumption. In NLP, however, even worst-case splits, maximizing bias, often under-estimate the error observed on new samples of in-domain data, i.e., the data that models should minimally generalize to at test time. This invalidates the covariate shift assumption. Instead of using multiple random splits, future benchmarks should ideally include multiple, independent test sets instead; if infeasible, we argue that multiple biased splits leads to more realistic performance estimates than multiple random splits.

There is a recent surge of interest in using attention as explanation of model predictions, with mixed evidence on whether attention can be used as such. While attention conveniently gives us one weight per input token and is easily extracted, it is often unclear toward what goal it is used as explanation. We find that often that goal, whether explicitly stated or not, is to find out what input tokens are the most relevant to a prediction, and that the implied user for the explanation is a model developer. For this goal and user, we argue that input saliency methods are better suited, and that there are no compelling reasons to use attention, despite the coincidence that it provides a weight for each input. With this position paper, we hope to shift some of the recent focus on attention to saliency methods, and for authors to clearly state the goal and user for their explanations.

We present the Language Interpretability Tool (LIT), an open-source platform for visualization and understanding of NLP models. We focus on core questions about model behavior: Why did my model make this prediction? When does it perform poorly? What happens under a controlled change in the input? LIT integrates local explanations, aggregate analysis, and counterfactual generation into a streamlined, browser-based interface to enable rapid exploration and error analysis. We include case studies for a diverse set of workflows, including exploring counterfactuals for sentiment analysis, measuring gender bias in coreference systems, and exploring local behavior in text generation. LIT supports a wide range of models—including classification, seq2seq, and structured prediction—and is highly extensible through a declarative, framework-agnostic API. LIT is under active development, with code and full documentation available at https://github.com/pair-code/lit.

The success of neural networks comes hand in hand with a desire for more interpretability. We focus on text classifiers and make them more interpretable by having them provide a justification–a rationale–for their predictions. We approach this problem by jointly training two neural network models: a latent model that selects a rationale (i.e. a short and informative part of the input text), and a classifier that learns from the words in the rationale alone. Previous work proposed to assign binary latent masks to input positions and to promote short selections via sparsity-inducing penalties such as L0 regularisation. We propose a latent model that mixes discrete and continuous behaviour allowing at the same time for binary selections and gradient-based training without REINFORCE. In our formulation, we can tractably compute the expected value of penalties such as L0, which allows us to directly optimise the model towards a pre-specified text selection rate. We show that our approach is competitive with previous work on rationale extraction, and explore further uses in attention mechanisms.

We present Joey NMT, a minimalist neural machine translation toolkit based on PyTorch that is specifically designed for novices. Joey NMT provides many popular NMT features in a small and simple code base, so that novices can easily and quickly learn to use it and adapt it to their needs. Despite its focus on simplicity, Joey NMT supports classic architectures (RNNs, transformers), fast beam search, weight tying, and more, and achieves performance comparable to more complex toolkits on standard benchmarks. We evaluate the accessibility of our toolkit in a user study where novices with general knowledge about Pytorch and NMT and experts work through a self-contained Joey NMT tutorial, showing that novices perform almost as well as experts in a subsequent code quiz. Joey NMT is available at https://github.com/joeynmt/joeynmt.

Semantic representations have long been argued as potentially useful for enforcing meaning preservation and improving generalization performance of machine translation methods. In this work, we are the first to incorporate information about predicate-argument structure of source sentences (namely, semantic-role representations) into neural machine translation. We use Graph Convolutional Networks (GCNs) to inject a semantic bias into sentence encoders and achieve improvements in BLEU scores over the linguistic-agnostic and syntax-aware versions on the English–German language pair.

Lake and Baroni (2018) recently introduced the SCAN data set, which consists of simple commands paired with action sequences and is intended to test the strong generalization abilities of recurrent sequence-to-sequence models. Their initial experiments suggested that such models may fail because they lack the ability to extract systematic rules. Here, we take a closer look at SCAN and show that it does not always capture the kind of generalization that it was designed for. To mitigate this we propose a complementary dataset, which requires mapping actions back to the original commands, called NACS. We show that models that do well on SCAN do not necessarily do well on NACS, and that NACS exhibits properties more closely aligned with realistic use-cases for sequence-to-sequence models.

We present a simple and effective approach to incorporating syntactic structure into neural attention-based encoder-decoder models for machine translation. We rely on graph-convolutional networks (GCNs), a recent class of neural networks developed for modeling graph-structured data. Our GCNs use predicted syntactic dependency trees of source sentences to produce representations of words (i.e. hidden states of the encoder) that are sensitive to their syntactic neighborhoods. GCNs take word representations as input and produce word representations as output, so they can easily be incorporated as layers into standard encoders (e.g., on top of bidirectional RNNs or convolutional neural networks). We evaluate their effectiveness with English-German and English-Czech translation experiments for different types of encoders and observe substantial improvements over their syntax-agnostic versions in all the considered setups.

PARSEVAL, the default paradigm for evaluating constituency parsers, calculates parsing success (Precision/Recall) as a function of the number of matching labeled brackets across the test set. Nodes in constituency trees, however, are connected together to reflect important linguistic relations such as predicate-argument and direct-dominance relations between categories. In this paper, we present FREVAL, a generalization of PARSEVAL, where the precision and recall are calculated not only for individual brackets, but also for co-occurring, connected brackets (i.e. fragments). FREVAL fragments precision (FLP) and recall (FLR) interpolate the match across the whole spectrum of fragment sizes ranging from those consisting of individual nodes (labeled brackets) to those consisting of full parse trees. We provide evidence that FREVAL is informative for inspecting relative parser performance by comparing a range of existing parsers.