Juraj Juraska

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.

Omkar Patil, Lena Reed, Kevin K. Bowden et al. · cmu

Conversational agents are consistently growing in popularity and many people interact with them every day. While many conversational agents act as personal assistants, they can have many different goals. Some are task-oriented, such as providing customer support for a bank or making a reservation. Others are designed to be empathetic and to form emotional connections with the user. The Alexa Prize Challenge aims to create a socialbot, which allows the user to engage in coherent conversations, on a range of popular topics that will interest the user. Here we describe Athena 2.0, UCSC's conversational agent for Amazon's Socialbot Grand Challenge 4. Athena 2.0 utilizes a novel knowledge-grounded discourse model that tracks the entity links that Athena introduces into the dialogue, and uses them to constrain named-entity recognition and linking, and coreference resolution. Athena 2.0 also relies on a user model to personalize topic selection and other aspects of the conversation to individual users.

Angela Ramirez, Karik Agarwal, Juraj Juraska et al.

Dialogue systems need to produce responses that realize multiple types of dialogue acts (DAs) with high semantic fidelity. In the past, natural language generators (NLGs) for dialogue were trained on large parallel corpora that map from a domain-specific DA and its semantic attributes to an output utterance. Recent work shows that pretrained language models (LLMs) offer new possibilities for controllable NLG using prompt-based learning. Here we develop a novel few-shot overgenerate-and-rank approach that achieves the controlled generation of DAs. We compare eight few-shot prompt styles that include a novel method of generating from textual pseudo-references using a textual style transfer approach. We develop six automatic ranking functions that identify outputs with both the correct DA and high semantic accuracy at generation time. We test our approach on three domains and four LLMs. To our knowledge, this is the first work on NLG for dialogue that automatically ranks outputs using both DA and attribute accuracy. For completeness, we compare our results to fine-tuned few-shot models trained with 5 to 100 instances per DA. Our results show that several prompt settings achieve perfect DA accuracy, and near perfect semantic accuracy (99.81%) and perform better than few-shot fine-tuning.

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.

Hamid Dadkhahi, Firas Trabelsi, Parker Riley et al.

Thinking Large Language Models (LLMs) used as judges for pairwise preferences remain noisy at the single-sample level, and common aggregation rules (majority vote, soft self-consistency, or instruction-based self-aggregation) are inconsistent when ties are allowed. We study inference-time compute (ITC) for evaluators that generate n independent thinking-rating samples per item, and propose a principled, distribution-calibrated aggregation scheme. Our method models three-way preferences with a Bradley-Terry-Davidson formulation on rating counts, leveraging both polarity (margin among non-ties) and decisiveness (non-tie rate) to distinguish narrow margins from strong consensus. Across various evaluation benchmarks, our approach consistently reduces MAE and increases pairwise accuracy versus standard baselines, and when evaluated against human-consensus meta-labels, matches or exceeds individual human raters. These results show that carefully allocating ITC and aggregating with distribution-aware methods turns noisy individual model judgments into reliable ratings for evaluation.

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.

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.

Mara Finkelstein, Dan Deutsch, Parker Riley et al.

As LLMs continue to become more powerful and versatile, human evaluation has quickly become intractable at scale and reliance on automatic metrics has become the norm. Recently, it has been shown that LLMs are themselves state-of-the-art evaluators for many tasks. These Autoraters are typically designed so that they generalize to new systems and test sets. In practice, however, evaluation is performed on a small set of fixed, canonical test sets, which are carefully curated to measure certain capabilities of interest and are not changed frequently. In this work, we design a method which specializes a prompted Autorater to a given test set, by leveraging historical ratings on the test set to construct in-context learning (ICL) examples. We evaluate our Specialist method on the task of fine-grained machine translation evaluation, and show that it dramatically outperforms the state-of-the-art XCOMET metric by 54% and 119% on the WMT'23 and WMT'24 test sets, respectively. We perform extensive analyses to understand the representations learned by our Specialist metrics, and how variability in rater behavior affects their performance. We also verify the generalizability and robustness of our Specialist method for designing automatic metrics across different numbers of ICL examples, LLM backbones, systems to evaluate, and evaluation tasks.

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.

Behzad Shayegh, Jan-Thorsten Peter, David Vilar et al.

We investigate the tradeoff between adequacy and fluency in machine translation. We show the severity of this tradeoff at the evaluation level and analyze where popular metrics fall within it. Essentially, current metrics generally lean toward adequacy, meaning that their scores correlate more strongly with the adequacy of translations than with fluency. More importantly, we find that this tradeoff also persists at the meta-evaluation level, and that the standard WMT meta-evaluation favors adequacy-oriented metrics over fluency-oriented ones. We show that this bias is partially attributed to the composition of the systems included in the meta-evaluation datasets. To control this bias, we propose a method that synthesizes translation systems in meta-evaluation. Our findings highlight the importance of understanding this tradeoff in meta-evaluation and its impact on metric rankings.

Juraj Juraska, Kevin K. Bowden, Lena Reed et al.

Athena 2.0 is an Alexa Prize SocialBot that has been a finalist in the last two Alexa Prize Grand Challenges. One reason for Athena's success is its novel dialogue management strategy, which allows it to dynamically construct dialogues and responses from component modules, leading to novel conversations with every interaction. Here we describe Athena's system design and performance in the Alexa Prize during the 20/21 competition. A live demo of Athena as well as video recordings will provoke discussion on the state of the art in conversational AI.

Juraj Juraska, Marilyn Walker

Ever since neural models were adopted in data-to-text language generation, they have invariably been reliant on extrinsic components to improve their semantic accuracy, because the models normally do not exhibit the ability to generate text that reliably mentions all of the information provided in the input. In this paper, we propose a novel decoding method that extracts interpretable information from encoder-decoder models' cross-attention, and uses it to infer which attributes are mentioned in the generated text, which is subsequently used to rescore beam hypotheses. Using this decoding method with T5 and BART, we show on three datasets its ability to dramatically reduce semantic errors in the generated outputs, while maintaining their state-of-the-art quality.

Vrindavan Harrison, Juraj Juraska, Wen Cui et al.

This report describes Athena, a dialogue system for spoken conversation on popular topics and current events. We develop a flexible topic-agnostic approach to dialogue management that dynamically configures dialogue based on general principles of entity and topic coherence. Athena's dialogue manager uses a contract-based method where discourse constraints are dispatched to clusters of response generators. This allows Athena to procure responses from dynamic sources, such as knowledge graph traversals and feature-based on-the-fly response retrieval methods. After describing the dialogue system architecture, we perform an analysis of conversations that Athena participated in during the 2019 Alexa Prize Competition. We conclude with a report on several user studies we carried out to better understand how individual user characteristics affect system ratings.

Juraj Juraska, Kevin K. Bowden, Marilyn Walker

The uptake of deep learning in natural language generation (NLG) led to the release of both small and relatively large parallel corpora for training neural models. The existing data-to-text datasets are, however, aimed at task-oriented dialogue systems, and often thus limited in diversity and versatility. They are typically crowdsourced, with much of the noise left in them. Moreover, current neural NLG models do not take full advantage of large training data, and due to their strong generalizing properties produce sentences that look template-like regardless. We therefore present a new corpus of 7K samples, which (1) is clean despite being crowdsourced, (2) has utterances of 9 generalizable and conversational dialogue act types, making it more suitable for open-domain dialogue systems, and (3) explores the domain of video games, which is new to dialogue systems despite having excellent potential for supporting rich conversations.

Kevin K. Bowden, Jiaqi Wu, Wen Cui et al.

Conversational systems typically focus on functional tasks such as scheduling appointments or creating todo lists. Instead we design and evaluate SlugBot (SB), one of 8 semifinalists in the 2018 AlexaPrize, whose goal is to support casual open-domain social inter-action. This novel application requires both broad topic coverage and engaging interactive skills. We developed a new technical approach to meet this demanding situation by crowd-sourcing novel content and introducing playful conversational strategies based on storytelling and games. We collected over 10,000 conversations during August 2018 as part of the Alexa Prize competition. We also conducted an in-lab follow-up qualitative evaluation. Over-all users found SB moderately engaging; conversations averaged 3.6 minutes and involved 26 user turns. However, users reacted very differently to different conversation subtypes. Storytelling and games were evaluated positively; these were seen as entertaining with predictable interactive structure. They also led users to impute personality and intelligence to SB. In contrast, search and general Chit-Chat induced coverage problems; here users found it hard to infer what topics SB could understand, with these conversations seen as being too system-driven. Theoretical and design implications suggest a move away from conversational systems that simply provide factual information. Future systems should be designed to have their own opinions with personal stories to share, and SB provides an example of how we might achieve this.

Kevin K. Bowden, Jiaqi Wu, Wen Cui et al.

One of the most interesting aspects of the Amazon Alexa Prize competition is that the framing of the competition requires the development of new computational models of dialogue and its structure. Traditional computational models of dialogue are of two types: (1) task-oriented dialogue, supported by AI planning models,or simplified planning models consisting of frames with slots to be filled; or (2)search-oriented dialogue where every user turn is treated as a search query that may elaborate and extend current search results. Alexa Prize dialogue systems such as SlugBot must support conversational capabilities that go beyond what these traditional models can do. Moreover, while traditional dialogue systems rely on theoretical computational models, there are no existing computational theories that circumscribe the expected system and user behaviors in the intended conversational genre of the Alexa Prize Bots. This paper describes how UCSC's SlugBot team has combined the development of a novel computational theoretical model, Discourse Relation Dialogue Model, with its implementation in a modular system in order to test and refine it. We highlight how our novel dialogue model has led us to create a novel ontological resource, UniSlug, and how the structure of UniSlug determine show we curate and structure content so that our dialogue manager implements and tests our novel computational dialogue model.

Juraj Juraska, Marilyn Walker

One of the biggest challenges of end-to-end language generation from meaning representations in dialogue systems is making the outputs more natural and varied. Here we take a large corpus of 50K crowd-sourced utterances in the restaurant domain and develop text analysis methods that systematically characterize types of sentences in the training data. We then automatically label the training data to allow us to conduct two kinds of experiments with a neural generator. First, we test the effect of training the system with different stylistic partitions and quantify the effect of smaller, but more stylistically controlled training data. Second, we propose a method of labeling the style variants during training, and show that we can modify the style of the generated utterances using our stylistic labels. We contrast and compare these methods that can be used with any existing large corpus, showing how they vary in terms of semantic quality and stylistic control.

Juraj Juraska, Panagiotis Karagiannis, Kevin K. Bowden et al.

Natural language generation lies at the core of generative dialogue systems and conversational agents. We describe an ensemble neural language generator, and present several novel methods for data representation and augmentation that yield improved results in our model. We test the model on three datasets in the restaurant, TV and laptop domains, and report both objective and subjective evaluations of our best model. Using a range of automatic metrics, as well as human evaluators, we show that our approach achieves better results than state-of-the-art models on the same datasets.