Recent large language models (LLM) areleveraging human feedback to improve theirgeneration quality. However, human feedbackis costly to obtain, especially during inference.In this work, we propose LLMRefine, aninference time optimization method to refineLLM’s output. The core idea is to usea learned fine-grained feedback model topinpoint defects and guide LLM to refinethem iteratively. Using original LLM as aproposal of edits, LLMRefine searches fordefect-less text via simulated annealing, tradingoff the exploration and exploitation. Weconduct experiments on three text generationtasks, including machine translation, long-form question answering (QA), and topicalsummarization. LLMRefine consistentlyoutperforms all baseline approaches, achievingimprovements up to 1.7 MetricX points ontranslation tasks, 8.1 ROUGE-L on ASQA, 2.2ROUGE-L on topical summarization.

Reliable human evaluation is critical to the development of successful natural language generation models, but achieving it is notoriously difficult. Stability is a crucial requirement when ranking systems by quality: consistent ranking of systems across repeated evaluations is not just desirable, but essential. Without it, there is no reliable foundation for hill-climbing or product launch decisions. In this paper, we use machine translation and its state-of-the-art human evaluation framework, MQM, as a case study to understand how to set up reliable human evaluations that yield stable conclusions. We investigate the optimal configurations for item allocation to raters, number of ratings per item, and score normalization. Our study on two language pairs provides concrete recommendations for designing replicable human evaluation studies. We also collect and release the largest publicly available dataset of multi-segment translations rated by multiple professional translators, consisting of nearly 140,000 segment annotations across two language pairs.

At the heart of the Pyramid evaluation method for text summarization lie human written summary content units (SCUs). These SCUs areconcise sentences that decompose a summary into small facts. Such SCUs can be used to judge the quality of a candidate summary, possibly partially automated via natural language inference (NLI) systems. Interestingly, with the aim to fully automate the Pyramid evaluation, Zhang and Bansal (2021) show that SCUs can be approximated by automatically generated semantic role triplets (STUs). However, several questions currently lack answers, in particular: i) Are there other ways of approximating SCUs that can offer advantages?ii) Under which conditions are SCUs (or their approximations) offering the most value? In this work, we examine two novel strategiesto approximate SCUs: generating SCU approximations from AMR meaning representations (SMUs) and from large language models (SGUs), respectively. We find that while STUs and SMUs are competitive, the best approximation quality is achieved by SGUs. We also show through a simple sentence-decomposition baseline (SSUs) that SCUs (and their approximations) offer the most value when rankingshort summaries, but may not help as much when ranking systems or longer summaries.

The WMT24 Metrics Shared Task evaluated the performance of automatic metrics for machine translation (MT), with a major focus on LLM-based translations that were generated as part of the WMT24 General MT Shared Task. As LLMs become increasingly popular in MT, it is crucial to determine whether existing evaluation metrics can accurately assess the output of these systems.To provide a robust benchmark for this evaluation, human assessments were collected using Multidimensional Quality Metrics (MQM), continuing the practice from recent years. Furthermore, building on the success of the previous year, a challenge set subtask was included, requiring participants to design contrastive test suites that specifically target a metric’s ability to identify and penalize different types of translation errors.Finally, the meta-evaluation procedure was refined to better reflect real-world usage of MT metrics, focusing on pairwise accuracy at both the system- and segment-levels.We present an extensive analysis on how well metrics perform on three language pairs: English to Spanish (Latin America), Japanese to Chinese, and English to German. The results strongly confirm the results reported last year, that fine-tuned neural metrics continue to perform well, even when used to evaluate LLM-based translation systems.

In this paper, we present the MetricX-24 submissions to the WMT24 Metrics Shared Task and provide details on the improvements we made over the previous version of MetricX. Our primary submission is a hybrid reference-based/-free metric, which can score a translation irrespective of whether it is given the source segment, the reference, or both. The metric is trained on previous WMT data in a two-stage fashion, first on the DA ratings only, then on a mixture of MQM and DA ratings. The training set in both stages is augmented with synthetic examples that we created to make the metric more robust to several common failure modes, such as fluent but unrelated translation, or undertranslation. We demonstrate the benefits of the individual modifications via an ablation study, and show a significant performance increase over MetricX-23 on the WMT23 MQM ratings, as well as our new synthetic challenge set.

While Minimum Bayes Risk (MBR) decoding using metrics such as COMET or MetricX has outperformed traditional decoding methods such as greedy or beam search, it introduces a challenge we refer to as metric bias. As MBR decoding aims to produce translations that score highly according to a specific utility metric, this very process makes it impossible to use the same metric for both decoding and evaluation, as any improvement might simply be due to reward hacking rather than reflecting real quality improvements. In this work we demonstrate that compared to human ratings, neural metrics not only overestimate the quality of MBR decoding when the same metric is used as the utility metric, but they also overestimate the quality of MBR/QE decoding with other neural utility metrics as well. We also show that the metric bias issue can be mitigated by using an ensemble of utility metrics during MBR decoding: human evaluations show that MBR decoding using an ensemble of utility metrics outperforms a single utility metric.

Collecting high-quality translations is crucial for the development and evaluation of machine translation systems. However, traditional human-only approaches are costly and slow. This study presents a comprehensive investigation of 11 approaches for acquiring translation data, including human-only, machine-only, and hybrid approaches. Our findings demonstrate that human-machine collaboration can match or even exceed the quality of human-only translations, while being more cost-efficient. Error analysis reveals the complementary strengths between human and machine contributions, highlighting the effectiveness of collaborative methods. Cost analysis further demonstrates the economic benefits of human-machine collaboration methods, with some approaches achieving top-tier quality at around 60% of the cost of traditional methods. We release a publicly available dataset containing nearly 18,000 segments of varying translation quality with corresponding human ratings to facilitate future research.

Selecting an automatic metric that best emulates human annotators is often non-trivial, because there is no clear definition of “best emulates.” A meta-metric is required to compare the human judgments to the automatic metric scores, and metric rankings depend on the choice of meta-metric. We propose Soft Pairwise Accuracy (SPA), a new meta-metric that builds on Pairwise Accuracy (PA) but incorporates the statistical significance of both the human judgments and the metric scores. We show that SPA is more stable than PA with respect to changes in the number of systems/segments used for evaluation. We also show that PA can only assign a small set of distinct output values to metrics, and this results in many metrics being artificially assigned the exact same PA score. We demonstrate that SPA fixes this issue. Finally, we show that SPA is more discriminative than PA, producing more statistically significant comparisons between metrics. SPA was selected as the official system-level metric for the 2024 WMT Metrics Shared Task.

To prevent the costly and inefficient use of resources on low-quality annotations, we want a method for creating a pool of dependable annotators who can effectively complete difficult tasks, such as evaluating automatic summarization. Thus, we investigate the recruitment of high-quality Amazon Mechanical Turk workers via a two-step pipeline. We show that we can successfully filter out subpar workers before they carry out the evaluations and obtain high-agreement annotations with similar constraints on resources. Although our workers demonstrate a strong consensus among themselves and CloudResearch workers, their alignment with expert judgments on a subset of the data is not as expected and needs further training in correctness. This paper still serves as a best practice for the recruitment of qualified annotators in other challenging annotation tasks.

Kendall’s tau is frequently used to meta-evaluate how well machine translation (MT) evaluation metrics score individual translations. Its focus on pairwise score comparisons is intuitive but raises the question of how ties should be handled, a gray area that has motivated different variants in the literature. We demonstrate that, in settings like modern MT meta-evaluation, existing variants have weaknesses arising from their handling of ties, and in some situations can even be gamed. We propose instead to meta-evaluate metrics with a version of pairwise accuracy that gives metrics credit for correctly predicting ties, in combination with a tie calibration procedure that automatically introduces ties into metric scores, enabling fair comparison between metrics that do and do not predict ties. We argue and provide experimental evidence that these modifications lead to fairer ranking-based assessments of metric performance.

In this work, we propose a method for incorporating question-answering (QA) signals into a summarization model. Our method identifies salient noun phrases (NPs) in the input document by automatically generating wh-questions that are answered by the NPs and automatically determining whether those questions are answered in the gold summaries. This QA-based signal is incorporated into a two-stage summarization model which first marks salient NPs in the input document using a classification model, then conditionally generates a summary. Our experiments demonstrate that the models trained using QA-based supervision generate higher-quality summaries than baseline methods of identifying salient spans on benchmark summarization datasets. Further, we show that the content of the generated summaries can be controlled based on which NPs are marked in the input document. Finally, we propose a method of augmenting the training data so the gold summaries are more consistent with the marked input spans used during training and show how this results in models which learn to better exclude unmarked document content.

Generative large language models (LLMs) have seen many breakthroughs over the last year. With an increasing number of parameters and pre-training data, they have shown remarkable capabilities to solve tasks with minimal or no task-related examples. Notably, LLMs have been successfully employed as evaluation metrics in text generation tasks. Strategies employed in this context differ in the choice of input prompts, the selection of samples for demonstration, and the methodology used to construct scores grading the generations. Approaches often differ in the input prompts, the samples that are selected for demonstration and the construction process of scores from the output. Within this context, we introduce the Eval4NLP 2023 shared task that asks participants to explore such approaches for machine translation evaluation and summarization eval- uation. Specifically, we select a list of allowed LLMs and disallow fine-tuning to ensure a focus on prompting. We test the approaches of the participants on a new reference-free test-set spanning 3 language pairs for machine transla- tion as well as a summarization dataset. Further, we present an overview of the approaches taken by the participants, present their results on the test set and analyze paths for future work. Fi- nally, as a separate track, we perform a human evaluation of the plausibility of explanations given by the LLMs and its effect on model performance. We make parts of our code and datasets available.

Quality Estimation (QE), the evaluation of machine translation output without the need of explicit references, has seen big improvements in the last years with the use of neural metrics. In this paper we analyze the viability of using QE metrics for filtering out bad quality sentence pairs in the training data of neural machine translation systems (NMT). While most corpus filtering methods are focused on detecting noisy examples in collections of texts, usually huge amounts of web crawled data, QE models are trained to discriminate more fine-grained quality differences. We show that by selecting the highest quality sentence pairs in the training data, we can improve translation quality while reducing the training size by half. We also provide a detailed analysis of the filtering results, which highlights the differences between both approaches.

This paper presents the results of the WMT23 Metrics Shared Task. Participants submitting automatic MT evaluation metrics were asked to score the outputs of the translation systems competing in the WMT23 News Translation Task. All metrics were evaluated on how well they correlate with human ratings at the system and segment level. Similar to last year, we acquired our own human ratings based on expert-based human evaluation via Multidimensional Quality Metrics (MQM). Following last year’s success, we also included a challenge set subtask, where participants had to create contrastive test suites for evaluating metrics’ ability to capture and penalise specific types of translation errors. Furthermore, we improved our meta-evaluation procedure by considering fewer tasks and calculating a global score by weighted averaging across the various tasks. We present an extensive analysis on how well metrics perform on three language pairs: Chinese-English, Hebrew-English on the sentence-level and English-German on the paragraph-level. The results strongly confirm the results reported last year, that neural-based metrics are significantly better than non-neural metrics in their levels of correlation with human judgments. Further, we investigate the impact of bad reference translations on the correlations of metrics with human judgment. We present a novel approach for generating synthetic reference translations based on the collection of MT system outputs and their corresponding MQM ratings, which has the potential to mitigate bad reference issues we observed this year for some language pairs. Finally, we also study the connections between the magnitude of metric differences and their expected significance in human evaluation, which should help the community to better understand and adopt new metrics.

This report details the MetricX-23 submission to the WMT23 Metrics Shared Task and provides an overview of the experiments that informed which metrics were submitted. Our 3 submissions—each with a quality estimation (or reference-free) version—are all learned regression-based metrics that vary in the data used for training and which pretrained language model was used for initialization. We report results related to understanding (1) which supervised training data to use, (2) the impact of how the training labels are normalized, (3) the amount of synthetic training data to use, (4) how metric performance is related to model size, and (5) the effect of initializing the metrics with different pretrained language models. The most successful training recipe for MetricX employs two-stage fine-tuning on DA and MQM ratings, and includes synthetic training data. Finally, one important takeaway from our extensive experiments is that optimizing for both segment- and system-level performance at the same time is a challenging task.

This report describes the Minimum Bayes Risk Quality Estimation (MBR-QE) submission to the Workshop on Machine Translation’s 2023 Metrics Shared Task. MBR decoding with neural utility metrics like BLEURT is known to be effective in generating high quality machine translations. We use the underlying technique of MBR decoding and develop an MBR based reference-free quality estimation metric. Our method uses an evaluator machine translation system and a reference-based utility metric (specifically BLEURT and MetricX) to calculate a quality estimation score of a model. We report results related to comparing different MBR configurations and utility metrics.

As research on machine translation moves to translating text beyond the sentence level, it remains unclear how effective automatic evaluation metrics are at scoring longer translations. In this work, we first propose a method for creating paragraph-level data for training and meta-evaluating metrics from existing sentence-level data. Then, we use these new datasets to benchmark existing sentence-level metrics as well as train learned metrics at the paragraph level. Interestingly, our experimental results demonstrate that using sentence-level metrics to score entire paragraphs is equally as effective as using a metric designed to work at the paragraph level. We speculate this result can be attributed to properties of the task of reference-based evaluation as well as limitations of our datasets with respect to capturing all types of phenomena that occur in paragraph-level translations.

Automatic evaluation of machine translation (MT) is a critical tool driving the rapid iterative development of MT systems. While considerable progress has been made on estimating a single scalar quality score, current metrics lack the informativeness of more detailed schemes that annotate individual errors, such as Multidimensional Quality Metrics (MQM). In this paper, we help fill this gap by proposing AutoMQM, a prompting technique which leverages the reasoning and in-context learning capabilities of large language models (LLMs) and asks them to identify and categorize errors in translations. We start by evaluating recent LLMs, such as PaLM and PaLM-2, through simple score prediction prompting, and we study the impact of labeled data through in-context learning and finetuning. We then evaluate AutoMQM with PaLM-2 models, and we find that it improves performance compared to just prompting for scores (with particularly large gains for larger models) while providing interpretability through error spans that align with human annotations.

How reliably an automatic summarization evaluation metric replicates human judgments of summary quality is quantified by system-level correlations. We identify two ways in which the definition of the system-level correlation is inconsistent with how metrics are used to evaluate systems in practice and propose changes to rectify this disconnect. First, we calculate the system score for an automatic metric using the full test set instead of the subset of summaries judged by humans, which is currently standard practice. We demonstrate how this small change leads to more precise estimates of system-level correlations. Second, we propose to calculate correlations only on pairs of systems that are separated by small differences in automatic scores which are commonly observed in practice. This allows us to demonstrate that our best estimate of the correlation of ROUGE to human judgments is near 0 in realistic scenarios. The results from the analyses point to the need to collect more high-quality human judgments and to improve automatic metrics when differences in system scores are small.

There is significant interest in developing evaluation metrics which accurately estimate the quality of generated text without the aid of a human-written reference text, which can be time consuming and expensive to collect or entirely unavailable in online applications. However, in this work, we demonstrate that these reference-free metrics are inherently biased and limited in their ability to evaluate generated text, and we argue that they should not be used to measure progress on tasks like machine translation or summarization. We show how reference-free metrics are equivalent to using one generation model to evaluate another, which has several limitations: (1) the metrics can be optimized at test time to find the approximate best-possible output, (2) they are inherently biased toward models which are more similar to their own, and (3) they can be biased against higher-quality outputs, including those written by humans. Therefore, we recommend that reference-free metrics should be used as diagnostic tools for analyzing and understanding model behavior instead of measures of how well models perform a task, in which the goal is to achieve as high of a score as possible.

Evaluations in machine learning rarely use the latest metrics, datasets, or human evaluation in favor of remaining compatible with prior work. The compatibility, often facilitated through leaderboards, thus leads to outdated but standardized evaluation practices. We pose that the standardization is taking place in the wrong spot. Evaluation infrastructure should enable researchers to use the latest methods and what should be standardized instead is how to incorporate these new evaluation advances. We introduce GEMv2, the new version of the Generation, Evaluation, and Metrics Benchmark which uses a modular infrastructure for dataset, model, and metric developers to benefit from each other’s work. GEMv2 supports 40 documented datasets in 51 languages, ongoing online evaluation for all datasets, and our interactive tools make it easier to add new datasets to the living benchmark.

Question answering-based summarization evaluation metrics must automatically determine whether the QA model’s prediction is correct or not, a task known as answer verification. In this work, we benchmark the lexical answer verification methods which have been used by current QA-based metrics as well as two more sophisticated text comparison methods, BERTScore and LERC. We find that LERC out-performs the other methods in some settings while remaining statistically indistinguishable from lexical overlap in others. However, our experiments reveal that improved verification performance does not necessarily translate to overall QA-based metric quality: In some scenarios, using a worse verification method — or using none at all — has comparable performance to using the best verification method, a result that we attribute to properties of the datasets.

A desirable property of a reference-based evaluation metric that measures the content quality of a summary is that it should estimate how much information that summary has in common with a reference. Traditional text overlap based metrics such as ROUGE fail to achieve this because they are limited to matching tokens, either lexically or via embeddings. In this work, we propose a metric to evaluate the content quality of a summary using question-answering (QA). QA-based methods directly measure a summary’s information overlap with a reference, making them fundamentally different than text overlap metrics. We demonstrate the experimental benefits of QA-based metrics through an analysis of our proposed metric, QAEval. QAEval outperforms current state-of-the-art metrics on most evaluations using benchmark datasets, while being competitive on others due to limitations of state-of-the-art models. Through a careful analysis of each component of QAEval, we identify its performance bottlenecks and estimate that its potential upper-bound performance surpasses all other automatic metrics, approaching that of the gold-standard Pyramid Method.1

The quality of a summarization evaluation metric is quantified by calculating the correlation between its scores and human annotations across a large number of summaries. Currently, it is unclear how precise these correlation estimates are, nor whether differences between two metrics’ correlations reflect a true difference or if it is due to mere chance. In this work, we address these two problems by proposing methods for calculating confidence intervals and running hypothesis tests for correlations using two resampling methods, bootstrapping and permutation. After evaluating which of the proposed methods is most appropriate for summarization through two simulation experiments, we analyze the results of applying these methods to several different automatic evaluation metrics across three sets of human annotations. We find that the confidence intervals are rather wide, demonstrating high uncertainty in the reliability of automatic metrics. Further, although many metrics fail to show statistical improvements over ROUGE, two recent works, QAEval and BERTScore, do so in some evaluation settings.1

Reference-based metrics such as ROUGE or BERTScore evaluate the content quality of a summary by comparing the summary to a reference. Ideally, this comparison should measure the summary’s information quality by calculating how much information the summaries have in common. In this work, we analyze the token alignments used by ROUGE and BERTScore to compare summaries and argue that their scores largely cannot be interpreted as measuring information overlap. Rather, they are better estimates of the extent to which the summaries discuss the same topics. Further, we provide evidence that this result holds true for many other summarization evaluation metrics. The consequence of this result is that the most frequently used summarization evaluation metrics do not align with the community’s research goal, to generate summaries with high-quality information. However, we conclude by demonstrating that a recently proposed metric, QAEval, which scores summaries using question-answering, appears to better capture information quality than current evaluations, highlighting a direction for future research.

Identifying the key events in a document is critical to holistically understanding its important information. Although measuring the salience of events is highly contextual, most previous work has used a limited representation of events that omits essential information. In this work, we propose a highly contextual model of event salience that uses a rich representation of events, incorporates document-level information and allows for interactions between latent event encodings. Our experimental results on an event salience dataset demonstrate that our model improves over previous work by an absolute 2-4% on standard metrics, establishing a new state-of-the-art performance for the task. We also propose a new evaluation metric that addresses flaws in previous evaluation methodologies. Finally, we discuss the importance of salient event detection for the downstream task of summarization.

We present SacreROUGE, an open-source library for using and developing summarization evaluation metrics. SacreROUGE removes many obstacles that researchers face when using or developing metrics: (1) The library provides Python wrappers around the official implementations of existing evaluation metrics so they share a common, easy-to-use interface; (2) it provides functionality to evaluate how well any metric implemented in the library correlates to human-annotated judgments, so no additional code needs to be written for a new evaluation metric; and (3) it includes scripts for loading datasets that contain human judgments so they can easily be used for evaluation. This work describes the design of the library, including the core Metric interface, the command-line API for evaluating summarization models and metrics, and the scripts to load and reformat publicly available datasets. The development of SacreROUGE is ongoing and open to contributions from the community.

Inference in structured prediction involves finding the best output structure for an input, subject to certain constraints. Many current approaches use sequential inference, which constructs the output in a left-to-right manner. However, there is no general framework to specify constraints in these approaches. We present a principled approach for incorporating constraints into sequential inference algorithms. Our approach expresses constraints using an automaton, which is traversed in lock-step during inference, guiding the search to valid outputs. We show that automata can express commonly used constraints and are easily incorporated into sequential inference. When it is more natural to represent constraints as a set of automata, our algorithm uses an active set method for demonstrably fast and efficient inference. We experimentally show the benefits of our algorithm on constituency parsing and semantic role labeling. For parsing, unlike unconstrained approaches, our algorithm always generates valid output, incurring only a small drop in performance. For semantic role labeling, imposing constraints using our algorithm corrects common errors, improving F1 by 1.5 points. These benefits increase in low-resource settings. Our active set method achieves a 5.2x relative speed-up over a naive approach.

A key challenge in topic-focused summarization is determining what information should be included in the summary, a problem known as content selection. In this work, we propose a new method for studying content selection in topic-focused summarization called the summary cloze task. The goal of the summary cloze task is to generate the next sentence of a summary conditioned on the beginning of the summary, a topic, and a reference document(s). The main challenge is deciding what information in the references is relevant to the topic and partial summary and should be included in the summary. Although the cloze task does not address all aspects of the traditional summarization problem, the more narrow scope of the task allows us to collect a large-scale datset of nearly 500k summary cloze instances from Wikipedia. We report experimental results on this new dataset using various extractive models and a two-step abstractive model that first extractively selects a small number of sentences and then abstractively summarizes them. Our results show that the topic and partial summary help the models identify relevant content, but the task remains a significant challenge.

Modeling derivational morphology to generate words with particular semantics is useful in many text generation tasks, such as machine translation or abstractive question answering. In this work, we tackle the task of derived word generation. That is, we attempt to generate the word “runner” for “someone who runs.” We identify two key problems in generating derived words from root words and transformations. We contribute a novel aggregation model of derived word generation that learns derivational transformations both as orthographic functions using sequence-to-sequence models and as functions in distributional word embedding space. The model then learns to choose between the hypothesis of each system. We also present two ways of incorporating corpus information into derived word generation.