Themis: Towards Flexible and Interpretable NLG Evaluation

The evaluation of natural language generation (NLG) tasks is a significant and longstanding research issue. With the recent emergence of powerful large language models (LLMs), some studies have turned to LLM-based automatic evaluation methods, which demonstrate great potential to become a new evaluation paradigm following traditional string-based and model-based metrics. However, despite the improved performance of existing methods, they still possess some deficiencies, such as dependency on references and limited evaluation flexibility. Therefore, in this paper, we meticulously construct a large-scale NLG evaluation corpus NLG-Eval with human and GPT-4 annotations to alleviate the lack of relevant data in this field. Furthermore, we propose Themis, an LLM dedicated to NLG evaluation, which has been trained with our designed multi-perspective consistency and rating-oriented preference alignment methods. Themis can conduct flexible and interpretable evaluations without references, and it exhibits superior evaluation performance on various NLG tasks, simultaneously generalizing well to unseen tasks and surpassing other evaluation models, including GPT-4.

PaCoST: Paired Confidence Significance Testing for Benchmark Contamination Detection in Large Language Models

Large language models (LLMs) are known to be trained on vast amounts of data, which may unintentionally or intentionally include data from commonly used benchmarks. This inclusion can lead to cheatingly high scores on model leaderboards, yet result in disappointing performance in real-world applications. To address this benchmark contamination problem, we first propose a set of requirements that practical contamination detection methods should follow. Following these proposed requirements, we introduce PaCoST, a Paired Confidence Significance Testing to effectively detect benchmark contamination in LLMs. Our method constructs a counterpart for each piece of data with the same distribution, and performs statistical analysis of the corresponding confidence to test whether the model is significantly more confident under the original benchmark. We validate the effectiveness of PaCoST and apply it on popular open-source models and benchmarks. We find that almost all models and benchmarks we tested are suspected contaminated more or less. We finally call for new LLM evaluation methods.

MC-MKE: A Fine-Grained Multimodal Knowledge Editing Benchmark Emphasizing Modality Consistency

Multimodal large language models (MLLMs) are prone to non-factual or outdated knowledge issues, which can manifest as misreading and misrecognition errors due to the complexity of multimodal knowledge. Previous benchmarks have not systematically analyzed the performance of editing methods in correcting these two error types. To better represent and correct these errors, we decompose multimodal knowledge into its visual and textual components. Different error types correspond to different editing formats, which edits distinct part of the multimodal knowledge. We present MC-MKE, a fine-grained Multimodal Knowledge Editing benchmark emphasizing Modality Consistency. Our benchmark facilitates independent correction of misreading and misrecognition errors by editing the corresponding knowledge component. We evaluate three multimodal knowledge editing methods on MC-MKE, revealing their limitations, particularly in terms of modality consistency. Our work highlights the challenges posed by multimodal knowledge editing and motivates further research in developing effective techniques for this task.

ContraSolver: Self-Alignment of Language Models by Resolving Internal Preference Contradictions

While substantial advancements have been made in developing large language models (LLMs), achieving control over their behavior can be difficult. Direct preference optimization (DPO) assumes the existence of a latent reward function to evaluate the responses of LLMs. This assumption indicates a strict preference ordering of different responses to the same input. However, there always exist contradictions of preference in LLMs according to our experimental observations. In this paper, we construct a graph structure of the preference relationship among different responses with self-annotation to find contradictions in the preference order. We propose ContraSolver, an algorithm that traverses all edges on the preference graph to identify those that might cause contradictions. ContraSolver initializes the graph with a maximum spanning tree and identifies contradictory edges, prioritizing the resolution of low-confidence preferences while preserving high-confidence ones. Experimental results on four different generation tasks show that the performance of different LLMs can be largely improved through our completely unsupervised self-alignment. Furthermore, by analyzing the preference graphs of LLMs with and without self-alignment by ContraSolver, we quantify the reduction in contradictions, suggesting that resolving preference contradictions is crucial for achieving better alignment performance.

Better than Random: Reliable NLG Human Evaluation with Constrained Active Sampling

Human evaluation is viewed as a reliable evaluation method for NLG which is expensive and time-consuming. To save labor and costs, researchers usually perform human evaluation on a small subset of data sampled from the whole dataset in practice. However, different selection subsets will lead to different rankings of the systems. To give a more correct inter-system ranking and make the gold standard human evaluation more reliable, we propose a Constrained Active Sampling Framework (CASF) for reliable human judgment. CASF operates through a Learner, a Systematic Sampler and a Constrained Controller to select representative samples for getting a more correct inter-system ranking.Experiment results on 137 real NLG evaluation setups with 44 human evaluation metrics across 16 datasets and 5 NLG tasks demonstrate CASF receives 93.18% top-ranked system recognition accuracy and ranks first or ranks second on 90.91% of the human metrics with 0.83 overall inter-system ranking Kendall correlation.Code and data are publicly available online.

Defining and Detecting Vulnerability in Human Evaluation Guidelines: A Preliminary Study Towards Reliable NLG Evaluation

Human evaluation serves as the gold standard for assessing the quality of Natural Language Generation (NLG) systems. Nevertheless, the evaluation guideline, as a pivotal element ensuring reliable and reproducible human assessment, has received limited attention.Our investigation revealed that only 29.84% of recent papers involving human evaluation at top conferences release their evaluation guidelines, with vulnerabilities identified in 77.09% of these guidelines. Unreliable evaluation guidelines can yield inaccurate assessment outcomes, potentially impeding the advancement of NLG in the right direction. To address these challenges, we take an initial step towards reliable evaluation guidelines and propose the first human evaluation guideline dataset by collecting annotations of guidelines extracted from existing papers as well as generated via Large Language Models (LLMs). We then introduce a taxonomy of eight vulnerabilities and formulate a principle for composing evaluation guidelines. Furthermore, a method for detecting guideline vulnerabilities has been explored using LLMs, and we offer a set of recommendations to enhance reliability in human evaluation. The annotated human evaluation guideline dataset and code for the vulnerability detection method are publicly available online.

WaterPool: A Watermark Mitigating Trade-offs among Imperceptibility, Efficacy and Robustness

With the increasing use of large language models (LLMs) in daily life, concerns have emerged regarding their potential misuse and societal impact. Watermarking is proposed to trace the usage of specific models by injecting patterns into their generated texts. An ideal watermark should produce outputs that are nearly indistinguishable from those of the original LLM (imperceptibility), while ensuring a high detection rate (efficacy), even when the text is partially altered (robustness). Despite many methods having been proposed, none have simultaneously achieved all three properties, revealing an inherent trade-off. This paper utilizes a key-centered scheme to unify existing watermarking techniques by decomposing a watermark into two distinct modules: a key module and a mark module. Through this decomposition, we demonstrate for the first time that the key module significantly contributes to the trade-off issues observed in prior methods. Specifically, this reflects the conflict between the scale of the key sampling space during generation and the complexity of key restoration during detection. To this end, we introduce \textbf{WaterPool}, a simple yet effective key module that preserves a complete key sampling space required by imperceptibility while utilizing semantics-based search to improve the key restoration process. WaterPool can integrate with most watermarks, acting as a plug-in. Our experiments with three well-known watermarking techniques show that WaterPool significantly enhances their performance, achieving near-optimal imperceptibility and markedly improving efficacy and robustness (+12.73\% for KGW, +20.27\% for EXP, +7.27\% for ITS).

Automated Similarity Metric Generation for Recommendation

The embedding-based architecture has become the dominant approach in modern recommender systems, mapping users and items into a compact vector space. It then employs predefined similarity metrics, such as the inner product, to calculate similarity scores between user and item embeddings, thereby guiding the recommendation of items that align closely with a user's preferences. Given the critical role of similarity metrics in recommender systems, existing methods mainly employ handcrafted similarity metrics to capture the complex characteristics of user-item interactions. Yet, handcrafted metrics may not fully capture the diverse range of similarity patterns that can significantly vary across different domains. To address this issue, we propose an Automated Similarity Metric Generation method for recommendations, named AutoSMG, which can generate tailored similarity metrics for various domains and datasets. Specifically, we first construct a similarity metric space by sampling from a set of basic embedding operators, which are then integrated into computational graphs to represent metrics. We employ an evolutionary algorithm to search for the optimal metrics within this metric space iteratively. To improve search efficiency, we utilize an early stopping strategy and a surrogate model to approximate the performance of candidate metrics instead of fully training models. Notably, our proposed method is model-agnostic, which can seamlessly plugin into different recommendation model architectures. The proposed method is validated on three public recommendation datasets across various domains in the Top-K recommendation task, and experimental results demonstrate that AutoSMG outperforms both commonly used handcrafted metrics and those generated by other search strategies.

WikiTableEdit: A Benchmark for Table Editing by Natural Language Instruction

Tabular data, as a crucial form of data representation, exists in diverse formats on the Web. When confronted with complex and irregular tables, manual modification becomes a laborious task. This paper investigates the performance of Large Language Models (LLMs) in the context of table editing tasks. Existing research mainly focuses on regular-shaped tables, wherein instructions are used to generate code in SQL, Python, or Excel Office-script for manipulating the tables. Nevertheless, editing tables with irregular structures, particularly those containing merged cells spanning multiple rows, poses a challenge when using code. To address this, we introduce the WikiTableEdit dataset. Leveraging 26,531 tables from the WikiSQL dataset, we automatically generate natural language instructions for six distinct basic operations and the corresponding outcomes, resulting in over 200,000 instances. Subsequently, we evaluate several representative large language models on the WikiTableEdit dataset to demonstrate the challenge of this task. The dataset will be released to the community to promote related researches.

Evaluating and Mitigating Number Hallucinations in Large Vision-Language Models: A Consistency Perspective

Large vision language models have demonstrated remarkable efficacy in addressing challenges related to both textual and visual content. Nevertheless, these models are susceptible to various hallucinations. In this paper, we focus on a new form of hallucination, specifically termed as number hallucination, which denotes instances where models fail to accurately identify the quantity of objects in an image. We establish a dataset and employ evaluation metrics to assess number hallucination, revealing a pronounced prevalence of this issue across mainstream large vision language models (LVLMs). Additionally, we delve into a thorough analysis of number hallucination, examining inner and outer inconsistency problem from two related perspectives. We assert that this inconsistency is one cause of number hallucination and propose a consistency training method as a means to alleviate such hallucination, which achieves an average improvement of 8\% compared with direct finetuning method.