Daniel Deutsch

Patrick Fernandes, Daniel Deutsch, Mara Finkelstein et al. · deepmind

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.

Sebastian Gehrmann, Abhik Bhattacharjee, Abinaya Mahendiran et al. · amazon-science, cmu

Evaluation in machine learning is usually informed by past choices, for example which datasets or metrics to use. This standardization enables the comparison on equal footing using leaderboards, but the evaluation choices become sub-optimal as better alternatives arise. This problem is especially pertinent in natural language generation which requires ever-improving suites of datasets, metrics, and human evaluation to make definitive claims. To make following best model evaluation practices easier, we introduce GEMv2. The new version of the Generation, Evaluation, and Metrics Benchmark introduces a modular infrastructure for dataset, model, and metric developers to benefit from each others work. GEMv2 supports 40 documented datasets in 51 languages. Models for all datasets can be evaluated online and our interactive data card creation and rendering tools make it easier to add new datasets to the living benchmark.

Wenda Xu, Daniel Deutsch, Mara Finkelstein et al. · cmu

Recent large language models (LLM) are leveraging human feedback to improve their generation quality. However, human feedback is costly to obtain, especially during inference. In this work, we propose LLMRefine, an inference time optimization method to refine LLM's output. The core idea is to use a learned fine-grained feedback model to pinpoint defects and guide LLM to refine them iteratively. Using original LLM as a proposal of edits, LLMRefine searches for defect-less text via simulated annealing, trading off the exploration and exploitation. We conduct experiments on three text generation tasks, including machine translation, long-form question answering (QA), and topical summarization. LLMRefine consistently outperforms all baseline approaches, achieving improvements up to 1.7 MetricX points on translation tasks, 8.1 ROUGE-L on ASQA, 2.2 ROUGE-L on topical summarization.

Daniel Deutsch, Dan Roth

We introduce Repro, an open-source library which aims at improving the reproducibility and usability of research code. The library provides a lightweight Python API for running software released by researchers within Docker containers which contain the exact required runtime configuration and dependencies for the code. Because the environment setup for each package is handled by Docker, users do not have to do any configuration themselves. Once Repro is installed, users can run the code for the 30+ papers currently supported by the library. We hope researchers see the value provided to others by including their research code in Repro and consider adding support for their own research code.

Lining Zhang, Simon Mille, Yufang Hou et al. · deepmind, uw

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.

Daniel Deutsch, Rotem Dror, Dan Roth

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.

Daniel Deutsch, Rotem Dror, Dan Roth

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.

Daniel Deutsch, Dan Roth

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.

Christoph Leiter, Juri Opitz, Daniel Deutsch et al.

With an increasing number of parameters and pre-training data, generative large language models (LLMs) 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. Within this context, we introduce the Eval4NLP 2023 shared task that asks participants to explore prompting and score extraction for machine translation (MT) and summarization evaluation. Specifically, we propose a novel competition setting in which we select a list of allowed LLMs and disallow fine-tuning to ensure a focus on prompting. We present an overview of participants' approaches and evaluate them on a new reference-free test set spanning three language pairs for MT and a summarization dataset. Notably, despite the task's restrictions, the best-performing systems achieve results on par with or even surpassing recent reference-free metrics developed using larger models, including GEMBA and Comet-Kiwi-XXL. Finally, as a separate track, we perform a small-scale human evaluation of the plausibility of explanations given by the LLMs.

Brian Thompson, Nitika Mathur, Daniel Deutsch et al.

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.

Jan-Thorsten Peter, David Vilar, Daniel Deutsch et al.

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.

Daniel Deutsch, Juraj Juraska, Mara Finkelstein et al.

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.

Gemma Team, Aishwarya Kamath, Johan Ferret et al. · deepmind, mit

We introduce Gemma 3, a multimodal addition to the Gemma family of lightweight open models, ranging in scale from 1 to 27 billion parameters. This version introduces vision understanding abilities, a wider coverage of languages and longer context - at least 128K tokens. We also change the architecture of the model to reduce the KV-cache memory that tends to explode with long context. This is achieved by increasing the ratio of local to global attention layers, and keeping the span on local attention short. The Gemma 3 models are trained with distillation and achieve superior performance to Gemma 2 for both pre-trained and instruction finetuned versions. In particular, our novel post-training recipe significantly improves the math, chat, instruction-following and multilingual abilities, making Gemma3-4B-IT competitive with Gemma2-27B-IT and Gemma3-27B-IT comparable to Gemini-1.5-Pro across benchmarks. We release all our models to the community.

Daniel Deutsch, Dan Roth

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.

Daniel Deutsch, Eleftheria Briakou, Isaac Caswell et al.

As large language models (LLM) become more and more capable in languages other than English, it is important to collect benchmark datasets in order to evaluate their multilingual performance, including on tasks like machine translation (MT). In this work, we extend the WMT24 dataset to cover 55 languages by collecting new human-written references and post-edits for 46 new languages and dialects in addition to post-edits of the references in 8 out of 9 languages in the original WMT24 dataset. The dataset covers four domains: literary, news, social, and speech. We benchmark a variety of MT providers and LLMs on the collected dataset using automatic metrics and find that LLMs are the best-performing MT systems in all 55 languages. These results should be confirmed using a human-based evaluation, which we leave for future work.

Geza Kovacs, Daniel Deutsch, Markus Freitag

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 improvements might simply be due to reward hacking rather than reflecting real quality improvements. In this work we find 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.

Parker Riley, Daniel Deutsch, George Foster et al.

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.

Marcel Nawrath, Agnieszka Nowak, Tristan Ratz et al. · amazon-science

At the heart of the Pyramid evaluation method for text summarization lie human written summary content units (SCUs). These SCUs are concise 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 strategies to 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 ranking short summaries, but may not help as much when ranking systems or longer summaries.

Muhammed Yusuf Kocyigit, Eleftheria Briakou, Daniel Deutsch et al. · mit

Data contamination -- the accidental consumption of evaluation examples within the pre-training data -- can undermine the validity of evaluation benchmarks. In this paper, we present a rigorous analysis of the effects of contamination on language models at 1B and 8B scales on the machine translation task. Starting from a carefully decontaminated train-test split, we systematically introduce contamination at various stages, scales, and data formats to isolate its effect and measure its impact on performance metrics. Our experiments reveal that contamination with both source and target substantially inflates BLEU scores, and this inflation is 2.5 times larger (up to 30 BLEU points) for 8B compared to 1B models. In contrast, source-only and target-only contamination generally produce smaller, less consistent over-estimations. Finally, we study how the temporal distribution and frequency of contaminated samples influence performance over-estimation across languages with varying degrees of data resources.

Zhongtao Liu, Parker Riley, Daniel Deutsch et al.

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, machineonly, 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.

Yixiao Song, Parker Riley, Daniel Deutsch et al.

Human evaluation is crucial for assessing rapidly evolving language models but is influenced by annotator proficiency and task design. This study explores the integration of comparative judgment into human annotation for machine translation (MT) and evaluates three annotation setups-point-wise Multidimensional Quality Metrics (MQM), side-by-side (SxS) MQM, and its simplified version SxS relative ranking (RR). In MQM, annotators mark error spans with categories and severity levels. SxS MQM extends MQM to pairwise error annotation for two translations of the same input, while SxS RR focuses on selecting the better output without labeling errors. Key findings are: (1) the SxS settings achieve higher inter-annotator agreement than MQM; (2) SxS MQM enhances inter-translation error marking consistency compared to MQM by, on average, 38.5% for explicitly compared MT systems and 19.5% for others; (3) all annotation settings return stable system rankings, with SxS RR offering a more efficient alternative to (SxS) MQM; (4) the SxS settings highlight subtle errors overlooked in MQM without altering absolute system evaluations. To spur further research, we will release the triply annotated datasets comprising 377 ZhEn and 104 EnDe annotation examples.

Mara Finkelstein, Isaac Caswell, Tobias Domhan et al.

We present TranslateGemma, a suite of open machine translation models based on the Gemma 3 foundation models. To enhance the inherent multilingual capabilities of Gemma 3 for the translation task, we employ a two-stage fine-tuning process. First, supervised fine-tuning is performed using a rich mixture of high-quality large-scale synthetic parallel data generated via state-of-the-art models and human-translated parallel data. This is followed by a reinforcement learning phase, where we optimize translation quality using an ensemble of reward models, including MetricX-QE and AutoMQM, targeting translation quality. We demonstrate the effectiveness of TranslateGemma with human evaluation on the WMT25 test set across 10 language pairs and with automatic evaluation on the WMT24++ benchmark across 55 language pairs. Automatic metrics show consistent and substantial gains over the baseline Gemma 3 models across all sizes. Notably, smaller TranslateGemma models often achieve performance comparable to larger baseline models, offering improved efficiency. We also show that TranslateGemma models retain strong multimodal capabilities, with enhanced performance on the Vistra image translation benchmark. The release of the open TranslateGemma models aims to provide the research community with powerful and adaptable tools for machine translation.

Senyu Li, Jiayi Wang, Felermino D. M. A. Ali et al.

Evaluating machine translation (MT) quality for under-resourced African languages remains a significant challenge, as existing metrics often suffer from limited language coverage and poor performance in low-resource settings. While recent efforts, such as AfriCOMET, have addressed some of the issues, they are still constrained by small evaluation sets, a lack of publicly available training data tailored to African languages, and inconsistent performance in extremely low-resource scenarios. In this work, we introduce SSA-MTE, a large-scale human-annotated MT evaluation (MTE) dataset covering 14 African language pairs from the News domain, with over 73,000 sentence-level annotations from a diverse set of MT systems. Based on this data, we develop SSA-COMET and SSA-COMET-QE, improved reference-based and reference-free evaluation metrics. We also benchmark prompting-based approaches using state-of-the-art LLMs like GPT-4o, Claude-3.7 and Gemini 2.5 Pro. Our experimental results show that SSA-COMET models significantly outperform AfriCOMET and are competitive with the strongest LLM Gemini 2.5 Pro evaluated in our study, particularly on low-resource languages such as Twi, Luo, and Yoruba. All resources are released under open licenses to support future research.

Juraj Juraska, Tobias Domhan, Mara Finkelstein et al.

In this paper, we present our submissions to the unified WMT25 Translation Evaluation Shared Task. For the Quality Score Prediction subtask, we create a new generation of MetricX with improvements in the input format and the training protocol, while for the Error Span Detection subtask we develop a new model, GemSpanEval, trained to predict error spans along with their severities and categories. Both systems are based on the state-of-the-art multilingual open-weights model Gemma 3, fine-tuned on publicly available WMT data. We demonstrate that MetricX-25, adapting Gemma 3 to an encoder-only architecture with a regression head on top, can be trained to effectively predict both MQM and ESA quality scores, and significantly outperforms its predecessor. Our decoder-only GemSpanEval model, on the other hand, we show to be competitive in error span detection with xCOMET, a strong encoder-only sequence-tagging baseline. With error span detection formulated as a generative task, we instruct the model to also output the context for each predicted error span, thus ensuring that error spans are identified unambiguously.

Vilém Zouhar, Wenda Xu, Parker Riley et al. · eth-zurich

Machine translation benchmarks sourced from the real world are quickly obsoleted, due to most examples being easy for state-of-the-art translation models. This limits the benchmark's ability to distinguish which model is better or to reveal models' weaknesses. Current methods for creating difficult test cases, such as subsampling or from-scratch synthesis, either fall short of identifying difficult examples or suffer from a lack of diversity and naturalness. Inspired by the iterative process of human experts probing for model failures, we propose MT-breaker, a method where a large language model iteratively refines a source text to increase its translation difficulty. The LLM iteratively queries a target machine translation model to guide its generation of difficult examples. Our approach generates examples that are more challenging for the target MT model while preserving the diversity of natural texts. While the examples are tailored to a particular machine translation model during the generation, the difficulty also transfers to other models and languages.

Parker Riley, Daniel Deutsch, Mara Finkelstein et al.

Human evaluation of machine translation is in an arms race with translation model quality: as our models get better, our evaluation methods need to be improved to ensure that quality gains are not lost in evaluation noise. To this end, we experiment with a two-stage version of the current state-of-the-art translation evaluation paradigm (MQM), which we call MQM re-annotation. In this setup, an MQM annotator reviews and edits a set of pre-existing MQM annotations, that may have come from themselves, another human annotator, or an automatic MQM annotation system. We demonstrate that rater behavior in re-annotation aligns with our goals, and that re-annotation results in higher-quality annotations, mostly due to finding errors that were missed during the first pass.

Wenda Xu, Vilém Zouhar, Parker Riley et al. · eth-zurich

NLP models require test data that are sufficiently challenging. The difficulty of an example is linked to the topic it originates from (''seed topic''). The relationship between the topic and the difficulty of its instances is stochastic in nature: an example about a difficult topic can happen to be easy, and vice versa. At the scale of the Internet, there are tens of thousands of potential topics, and finding the most difficult one by drawing and evaluating a large number of examples across all topics is computationally infeasible. We formalize this task and treat it as a multi-armed bandit problem. In this framework, each topic is an ''arm,'' and pulling an arm (at a cost) involves drawing a single example, evaluating it, and measuring its difficulty. The goal is to efficiently identify the most difficult topics within a fixed computational budget. We illustrate the bandit problem setup of finding difficult examples for the task of machine translation. We find that various bandit strategies vastly outperform baseline methods like brute-force searching the most challenging topics.

Wenda Xu, Sweta Agrawal, Vilém Zouhar et al. · eth-zurich

As large language models (LLMs) begin to saturate existing benchmarks, automated benchmark creation using LLMs (LLM as a benchmark) has emerged as a scalable alternative to slow and costly human curation. While these generated test sets have to potential to cheaply rank models, we demonstrate a critical flaw. LLM generated benchmarks systematically favor the model that created the benchmark, they exhibit self bias on low resource languages to English translation tasks. We show three key findings on automatic benchmarking of LLMs for translation: First, this bias originates from two sources: the generated test data (LLM as a testset) and the evaluation method (LLM as an evaluator), with their combination amplifying the effect. Second, self bias in LLM as a benchmark is heavily influenced by the model's generation capabilities in the source language. For instance, we observe more pronounced bias in into English translation, where the model's generation system is developed, than in out of English translation tasks. Third, we observe that low diversity in source text is one attribution to self bias. Our results suggest that improving the diversity of these generated source texts can mitigate some of the observed self bias.

Colten DiIanni, Daniel Deutsch

This paper introduces Pairwise Difference Pearson (PDP), a novel segment-level meta-evaluation metric for Machine Translation (MT) that address limitations in previous Pearson's $ρ$-based and and Kendall's $τ$-based meta-evaluation approaches. PDP is a correlation-based metric that utilizes pairwise differences rather than raw scores. It draws on information from all segments for a more robust understanding of score distributions and uses segment-wise pairwise differences to refine Global Pearson to intra-segment score comparisons. Analysis on the WMT'24 shared task shows PDP properly ranks sentinel evaluation metrics and better aligns with human error weightings than previous work. Noise injection analysis demonstrates PDP's robustness to random noise, segment bias, and system bias while highlighting its sensitivity to extreme outliers.

Daniel Deutsch, George Foster, Markus Freitag

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.

Daniel Deutsch, Dan Roth

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.

Daniel Deutsch, Rotem Dror, Dan Roth

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 in some evaluation settings.

Daniel Deutsch, Dan Roth

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, but rather the extent to which they 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 it means the summarization community has not yet found a reliable automatic metric that aligns with its research goal, to generate summaries with high-quality information. Then, we propose a simple and interpretable method of evaluating summaries which does directly measure information overlap and demonstrate how it can be used to gain insights into model behavior that could not be provided by other methods alone.

Daniel Deutsch, Tania Bedrax-Weiss, Dan Roth

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 out-performs 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.