Steffen Eger

Christoph Leiter, Ran Zhang, Yanran Chen et al.

ChatGPT, a chatbot developed by OpenAI, has gained widespread popularity and media attention since its release in November 2022. However, little hard evidence is available regarding its perception in various sources. In this paper, we analyze over 300,000 tweets and more than 150 scientific papers to investigate how ChatGPT is perceived and discussed. Our findings show that ChatGPT is generally viewed as of high quality, with positive sentiment and emotions of joy dominating in social media. Its perception has slightly decreased since its debut, however, with joy decreasing and (negative) surprise on the rise, and it is perceived more negatively in languages other than English. In recent scientific papers, ChatGPT is characterized as a great opportunity across various fields including the medical domain, but also as a threat concerning ethics and receives mixed assessments for education. Our comprehensive meta-analysis of ChatGPT's current perception after 2.5 months since its release can contribute to shaping the public debate and informing its future development. We make our data available.

Christoph Leiter, Yuki M. Asano, Margret Keuper et al.

The assessment of evaluation metrics (meta-evaluation) is crucial for determining the suitability of existing metrics in text-to-image (T2I) generation tasks. Human-based meta-evaluation is costly and time-intensive, and automated alternatives are scarce. We address this gap and propose CROC: a scalable framework for automated Contrastive Robustness Checks that systematically probes and quantifies metric robustness by synthesizing contrastive test cases across a comprehensive taxonomy of image properties. With CROC, we generate a pseudo-labeled dataset (CROC$^{syn}$) of over 1 million contrastive prompt-image pairs to enable a fine-grained comparison of evaluation metrics. We also use this dataset to train CROCScore, a new metric that achieves state-of-the-art performance among open-source methods, demonstrating an additional key application of our framework. To complement this dataset, we introduce a human-supervised benchmark (CROC$^{hum}$) targeting especially challenging categories. Our results highlight robustness issues in existing metrics: for example, many fail on prompts involving negation, and all tested open-source metrics fail on at least 24% of cases involving correct identification of body parts.

Jonas Belouadi, Anne Lauscher, Steffen Eger

Generating bitmap graphics from text has gained considerable attention, yet for scientific figures, vector graphics are often preferred. Given that vector graphics are typically encoded using low-level graphics primitives, generating them directly is difficult. To address this, we propose the use of TikZ, a well-known abstract graphics language that can be compiled to vector graphics, as an intermediate representation of scientific figures. TikZ offers human-oriented, high-level commands, thereby facilitating conditional language modeling with any large language model. To this end, we introduce DaTikZ, the first large-scale TikZ dataset consisting of 120k TikZ drawings aligned with captions. We fine-tune LLaMA on DaTikZ, as well as our new model CLiMA, which augments LLaMA with multimodal CLIP embeddings. In both human and automatic evaluation, CLiMA and LLaMA outperform commercial GPT-4 and Claude 2 in terms of similarity to human-created figures, with CLiMA additionally improving text-image alignment. Our detailed analysis shows that all models generalize well and are not susceptible to memorization. GPT-4 and Claude 2, however, tend to generate more simplistic figures compared to both humans and our models. We make our framework, AutomaTikZ, along with model weights and datasets, publicly available.

Jonas Belouadi, Steffen Eger

State-of-the-art poetry generation systems are often complex. They either consist of task-specific model pipelines, incorporate prior knowledge in the form of manually created constraints, or both. In contrast, end-to-end models would not suffer from the overhead of having to model prior knowledge and could learn the nuances of poetry from data alone, reducing the degree of human supervision required. In this work, we investigate end-to-end poetry generation conditioned on styles such as rhyme, meter, and alliteration. We identify and address lack of training data and mismatching tokenization algorithms as possible limitations of past attempts. In particular, we successfully pre-train ByGPT5, a new token-free decoder-only language model, and fine-tune it on a large custom corpus of English and German quatrains annotated with our styles. We show that ByGPT5 outperforms other models such as mT5, ByT5, GPT-2 and ChatGPT, while also being more parameter efficient and performing favorably compared to humans. In addition, we analyze its runtime performance and demonstrate that it is not prone to memorization. We make our code, models, and datasets publicly available.

Yanran Chen, Steffen Eger

Recently proposed BERT-based evaluation metrics for text generation perform well on standard benchmarks but are vulnerable to adversarial attacks, e.g., relating to information correctness. We argue that this stems (in part) from the fact that they are models of semantic similarity. In contrast, we develop evaluation metrics based on Natural Language Inference (NLI), which we deem a more appropriate modeling. We design a preference-based adversarial attack framework and show that our NLI based metrics are much more robust to the attacks than the recent BERT-based metrics. On standard benchmarks, our NLI based metrics outperform existing summarization metrics, but perform below SOTA MT metrics. However, when combining existing metrics with our NLI metrics, we obtain both higher adversarial robustness (15%-30%) and higher quality metrics as measured on standard benchmarks (+5% to 30%).

Christoph Leiter, Piyawat Lertvittayakumjorn, Marina Fomicheva et al.

Unlike classical lexical overlap metrics such as BLEU, most current evaluation metrics for machine translation (for example, COMET or BERTScore) are based on black-box large language models. They often achieve strong correlations with human judgments, but recent research indicates that the lower-quality classical metrics remain dominant, one of the potential reasons being that their decision processes are more transparent. To foster more widespread acceptance of novel high-quality metrics, explainability thus becomes crucial. In this concept paper, we identify key properties as well as key goals of explainable machine translation metrics and provide a comprehensive synthesis of recent techniques, relating them to our established goals and properties. In this context, we also discuss the latest state-of-the-art approaches to explainable metrics based on generative models such as ChatGPT and GPT4. Finally, we contribute a vision of next-generation approaches, including natural language explanations. We hope that our work can help catalyze and guide future research on explainable evaluation metrics and, mediately, also contribute to better and more transparent machine translation systems.

Yanran Chen, Steffen Eger

We consider the end-to-end abstract-to-title generation problem, exploring seven recent transformer based models (including ChatGPT) fine-tuned on more than 30k abstract-title pairs from NLP and machine learning (ML) venues. As an extension, we also consider the harder problem of generating humorous paper titles. For the latter, we compile the first large-scale humor annotated dataset for scientific papers in the NLP/ML domains, comprising almost ~2.6k titles. We evaluate all models using human and automatic metrics. Our human evaluation suggests that our best end-to-end system performs similarly to human authors (but arguably slightly worse). Generating funny titles is more difficult, however, and our automatic systems clearly underperform relative to humans and often learn dataset artefacts of humor. Finally, ChatGPT, without any fine-tuning, performs on the level of our best fine-tuned system.

Christoph Leiter, Piyawat Lertvittayakumjorn, Marina Fomicheva et al.

Unlike classical lexical overlap metrics such as BLEU, most current evaluation metrics (such as BERTScore or MoverScore) are based on black-box language models such as BERT or XLM-R. They often achieve strong correlations with human judgments, but recent research indicates that the lower-quality classical metrics remain dominant, one of the potential reasons being that their decision processes are transparent. To foster more widespread acceptance of the novel high-quality metrics, explainability thus becomes crucial. In this concept paper, we identify key properties and propose key goals of explainable machine translation evaluation metrics. We also provide a synthesizing overview over recent approaches for explainable machine translation metrics and discuss how they relate to those goals and properties. Further, we conduct own novel experiments, which (among others) find that current adversarial NLP techniques are unsuitable for automatically identifying limitations of high-quality black-box evaluation metrics, as they are not meaning-preserving. Finally, we provide a vision of future approaches to explainable evaluation metrics and their evaluation. We hope that our work can help catalyze and guide future research on explainable evaluation metrics and, mediately, also contribute to better and more transparent text generation systems.

Ran Zhang, Jihed Ouni, Steffen Eger

While summarization has been extensively researched in natural language processing (NLP), cross-lingual cross-temporal summarization (CLCTS) is a largely unexplored area that has the potential to improve cross-cultural accessibility and understanding. This paper comprehensively addresses the CLCTS task, including dataset creation, modeling, and evaluation. We (1) build the first CLCTS corpus with 328 instances for hDe-En (extended version with 455 instances) and 289 for hEn-De (extended version with 501 instances), leveraging historical fiction texts and Wikipedia summaries in English and German; (2) examine the effectiveness of popular transformer end-to-end models with different intermediate finetuning tasks; (3) explore the potential of GPT-3.5 as a summarizer; (4) report evaluations from humans, GPT-4, and several recent automatic evaluation metrics. Our results indicate that intermediate task finetuned end-to-end models generate bad to moderate quality summaries while GPT-3.5, as a zero-shot summarizer, provides moderate to good quality outputs. GPT-3.5 also seems very adept at normalizing historical text. To assess data contamination in GPT-3.5, we design an adversarial attack scheme in which we find that GPT-3.5 performs slightly worse for unseen source documents compared to seen documents. Moreover, it sometimes hallucinates when the source sentences are inverted against its prior knowledge with a summarization accuracy of 0.67 for plot omission, 0.71 for entity swap, and 0.53 for plot negation. Overall, our regression results of model performances suggest that longer, older, and more complex source texts (all of which are more characteristic for historical language variants) are harder to summarize for all models, indicating the difficulty of the CLCTS task.

Doan Nam Long Vu, Nafise Sadat Moosavi, Steffen Eger

The evaluation of recent embedding-based evaluation metrics for text generation is primarily based on measuring their correlation with human evaluations on standard benchmarks. However, these benchmarks are mostly from similar domains to those used for pretraining word embeddings. This raises concerns about the (lack of) generalization of embedding-based metrics to new and noisy domains that contain a different vocabulary than the pretraining data. In this paper, we examine the robustness of BERTScore, one of the most popular embedding-based metrics for text generation. We show that (a) an embedding-based metric that has the highest correlation with human evaluations on a standard benchmark can have the lowest correlation if the amount of input noise or unknown tokens increases, (b) taking embeddings from the first layer of pretrained models improves the robustness of all metrics, and (c) the highest robustness is achieved when using character-level embeddings, instead of token-based embeddings, from the first layer of the pretrained model.

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.

Gil Rocha, Henrique Lopes Cardoso, Jonas Belouadi et al.

Available corpora for Argument Mining differ along several axes, and one of the key differences is the presence (or absence) of discourse markers to signal argumentative content. Exploring effective ways to use discourse markers has received wide attention in various discourse parsing tasks, from which it is well-known that discourse markers are strong indicators of discourse relations. To improve the robustness of Argument Mining systems across different genres, we propose to automatically augment a given text with discourse markers such that all relations are explicitly signaled. Our analysis unveils that popular language models taken out-of-the-box fail on this task; however, when fine-tuned on a new heterogeneous dataset that we construct (including synthetic and real examples), they perform considerably better. We demonstrate the impact of our approach on an Argument Mining downstream task, evaluated on different corpora, showing that language models can be trained to automatically fill in discourse markers across different corpora, improving the performance of a downstream model in some, but not all, cases. Our proposed approach can further be employed as an assistive tool for better discourse understanding.

Daniil Larionov, Jens Grünwald, Christoph Leiter et al.

Efficiency is a key property to foster inclusiveness and reduce environmental costs, especially in an era of LLMs. In this work, we provide a comprehensive evaluation of efficiency for MT evaluation metrics. Our approach involves replacing computation-intensive transformers with lighter alternatives and employing linear and quadratic approximations for alignment algorithms on top of LLM representations. We evaluate six (reference-free and reference-based) metrics across three MT datasets and examine 16 lightweight transformers. In addition, we look into the training efficiency of metrics like COMET by utilizing adapters. Our results indicate that (a) TinyBERT provides the optimal balance between quality and efficiency, (b) CPU speed-ups are more substantial than those on GPU; (c) WMD approximations yield no efficiency gains while reducing quality and (d) adapters enhance training efficiency (regarding backward pass speed and memory requirements) as well as, in some cases, metric quality. These findings can help to strike a balance between evaluation speed and quality, which is essential for effective NLG systems. Furthermore, our research contributes to the ongoing efforts to optimize NLG evaluation metrics with minimal impact on performance. To our knowledge, ours is the most comprehensive analysis of different aspects of efficiency for MT metrics conducted so far.

Steffen Eger, Christoph Leiter, Jonas Belouadi et al.

The rapid growth of information in the field of Generative Artificial Intelligence (AI), particularly in the subfields of Natural Language Processing (NLP) and Machine Learning (ML), presents a significant challenge for researchers and practitioners to keep pace with the latest developments. To address the problem of information overload, this report by the Natural Language Learning Group at Bielefeld University focuses on identifying the most popular papers on arXiv, with a specific emphasis on NLP and ML. The objective is to offer a quick guide to the most relevant and widely discussed research, aiding both newcomers and established researchers in staying abreast of current trends. In particular, we compile a list of the 40 most popular papers based on normalized citation counts from the first half of 2023. We observe the dominance of papers related to Large Language Models (LLMs) and specifically ChatGPT during the first half of 2023, with the latter showing signs of declining popularity more recently, however. Further, NLP related papers are the most influential (around 60\% of top papers) even though there are twice as many ML related papers in our data. Core issues investigated in the most heavily cited papers are: LLM efficiency, evaluation techniques, ethical considerations, embodied agents, and problem-solving with LLMs. Additionally, we examine the characteristics of top papers in comparison to others outside the top-40 list (noticing the top paper's focus on LLM related issues and higher number of co-authors) and analyze the citation distributions in our dataset, among others.

Christoph Leiter, Hoa Nguyen, Steffen Eger

State-of-the-art natural language generation evaluation metrics are based on black-box language models. Hence, recent works consider their explainability with the goals of better understandability for humans and better metric analysis, including failure cases. In contrast, our proposed method BMX: Boosting Natural Language Generation Metrics with explainability explicitly leverages explanations to boost the metrics' performance. In particular, we perceive feature importance explanations as word-level scores, which we convert, via power means, into a segment-level score. We then combine this segment-level score with the original metric to obtain a better metric. Our tests show improvements for multiple metrics across MT and summarization datasets. While improvements in machine translation are small, they are strong for summarization. Notably, BMX with the LIME explainer and preselected parameters achieves an average improvement of 0.087 points in Spearman correlation on the system-level evaluation of SummEval.

Ran Zhang, Steffen Eger

Despite substantial progress of large language models (LLMs) for automatic poetry generation, the generated poetry lacks diversity while the training process differs greatly from human learning. Under the rationale that the learning process of the poetry generation systems should be more human-like and their output more diverse and novel, we introduce a framework based on social learning where we emphasize non-cooperative interactions besides cooperative interactions to encourage diversity. Our experiments are the first attempt at LLM-based multi-agent systems in non-cooperative environments for poetry generation employing both TRAINING-BASED agents (GPT-2) and PROMPTING-BASED agents (GPT-3 and GPT-4). Our evaluation based on 96k generated poems shows that our framework benefits the poetry generation process for TRAINING-BASED agents resulting in 1) a 3.0-3.7 percentage point (pp) increase in diversity and a 5.6-11.3 pp increase in novelty according to distinct and novel n-grams. The generated poetry from TRAINING-BASED agents also exhibits group divergence in terms of lexicons, styles and semantics. PROMPTING-BASED agents in our framework also benefit from non-cooperative environments and a more diverse ensemble of models with non-homogeneous agents has the potential to further enhance diversity, with an increase of 7.0-17.5 pp according to our experiments. However, PROMPTING-BASED agents show a decrease in lexical diversity over time and do not exhibit the group-based divergence intended in the social network. Our paper argues for a paradigm shift in creative tasks such as automatic poetry generation to include social learning processes (via LLM-based agent modeling) similar to human interaction.

Aida Kostikova, Benjamin Paassen, Dominik Beese et al.

Solidarity is a crucial concept to understand social relations in societies. In this paper, we explore fine-grained solidarity frames to study solidarity towards women and migrants in German parliamentary debates between 1867 and 2022. Using 2,864 manually annotated text snippets (with a cost exceeding 18k Euro), we evaluate large language models (LLMs) like Llama 3, GPT-3.5, and GPT-4. We find that GPT-4 outperforms other LLMs, approaching human annotation quality. Using GPT-4, we automatically annotate more than 18k further instances (with a cost of around 500 Euro) across 155 years and find that solidarity with migrants outweighs anti-solidarity but that frequencies and solidarity types shift over time. Most importantly, group-based notions of (anti-)solidarity fade in favor of compassionate solidarity, focusing on the vulnerability of migrant groups, and exchange-based anti-solidarity, focusing on the lack of (economic) contribution. Our study highlights the interplay of historical events, socio-economic needs, and political ideologies in shaping migration discourse and social cohesion. We also show that powerful LLMs, if carefully prompted, can be cost-effective alternatives to human annotation for hard social scientific tasks.

Daniil Larionov, Sotaro Takeshita, Ran Zhang et al.

Reasoning-enabled large language models (LLMs) excel in logical tasks, yet their utility for evaluating natural language generation remains unexplored. This study systematically compares reasoning LLMs with non-reasoning counterparts across machine translation and text summarization evaluation tasks. We evaluate eight models spanning state-of-the-art reasoning models (DeepSeek-R1, OpenAI o3), their distilled variants (8B-70B parameters), and equivalent non-reasoning LLMs. Experiments on WMT23 and SummEval benchmarks reveal architecture and task-dependent benefits: OpenAI o3-mini models show improved performance with increased reasoning on MT, while DeepSeek-R1 and generally underperforms compared to its non-reasoning variant except in summarization consistency evaluation. Correlation analysis demonstrates that reasoning token usage correlates with evaluation quality only in specific models, while almost all models generally allocate more reasoning tokens when identifying more quality issues. Distillation maintains reasonable performance up to 32B parameter models but degrades substantially at 8B scale. This work provides the first assessment of reasoning LLMs for NLG evaluation and comparison to non-reasoning models. We share our code to facilitate further research: https://github.com/NL2G/reasoning-eval.

Daniil Larionov, Steffen Eger

Recent advancements in Large Language Model (LLM)-based Natural Language Generation evaluation have largely focused on single-example prompting, resulting in significant token overhead and computational inefficiencies. In this work, we introduce BatchGEMBA-MQM, a framework that integrates batched prompting with the GEMBA-MQM metric for machine translation evaluation. Our approach aggregates multiple translation examples into a single prompt, reducing token usage by 2-4 times (depending on the batch size) relative to single-example prompting. Furthermore, we propose a batching-aware prompt compression model that achieves an additional token reduction of 13-15% on average while also showing ability to help mitigate batching-induced quality degradation. Evaluations across several LLMs (GPT-4o, GPT-4o-mini, Mistral Small, Phi4, and CommandR7B) and varying batch sizes reveal that while batching generally negatively affects quality (but sometimes not substantially), prompt compression does not degrade further, and in some cases, recovers quality loss. For instance, GPT-4o retains over 90% of its baseline performance at a batch size of 4 when compression is applied, compared to a 44.6% drop without compression. We plan to release our code and trained models at https://github.com/NL2G/batchgemba to support future research in this domain.

Christoph Leiter, Steffen Eger

Large language models (LLMs) have revolutionized NLP research. Notably, in-context learning enables their use as evaluation metrics for natural language generation, making them particularly advantageous in low-resource scenarios and time-restricted applications. In this work, we introduce PrExMe, a large-scale Prompt Exploration for Metrics, where we evaluate more than 720 prompt templates for open-source LLM-based metrics on machine translation (MT) and summarization datasets, totalling over 6.6M evaluations. This extensive comparison (1) benchmarks recent open-source LLMs as metrics and (2) explores the stability and variability of different prompting strategies. We discover that, on the one hand, there are scenarios for which prompts are stable. For instance, some LLMs show idiosyncratic preferences and favor to grade generated texts with textual labels while others prefer to return numeric scores. On the other hand, the stability of prompts and model rankings can be susceptible to seemingly innocuous changes. For example, changing the requested output format from "0 to 100" to "-1 to +1" can strongly affect the rankings in our evaluation. Our study contributes to understanding the impact of different prompting approaches on LLM-based metrics for MT and summarization evaluation, highlighting the most stable prompting patterns and potential limitations.

Daniil Larionov, Mikhail Seleznyov, Vasiliy Viskov et al.

State-of-the-art trainable machine translation evaluation metrics like xCOMET achieve high correlation with human judgment but rely on large encoders (up to 10.7B parameters), making them computationally expensive and inaccessible to researchers with limited resources. To address this issue, we investigate whether the knowledge stored in these large encoders can be compressed while maintaining quality. We employ distillation, quantization, and pruning techniques to create efficient xCOMET alternatives and introduce a novel data collection pipeline for efficient black-box distillation. Our experiments show that, using quantization, xCOMET can be compressed up to three times with no quality degradation. Additionally, through distillation, we create an 278M-sized xCOMET-lite metric, which has only 2.6% of xCOMET-XXL parameters, but retains 92.1% of its quality. Besides, it surpasses strong small-scale metrics like COMET-22 and BLEURT-20 on the WMT22 metrics challenge dataset by 6.4%, despite using 50% fewer parameters. All code, dataset, and models are available online at https://github.com/NL2G/xCOMET-lite.

Wei Zhao, Michael Strube, Steffen Eger

Recently, there has been a growing interest in designing text generation systems from a discourse coherence perspective, e.g., modeling the interdependence between sentences. Still, recent BERT-based evaluation metrics are weak in recognizing coherence, and thus are not reliable in a way to spot the discourse-level improvements of those text generation systems. In this work, we introduce DiscoScore, a parametrized discourse metric, which uses BERT to model discourse coherence from different perspectives, driven by Centering theory. Our experiments encompass 16 non-discourse and discourse metrics, including DiscoScore and popular coherence models, evaluated on summarization and document-level machine translation (MT). We find that (i) the majority of BERT-based metrics correlate much worse with human rated coherence than early discourse metrics, invented a decade ago; (ii) the recent state-of-the-art BARTScore is weak when operated at system level -- which is particularly problematic as systems are typically compared in this manner. DiscoScore, in contrast, achieves strong system-level correlation with human ratings, not only in coherence but also in factual consistency and other aspects, and surpasses BARTScore by over 10 correlation points on average. Further, aiming to understand DiscoScore, we provide justifications to the importance of discourse coherence for evaluation metrics, and explain the superiority of one variant over another. Our code is available at \url{https://github.com/AIPHES/DiscoScore}.

Marina Fomicheva, Piyawat Lertvittayakumjorn, Wei Zhao et al.

In this paper, we introduce the Eval4NLP-2021shared task on explainable quality estimation. Given a source-translation pair, this shared task requires not only to provide a sentence-level score indicating the overall quality of the translation, but also to explain this score by identifying the words that negatively impact translation quality. We present the data, annotation guidelines and evaluation setup of the shared task, describe the six participating systems, and analyze the results. To the best of our knowledge, this is the first shared task on explainable NLP evaluation metrics. Datasets and results are available at https://github.com/eval4nlp/SharedTask2021.

Yang Gao, Wei Zhao, Steffen Eger

We study unsupervised multi-document summarization evaluation metrics, which require neither human-written reference summaries nor human annotations (e.g. preferences, ratings, etc.). We propose SUPERT, which rates the quality of a summary by measuring its semantic similarity with a pseudo reference summary, i.e. selected salient sentences from the source documents, using contextualized embeddings and soft token alignment techniques. Compared to the state-of-the-art unsupervised evaluation metrics, SUPERT correlates better with human ratings by 18-39%. Furthermore, we use SUPERT as rewards to guide a neural-based reinforcement learning summarizer, yielding favorable performance compared to the state-of-the-art unsupervised summarizers. All source code is available at https://github.com/yg211/acl20-ref-free-eval.

Thomas Haider, Steffen Eger, Evgeny Kim et al.

Most approaches to emotion analysis of social media, literature, news, and other domains focus exclusively on basic emotion categories as defined by Ekman or Plutchik. However, art (such as literature) enables engagement in a broader range of more complex and subtle emotions. These have been shown to also include mixed emotional responses. We consider emotions in poetry as they are elicited in the reader, rather than what is expressed in the text or intended by the author. Thus, we conceptualize a set of aesthetic emotions that are predictive of aesthetic appreciation in the reader, and allow the annotation of multiple labels per line to capture mixed emotions within their context. We evaluate this novel setting in an annotation experiment both with carefully trained experts and via crowdsourcing. Our annotation with experts leads to an acceptable agreement of kappa = .70, resulting in a consistent dataset for future large scale analysis. Finally, we conduct first emotion classification experiments based on BERT, showing that identifying aesthetic emotions is challenging in our data, with up to .52 F1-micro on the German subset. Data and resources are available at https://github.com/tnhaider/poetry-emotion

Steffen Eger, Johannes Daxenberger, Christian Stab et al.

Argumentation mining (AM) requires the identification of complex discourse structures and has lately been applied with success monolingually. In this work, we show that the existing resources are, however, not adequate for assessing cross-lingual AM, due to their heterogeneity or lack of complexity. We therefore create suitable parallel corpora by (human and machine) translating a popular AM dataset consisting of persuasive student essays into German, French, Spanish, and Chinese. We then compare (i) annotation projection and (ii) bilingual word embeddings based direct transfer strategies for cross-lingual AM, finding that the former performs considerably better and almost eliminates the loss from cross-lingual transfer. Moreover, we find that annotation projection works equally well when using either costly human or cheap machine translations. Our code and data are available at \url{http://github.com/UKPLab/coling2018-xling_argument_mining}.

Steffen Eger, Erik-Lân Do Dinh, Ilia Kuznetsov et al.

This paper describes our approach to the SemEval 2017 Task 10: "Extracting Keyphrases and Relations from Scientific Publications", specifically to Subtask (B): "Classification of identified keyphrases". We explored three different deep learning approaches: a character-level convolutional neural network (CNN), a stacked learner with an MLP meta-classifier, and an attention based Bi-LSTM. From these approaches, we created an ensemble of differently hyper-parameterized systems, achieving a micro-F1-score of 0.63 on the test data. Our approach ranks 2nd (score of 1st placed system: 0.64) out of four according to this official score. However, we erroneously trained 2 out of 3 neural nets (the stacker and the CNN) on only roughly 15% of the full data, namely, the original development set. When trained on the full data (training+development), our ensemble has a micro-F1-score of 0.69. Our code is available from https://github.com/UKPLab/semeval2017-scienceie.

Steffen Eger, Yong Cao, Jennifer D'Souza et al.

With the advent of large multimodal language models, science is now at a threshold of an AI-based technological transformation. Recently, a plethora of new AI models and tools has been proposed, promising to empower researchers and academics worldwide to conduct their research more effectively and efficiently. This includes all aspects of the research cycle, especially (1) searching for relevant literature; (2) generating research ideas and conducting experimentation; generating (3) text-based and (4) multimodal content (e.g., scientific figures and diagrams); and (5) AI-based automatic peer review. In this survey, we provide an in-depth overview over these exciting recent developments, which promise to fundamentally alter the scientific research process for good. Our survey covers the five aspects outlined above, indicating relevant datasets, methods and results (including evaluation) as well as limitations and scope for future research. Ethical concerns regarding shortcomings of these tools and potential for misuse (fake science, plagiarism, harms to research integrity) take a particularly prominent place in our discussion. We hope that our survey will not only become a reference guide for newcomers to the field but also a catalyst for new AI-based initiatives in the area of "AI4Science".

Ran Zhang, Wei Zhao, Steffen Eger

Recent research has focused on literary machine translation (MT) as a new challenge in MT. However, the evaluation of literary MT remains an open problem. We contribute to this ongoing discussion by introducing LITEVAL-CORPUS, a paragraph-level parallel corpus containing verified human translations and outputs from 9 MT systems, which totals over 2k translations and 13k evaluated sentences across four language pairs, costing 4.5k C. This corpus enables us to (i) examine the consistency and adequacy of human evaluation schemes with various degrees of complexity, (ii) compare evaluations by students and professionals, assess the effectiveness of (iii) LLM-based metrics and (iv) LLMs themselves. Our findings indicate that the adequacy of human evaluation is controlled by two factors: the complexity of the evaluation scheme (more complex is less adequate) and the expertise of evaluators (higher expertise yields more adequate evaluations). For instance, MQM (Multidimensional Quality Metrics), a complex scheme and the de facto standard for non-literary human MT evaluation, is largely inadequate for literary translation evaluation: with student evaluators, nearly 60% of human translations are misjudged as indistinguishable or inferior to machine translations. In contrast, BWS (BEST-WORST SCALING), a much simpler scheme, identifies human translations at a rate of 80-100%. Automatic metrics fare dramatically worse, with rates of at most 20%. Our overall evaluation indicates that published human translations consistently outperform LLM translations, where even the most recent LLMs tend to produce considerably more literal and less diverse translations compared to humans.

Yanran Chen, Steffen Eger

Emotions have been shown to play a role in argument convincingness, yet this aspect is underexplored in the natural language processing (NLP) community. Unlike prior studies that use static analyses, focus on a single text domain or language, or treat emotion as just one of many factors, we introduce a dynamic framework inspired by manipulation checks commonly used in psychology and social science; leveraging LLM-based manipulation checks, this framework examines the extent to which perceived emotional intensity influences perceived convincingness. Through human evaluation of arguments across different languages, text domains, and topics, we find that in over half of cases, human judgments of convincingness remain unchanged despite variations in perceived emotional intensity; when emotions do have an impact, they more often enhance rather than weaken convincingness. We further analyze whether 11 LLMs behave like humans in the same scenario, finding that while LLMs generally mirror human patterns, they struggle to capture nuanced emotional effects in individual judgments.

Daniil Larionov, Steffen Eger

Evaluating the quality of machine-generated natural language content is a challenging task in Natural Language Processing (NLP). Recently, large language models (LLMs) like GPT-4 have been employed for this purpose, but they are computationally expensive due to the extensive token usage required by complex evaluation prompts. In this paper, we propose a prompt optimization approach that uses a smaller, fine-tuned language model to compress input data for evaluation prompt, thus reducing token usage and computational cost when using larger LLMs for downstream evaluation. Our method involves a two-stage fine-tuning process: supervised fine-tuning followed by preference optimization to refine the model's outputs based on human preferences. We focus on Machine Translation (MT) evaluation and utilize the GEMBA-MQM metric as a starting point. Our results show a $2.37\times$ reduction in token usage without any loss in evaluation quality. This work makes state-of-the-art LLM-based metrics like GEMBA-MQM more cost-effective and efficient, enhancing their accessibility for broader use.

Vladyslav Nechakhin, Jennifer D'Souza, Steffen Eger

Structured science summaries or research contributions using properties or dimensions beyond traditional keywords enhances science findability. Current methods, such as those used by the Open Research Knowledge Graph (ORKG), involve manually curating properties to describe research papers' contributions in a structured manner, but this is labor-intensive and inconsistent between the domain expert human curators. We propose using Large Language Models (LLMs) to automatically suggest these properties. However, it's essential to assess the readiness of LLMs like GPT-3.5, Llama 2, and Mistral for this task before application. Our study performs a comprehensive comparative analysis between ORKG's manually curated properties and those generated by the aforementioned state-of-the-art LLMs. We evaluate LLM performance through four unique perspectives: semantic alignment and deviation with ORKG properties, fine-grained properties mapping accuracy, SciNCL embeddings-based cosine similarity, and expert surveys comparing manual annotations with LLM outputs. These evaluations occur within a multidisciplinary science setting. Overall, LLMs show potential as recommendation systems for structuring science, but further finetuning is recommended to improve their alignment with scientific tasks and mimicry of human expertise.

Leixin Zhang, Steffen Eger, Yinjie Cheng et al.

Multimodal large language models (LLMs) have demonstrated impressive capabilities in generating high-quality images from textual instructions. However, their performance in generating scientific images--a critical application for accelerating scientific progress--remains underexplored. In this work, we address this gap by introducing ScImage, a benchmark designed to evaluate the multimodal capabilities of LLMs in generating scientific images from textual descriptions. ScImage assesses three key dimensions of understanding: spatial, numeric, and attribute comprehension, as well as their combinations, focusing on the relationships between scientific objects (e.g., squares, circles). We evaluate five models, GPT-4o, Llama, AutomaTikZ, Dall-E, and StableDiffusion, using two modes of output generation: code-based outputs (Python, TikZ) and direct raster image generation. Additionally, we examine four different input languages: English, German, Farsi, and Chinese. Our evaluation, conducted with 11 scientists across three criteria (correctness, relevance, and scientific accuracy), reveals that while GPT-4o produces outputs of decent quality for simpler prompts involving individual dimensions such as spatial, numeric, or attribute understanding in isolation, all models face challenges in this task, especially for more complex prompts.

Yanran Chen, Wei Zhao, Anne Breitbarth et al.

Many studies have shown that human languages tend to optimize for lower complexity and increased communication efficiency. Syntactic dependency distance, which measures the linear distance between dependent words, is often considered a key indicator of language processing difficulty and working memory load. The current paper looks at diachronic trends in syntactic language change in both English and German, using corpora of parliamentary debates from the last c. 160 years. We base our observations on five dependency parsers, including the widely used Stanford CoreNLP as well as 4 newer alternatives. Our analysis of syntactic language change goes beyond linear dependency distance and explores 15 metrics relevant to dependency distance minimization (DDM) and/or based on tree graph properties, such as the tree height and degree variance. Even though we have evidence that recent parsers trained on modern treebanks are not heavily affected by data 'noise' such as spelling changes and OCR errors in our historic data, we find that results of syntactic language change are sensitive to the parsers involved, which is a caution against using a single parser for evaluating syntactic language change as done in previous work. We also show that syntactic language change over the time period investigated is largely similar between English and German for the different metrics explored: only 4% of cases we examine yield opposite conclusions regarding upwards and downtrends of syntactic metrics across German and English. We also show that changes in syntactic measures seem to be more frequent at the tails of sentence length distributions. To our best knowledge, ours is the most comprehensive analysis of syntactic language change using modern NLP technology in recent corpora of English and German.

Hoa Nguyen, Steffen Eger

Citations are a key ingredient of scientific research to relate a paper to others published in the community. Recently, it has been noted that there is a citation age bias in the Natural Language Processing (NLP) community, one of the currently fastest growing AI subfields, in that the mean age of the bibliography of NLP papers has become ever younger in the last few years, leading to `citation amnesia' in which older knowledge is increasingly forgotten. In this work, we put such claims into perspective by analyzing the bibliography of $\sim$300k papers across 15 different scientific fields submitted to the popular preprint server Arxiv in the time period from 2013 to 2022. We find that all AI subfields (in particular: cs.AI, cs.CL, cs.CV, cs.LG) have similar trends of citation amnesia, in which the age of the bibliography has roughly halved in the last 10 years (from above 12 in 2013 to below 7 in 2022), on average. Rather than diagnosing this as a citation age bias in the NLP community, we believe this pattern is an artefact of the dynamics of these research fields, in which new knowledge is produced in ever shorter time intervals.

Xiao Wang, Daniil Larionov, Siwei Wu et al.

Evaluating the quality of generated text automatically remains a significant challenge. Conventional reference-based metrics have been shown to exhibit relatively weak correlation with human evaluations. Recent research advocates the use of large language models (LLMs) as source-based metrics for natural language generation (NLG) assessment. While promising, LLM-based metrics, particularly those using smaller models, still fall short in aligning with human judgments. In this work, we introduce ContrastScore, a contrastive evaluation metric designed to enable higher-quality, less biased, and more efficient assessment of generated text. We evaluate ContrastScore on two NLG tasks: machine translation and summarization. Experimental results show that ContrastScore consistently achieves stronger correlation with human judgments than both single-model and ensemble-based baselines. Notably, ContrastScore based on Qwen 3B and 0.5B even outperforms Qwen 7B, despite having only half as many parameters, demonstrating its efficiency. Furthermore, it effectively mitigates common evaluation biases such as length and likelihood preferences, resulting in more robust automatic evaluation.

Shuzhou Yuan, Jingyi Sun, Ran Zhang et al.

Natural language explanations (NLEs) are commonly used to provide plausible free-text explanations of a model's reasoning about its predictions. However, recent work has questioned their faithfulness, as they may not accurately reflect the model's internal reasoning process regarding its predicted answer. In contrast, highlight explanations--input fragments critical for the model's predicted answers--exhibit measurable faithfulness. Building on this foundation, we propose G-Tex, a Graph-Guided Textual Explanation Generation framework designed to enhance the faithfulness of NLEs. Specifically, highlight explanations are first extracted as faithful cues reflecting the model's reasoning logic toward answer prediction. They are subsequently encoded through a graph neural network layer to guide the NLE generation, which aligns the generated explanations with the model's underlying reasoning toward the predicted answer. Experiments on T5 and BART using three reasoning datasets show that G-Tex improves NLE faithfulness by up to 12.18% compared to baseline methods. Additionally, G-Tex generates NLEs with greater semantic and lexical similarity to human-written ones. Human evaluations show that G-Tex can decrease redundant content and enhance the overall quality of NLEs. Our work presents a novel method for explicitly guiding NLE generation to enhance faithfulness, serving as a foundation for addressing broader criteria in NLE and generated text.

Christoph Leiter, Jonas Belouadi, Yanran Chen et al.

The NLLG (Natural Language Learning & Generation) arXiv reports assist in navigating the rapidly evolving landscape of NLP and AI research across cs.CL, cs.CV, cs.AI, and cs.LG categories. This fourth installment captures a transformative period in AI history - from January 1, 2023, following ChatGPT's debut, through September 30, 2024. Our analysis reveals substantial new developments in the field - with 45% of the top 40 most-cited papers being new entries since our last report eight months ago and offers insights into emerging trends and major breakthroughs, such as novel multimodal architectures, including diffusion and state space models. Natural Language Processing (NLP; cs.CL) remains the dominant main category in the list of our top-40 papers but its dominance is on the decline in favor of Computer vision (cs.CV) and general machine learning (cs.LG). This report also presents novel findings on the integration of generative AI in academic writing, documenting its increasing adoption since 2022 while revealing an intriguing pattern: top-cited papers show notably fewer markers of AI-generated content compared to random samples. Furthermore, we track the evolution of AI-associated language, identifying declining trends in previously common indicators such as "delve".

Aida Kostikova, Ole Pütz, Steffen Eger et al.

Migration has been a core topic in German political debate, from millions of expellees post World War II over labor migration to refugee movements in the recent past. Studying political speech regarding such wide-ranging phenomena in depth traditionally required extensive manual annotations, limiting the scope of analysis to small subsets of the data. Large language models (LLMs) have the potential to partially automate even complex annotation tasks. We provide an extensive evaluation of a multiple LLMs in annotating (anti-)solidarity subtypes in German parliamentary debates compared to a large set of thousands of human reference annotations (gathered over a year). We evaluate the influence of model size, prompting differences, fine-tuning, historical versus contemporary data; and we investigate systematic errors. Beyond methodological evaluation, we also interpret the resulting annotations from a social science lense, gaining deeper insight into (anti-)solidarity trends towards migrants in the German post-World War II period and recent past. Our data reveals a high degree of migrant-directed solidarity in the postwar period, as well as a strong trend towards anti-solidarity in the German parliament since 2015, motivating further research. These findings highlight the promise of LLMs for political text analysis and the importance of migration debates in Germany, where demographic decline and labor shortages coexist with rising polarization.

Yanran Chen, Hannes Gröner, Sina Zarrieß et al.

Natural Language Generation (NLG), and more generally generative AI, are among the currently most impactful research fields. Creative NLG, such as automatic poetry generation, is a fascinating niche in this area. While most previous research has focused on forms of the Turing test when evaluating automatic poetry generation -- can humans distinguish between automatic and human generated poetry -- we evaluate the diversity of automatically generated poetry (with a focus on quatrains), by comparing distributions of generated poetry to distributions of human poetry along structural, lexical, semantic and stylistic dimensions, assessing different model types (word vs. character-level, general purpose LLMs vs. poetry-specific models), including the very recent LLaMA3-8B, and types of fine-tuning (conditioned vs. unconditioned). We find that current automatic poetry systems are considerably underdiverse along multiple dimensions -- they often do not rhyme sufficiently, are semantically too uniform and even do not match the length distribution of human poetry. Our experiments reveal, however, that style-conditioning and character-level modeling clearly increases diversity across virtually all dimensions we explore. Our identified limitations may serve as the basis for more genuinely diverse future poetry generation models.

Jonas Belouadi, Simone Paolo Ponzetto, Steffen Eger

Creating high-quality scientific figures can be time-consuming and challenging, even though sketching ideas on paper is relatively easy. Furthermore, recreating existing figures that are not stored in formats preserving semantic information is equally complex. To tackle this problem, we introduce DeTikZify, a novel multimodal language model that automatically synthesizes scientific figures as semantics-preserving TikZ graphics programs based on sketches and existing figures. To achieve this, we create three new datasets: DaTikZv2, the largest TikZ dataset to date, containing over 360k human-created TikZ graphics; SketchFig, a dataset that pairs hand-drawn sketches with their corresponding scientific figures; and MetaFig, a collection of diverse scientific figures and associated metadata. We train DeTikZify on MetaFig and DaTikZv2, along with synthetically generated sketches learned from SketchFig. We also introduce an MCTS-based inference algorithm that enables DeTikZify to iteratively refine its outputs without the need for additional training. Through both automatic and human evaluation, we demonstrate that DeTikZify outperforms commercial Claude 3 and GPT-4V in synthesizing TikZ programs, with the MCTS algorithm effectively boosting its performance. We make our code, models, and datasets publicly available.

Ran Zhang, Aida Kostikova, Christoph Leiter et al.

Artificial Intelligence (AI) has witnessed rapid growth, especially in the subfields Natural Language Processing (NLP), Machine Learning (ML) and Computer Vision (CV). Keeping pace with this rapid progress poses a considerable challenge for researchers and professionals in the field. In this arXiv report, the second of its kind, which covers the period from January to September 2023, we aim to provide insights and analysis that help navigate these dynamic areas of AI. We accomplish this by 1) identifying the top-40 most cited papers from arXiv in the given period, comparing the current top-40 papers to the previous report, which covered the period January to June; 2) analyzing dataset characteristics and keyword popularity; 3) examining the global sectoral distribution of institutions to reveal differences in engagement across geographical areas. Our findings highlight the continued dominance of NLP: while only 16% of all submitted papers have NLP as primary category (more than 25% have CV and ML as primary category), 50% of the most cited papers have NLP as primary category, 90% of which target LLMs. Additionally, we show that i) the US dominates among both top-40 and top-9k papers, followed by China; ii) Europe clearly lags behind and is hardly represented in the top-40 most cited papers; iii) US industry is largely overrepresented in the top-40 most influential papers.

Cleo Matzken, Steffen Eger, Ivan Habernal

Protecting privacy in contemporary NLP models is gaining in importance. So does the need to mitigate social biases of such models. But can we have both at the same time? Existing research suggests that privacy preservation comes at the price of worsening biases in classification tasks. In this paper, we explore the extent to which this tradeoff really holds when we incorporate both privacy preservation and de-biasing techniques into training text generation models. How does improving the model along one dimension affect the other dimension as well as the utility of the model? We conduct an extensive set of experiments that include bias detection, privacy attacks, language modeling, and performance on downstream tasks.

Anya Belz, Craig Thomson, Ehud Reiter et al.

We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13\% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.

Yanran Chen, Jonas Belouadi, Steffen Eger

Reproducibility is of utmost concern in machine learning and natural language processing (NLP). In the field of natural language generation (especially machine translation), the seminal paper of Post (2018) has pointed out problems of reproducibility of the dominant metric, BLEU, at the time of publication. Nowadays, BERT-based evaluation metrics considerably outperform BLEU. In this paper, we ask whether results and claims from four recent BERT-based metrics can be reproduced. We find that reproduction of claims and results often fails because of (i) heavy undocumented preprocessing involved in the metrics, (ii) missing code and (iii) reporting weaker results for the baseline metrics. (iv) In one case, the problem stems from correlating not to human scores but to a wrong column in the csv file, inflating scores by 5 points. Motivated by the impact of preprocessing, we then conduct a second study where we examine its effects more closely (for one of the metrics). We find that preprocessing can have large effects, especially for highly inflectional languages. In this case, the effect of preprocessing may be larger than the effect of the aggregation mechanism (e.g., greedy alignment vs. Word Mover Distance).

Dominik Beese, Begüm Altunbaş, Görkem Güzeler et al.

In this paper, we classify scientific articles in the domain of natural language processing (NLP) and machine learning (ML), as core subfields of artificial intelligence (AI), into whether (i) they extend the current state-of-the-art by the introduction of novel techniques which beat existing models or whether (ii) they mainly criticize the existing state-of-the-art, i.e. that it is deficient with respect to some property (e.g. wrong evaluation, wrong datasets, misleading task specification). We refer to contributions under (i) as having a 'positive stance' and contributions under (ii) as having a 'negative stance' (to related work). We annotate over 1.5 k papers from NLP and ML to train a SciBERT-based model to automatically predict the stance of a paper based on its title and abstract. We then analyse large-scale trends on over 41 k papers from the last approximately 35 years in NLP and ML, finding that papers have become substantially more positive over time, but negative papers also got more negative and we observe considerably more negative papers in recent years. Negative papers are also more influential in terms of citations they receive.

Jonas Belouadi, Steffen Eger

The vast majority of evaluation metrics for machine translation are supervised, i.e., (i) are trained on human scores, (ii) assume the existence of reference translations, or (iii) leverage parallel data. This hinders their applicability to cases where such supervision signals are not available. In this work, we develop fully unsupervised evaluation metrics. To do so, we leverage similarities and synergies between evaluation metric induction, parallel corpus mining, and MT systems. In particular, we use an unsupervised evaluation metric to mine pseudo-parallel data, which we use to remap deficient underlying vector spaces (in an iterative manner) and to induce an unsupervised MT system, which then provides pseudo-references as an additional component in the metric. Finally, we also induce unsupervised multilingual sentence embeddings from pseudo-parallel data. We show that our fully unsupervised metrics are effective, i.e., they beat supervised competitors on 4 out of our 5 evaluation datasets. We make our code publicly available.

Wei Zhao, Steffen Eger

Multilingual representations pre-trained with monolingual data exhibit considerably unequal task performances across languages. Previous studies address this challenge with resource-intensive contextualized alignment, which assumes the availability of large parallel data, thereby leaving under-represented language communities behind. In this work, we attribute the data hungriness of previous alignment techniques to two limitations: (i) the inability to sufficiently leverage data and (ii) these techniques are not trained properly. To address these issues, we introduce supervised and unsupervised density-based approaches named Real-NVP and GAN-Real-NVP, driven by Normalizing Flow, to perform alignment, both dissecting the alignment of multilingual subspaces into density matching and density modeling. We complement these approaches with our validation criteria in order to guide the training process. Our experiments encompass 16 alignments, including our approaches, evaluated across 6 language pairs, synthetic data and 5 NLP tasks. We demonstrate the effectiveness of our approaches in the scenarios of limited and no parallel data. First, our supervised approach trained on 20k parallel data (sentences) mostly surpasses Joint-Align and InfoXLM trained on over 100k parallel sentences. Second, parallel data can be removed without sacrificing performance when integrating our unsupervised approach in our bootstrapping procedure, which is theoretically motivated to enforce equality of multilingual subspaces. Moreover, we demonstrate the advantages of validation criteria over validation data for guiding supervised training.

Maxime Peyrard, Wei Zhao, Steffen Eger et al.

Evaluation in NLP is usually done by comparing the scores of competing systems independently averaged over a common set of test instances. In this work, we question the use of averages for aggregating evaluation scores into a final number used to decide which system is best, since the average, as well as alternatives such as the median, ignores the pairing arising from the fact that systems are evaluated on the same test instances. We illustrate the importance of taking the instance-level pairing of evaluation scores into account and demonstrate, both theoretically and empirically, the advantages of aggregation methods based on pairwise comparisons, such as the Bradley-Terry (BT) model, a mechanism based on the estimated probability that a given system scores better than another on the test set. By re-evaluating 296 real NLP evaluation setups across four tasks and 18 evaluation metrics, we show that the choice of aggregation mechanism matters and yields different conclusions as to which systems are state of the art in about 30% of the setups. To facilitate the adoption of pairwise evaluation, we release a practical tool for performing the full analysis of evaluation scores with the mean, median, BT, and two variants of BT (Elo and TrueSkill), alongside functionality for appropriate statistical testing.

Marvin Kaster, Wei Zhao, Steffen Eger

Evaluation metrics are a key ingredient for progress of text generation systems. In recent years, several BERT-based evaluation metrics have been proposed (including BERTScore, MoverScore, BLEURT, etc.) which correlate much better with human assessment of text generation quality than BLEU or ROUGE, invented two decades ago. However, little is known what these metrics, which are based on black-box language model representations, actually capture (it is typically assumed they model semantic similarity). In this work, we use a simple regression based global explainability technique to disentangle metric scores along linguistic factors, including semantics, syntax, morphology, and lexical overlap. We show that the different metrics capture all aspects to some degree, but that they are all substantially sensitive to lexical overlap, just like BLEU and ROUGE. This exposes limitations of these novelly proposed metrics, which we also highlight in an adversarial test scenario.