Despite the impressive performance of large language models (LLMs), theyoften lag behind specialized models in various tasks. LLMs only use a fractionof the existing training data for in-context learning, while task-specificmodels harness the full dataset for fine-tuning. In this work, we tackle theproblem of leveraging training data to improve the performance of LLMs withoutfine-tuning. Our approach directly targets LLM predictions without requiringaccess to their weights. We create a pool of candidates from the LLM throughfew-shot prompting and we employ a compact model, the LM-corrector (LMCor),specifically trained to merge these candidates to produce an enhanced output.Our experiments on four natural language generation tasks demonstrate that evena small LMCor model (250M) substantially improves the few-shot performance ofLLMs (62B), matching and even outperforming standard fine-tuning. Furthermore,we illustrate the robustness of LMCor against different prompts, therebyminimizing the need for extensive prompt engineering. Finally, we show thatLMCor can be seamlessly integrated with different LLMs at inference, serving asa plug-and-play module to improve their performance.

Neural machine translation systems estimate probabilities of target sentences given source sentences, yet these estimates may not align with human preferences. This work introduces QE-fusion, a method that synthesizes translations using a quality estimation metric (QE), which correlates better with human judgments. QE-fusion leverages a pool of candidates sampled from a model, combining spans from different candidates using a QE metric such as CometKiwi. We compare QE-fusion against beam search and recent reranking techniques, such as Minimum Bayes Risk decoding or QE-reranking. Our method consistently improves translation quality in terms of COMET and BLEURT scores when applied to large language models (LLMs) used for translation (PolyLM, XGLM, Llama2, Mistral, ALMA, and Tower) and to multilingual translation models (NLLB), over five language pairs. Notably, QE-fusion exhibits larger improvements for LLMs due to their ability to generate diverse outputs. We demonstrate that our approach generates novel translations in over half of the cases and consistently outperforms other methods across varying numbers of candidates (5–200). Furthermore, we empirically establish that QE-fusion scales linearly with the number of candidates in the pool.

Poem generation with language models requires the modeling of rhyming patterns. We propose a novel solution for learning to rhyme, based on synthetic data generated with a rule-based rhyming algorithm. The algorithm and an evaluation metric use a phonetic dictionary and the definitions of perfect and assonant rhymes. We fine-tune a GPT-2 English model with 124M parameters on 142 MB of natural poems and find that this model generates consecutive rhymes infrequently (11%). We then fine-tune the model on 6 MB of synthetic quatrains with consecutive rhymes (AABB) and obtain nearly 60% of rhyming lines in samples generated by the model. Alternating rhymes (ABAB) are more difficult to model because of longer-range dependencies, but they are still learnable from synthetic data, reaching 45% of rhyming lines in generated samples.

We present a very simple method for extending pretrained machine translation metrics to incorporate document-level context. We apply our method to four popular metrics: BERTScore, Prism, COMET, and the reference-free metric COMET-QE. We evaluate our document-level metrics on the MQM annotations from the WMT 2021 metrics shared task and find that the document-level metrics outperform their sentence-level counterparts in about 85% of the tested conditions, when excluding results on low-quality human references. Additionally, we show that our document-level extension of COMET-QE dramatically improves accuracy on discourse phenomena tasks, supporting our hypothesis that our document-level metrics are resolving ambiguities in the reference sentence by using additional context.

In this paper, we present the systems submitted by our team from the Institute of ICT (HEIG-VD / HES-SO) to the Unsupervised MT and Very Low Resource Supervised MT task. We first study the improvements brought to a baseline system by techniques such as back-translation and initialization from a parent model. We find that both techniques are beneficial and suffice to reach performance that compares with more sophisticated systems from the 2020 task. We then present the application of this system to the 2021 task for low-resource supervised Upper Sorbian (HSB) to German translation, in both directions. Finally, we present a contrastive system for HSB-DE in both directions, and for unsupervised German to Lower Sorbian (DSB) translation, which uses multi-task training with various training schedules to improve over the baseline.

State-of-the-art multilingual systems rely on shared vocabularies that sufficiently cover all considered languages. To this end, a simple and frequently used approach makes use of subword vocabularies constructed jointly over several languages. We hypothesize that such vocabularies are suboptimal due to false positives (identical subwords with different meanings across languages) and false negatives (different subwords with similar meanings). To address these issues, we propose Subword Mapping and Anchoring across Languages (SMALA), a method to construct bilingual subword vocabularies. SMALA extracts subword alignments using an unsupervised state-of-the-art mapping technique and uses them to create cross-lingual anchors based on subword similarities. We demonstrate the benefits of SMALA for cross-lingual natural language inference (XNLI), where it improves zero-shot transfer to an unseen language without task-specific data, but only by sharing subword embeddings. Moreover, in neural machine translation, we show that joint subword vocabularies obtained with SMALA lead to higher BLEU scores on sentences that contain many false positives and false negatives.

Common acquisition functions for active learning use either uncertainty or diversity sampling, aiming to select difficult and diverse data points from the pool of unlabeled data, respectively. In this work, leveraging the best of both worlds, we propose an acquisition function that opts for selecting contrastive examples, i.e. data points that are similar in the model feature space and yet the model outputs maximally different predictive likelihoods. We compare our approach, CAL (Contrastive Active Learning), with a diverse set of acquisition functions in four natural language understanding tasks and seven datasets. Our experiments show that CAL performs consistently better or equal than the best performing baseline across all tasks, on both in-domain and out-of-domain data. We also conduct an extensive ablation study of our method and we further analyze all actively acquired datasets showing that CAL achieves a better trade-off between uncertainty and diversity compared to other strategies.

In Natural Language Processing (NLP), pretrained language models (LMs) that are transferred to downstream tasks have been recently shown to achieve state-of-the-art results. However, standard fine-tuning can degrade the general-domain representations captured during pretraining. To address this issue, we introduce a new regularization technique, AFTER; domain Adversarial Fine-Tuning as an Effective Regularizer. Specifically, we complement the task-specific loss used during fine-tuning with an adversarial objective. This additional loss term is related to an adversarial classifier, that aims to discriminate between in-domain and out-of-domain text representations. Indomain refers to the labeled dataset of the task at hand while out-of-domain refers to unlabeled data from a different domain. Intuitively, the adversarial classifier acts as a regularize which prevents the model from overfitting to the task-specific domain. Empirical results on various natural language understanding tasks show that AFTER leads to improved performance compared to standard fine-tuning.