Adopting human and large language models (LLM) as judges (*a.k.a* human- and LLM-as-a-judge) for evaluating the performance of LLMs has recently gained attention. Nonetheless, this approach concurrently introduces potential biases from human and LLMs, questioning the reliability of the evaluation results. In this paper, we propose a novel framework that is free from referencing groundtruth annotations for investigating **Misinformation Oversight Bias**, **Gender Bias**, **Authority Bias** and **Beauty Bias** on LLM and human judges. We curate a dataset referring to the revised Bloom’s Taxonomy and conduct thousands of evaluations. Results show that human and LLM judges are vulnerable to perturbations to various degrees, and that even the cutting-edge judges possess considerable biases. We further exploit these biases to conduct attacks on LLM judges. We hope that our work can notify the community of the bias and vulnerability of human- and LLM-as-a-judge, as well as the urgency of developing robust evaluation systems.
Mainstream approaches to aligning large language models (LLMs) heavily rely on human preference data, particularly when models require periodic updates. The standard process for iterative alignment of LLMs involves collecting new human feedback for each update. However, the data collection process is costly and challenging to scale. To address this issue, we introduce the “TS-Align” framework, which fine-tunes a policy model using pairwise feedback data automatically mined from its outputs. This automatic mining process is efficiently accomplished through the collaboration between a large-scale teacher model and a small-scale student model. The policy fine-tuning process can be iteratively repeated using on-policy generations within our proposed teacher-student collaborative framework. Through extensive experiments, we demonstrate that our final aligned policy outperforms the base policy model with an average win rate of 69.7% across seven conversational or instruction-following datasets. Furthermore, we show that the ranking capability of the teacher is effectively distilled into the student through our pipeline, resulting in a small-scale yet effective reward model for policy model alignment.
Large Language Models (LLMs) provide a possibility to make a great breakthrough in medicine. The establishment of a standardized medical benchmark becomes a fundamental cornerstone to measure progression. However, medical environments in different regions have their local characteristics, e.g., the ubiquity and significance of traditional Chinese medicine within China. Therefore, merely translating English-based medical evaluation may result in contextual incongruities to a local region. To solve the issue, we propose a localized medical benchmark called CMB, a Comprehensive Medical Benchmark in Chinese, designed and rooted entirely within the native Chinese linguistic and cultural framework. While traditional Chinese medicine is integral to this evaluation, it does not constitute its entirety. Using this benchmark, we have evaluated several prominent large-scale LLMs, including ChatGPT, GPT-4, dedicated Chinese LLMs, and LLMs specialized in the medical domain. We hope this benchmark provide first-hand experience in existing LLMs for medicine and also facilitate the widespread adoption and enhancement of medical LLMs within China. Our data and code are publicly available at https://github.com/FreedomIntelligence/CMB.
The unparalleled performance of closed-sourced ChatGPT has sparked efforts towards its democratization, with notable strides made by leveraging real user and ChatGPT dialogues, as evidenced by Vicuna. However, due to challenges in gathering dialogues involving human participation, current endeavors like Baize and UltraChat rely on ChatGPT conducting roleplay to simulate humans based on instructions, resulting in overdependence on seeds, diminished human-likeness, limited topic diversity, and an absence of genuine multi-round conversational dynamics. To address the above issues, we propose a paradigm to simulate human behavior better and explore the benefits of incorporating more human-like questions in multi-turn conversations. Specifically, we directly target human questions extracted from genuine human-machine conversations as a learning goal and provide a novel user simulator called ‘Socratic‘. The experimental results show our response model, ‘PlatoLM‘, achieves SoTA performance among LLaMA-based 7B models in MT-Bench. Our findings further demonstrate that our method introduces highly human-like questioning patterns and rich topic structures, which can teach the response model better than previous works in multi-round conversations.
Topic segmentation and outline generation strive to divide a document into coherent topic sections and generate corresponding subheadings, unveiling the discourse topic structure of a document. Compared with sentence-level topic structure, the paragraph-level topic structure can quickly grasp and understand the overall context of the document from a higher level, benefitting many downstream tasks such as summarization, discourse parsing, and information retrieval. However, the lack of large-scale, high-quality Chinese paragraph-level topic structure corpora restrained relative research and applications. To fill this gap, we build the Chinese paragraph-level topic representation, corpus, and benchmark in this paper. Firstly, we propose a hierarchical paragraph-level topic structure representation with three layers to guide the corpus construction. Then, we employ a two-stage man-machine collaborative annotation method to construct the largest Chinese Paragraph-level Topic Structure corpus (CPTS), achieving high quality. We also build several strong baselines, including ChatGPT, to validate the computability of CPTS on two fundamental tasks (topic segmentation and outline generation) and preliminarily verified its usefulness for the downstream task (discourse parsing).
Large language models, like ChatGPT, have shown remarkable capability in many downstream tasks, yet their ability to understand discourse structures of dialogues remains less explored, where it requires higher level capabilities of understanding and reasoning. In this paper, we aim to systematically inspect ChatGPT’s performance in two discourse analysis tasks: topic segmentation and discourse parsing, focusing on its deep semantic understanding of linear and hierarchical discourse structures underlying dialogue. To instruct ChatGPT to complete these tasks, we initially craft a prompt template consisting of the task description, output format, and structured input. Then, we conduct experiments on four popular topic segmentation datasets and two discourse parsing datasets. The experimental results showcase that ChatGPT demonstrates proficiency in identifying topic structures in general-domain conversations yet struggles considerably in specific-domain conversations. We also found that ChatGPT hardly understands rhetorical structures that are more complex than topic structures. Our deeper investigation indicates that ChatGPT can give more reasonable topic structures than human annotations but only linearly parses the hierarchical rhetorical structures. In addition, we delve into the impact of in-context learning (e.g., chain-of-thought) on ChatGPT and conduct the ablation study on various prompt components, which can provide a research foundation for future work. The code is available at https://github.com/yxfanSuda/GPTforDDA.
Most neural abstractive summarization models are capable of producing high-quality summaries. However, they still frequently contain factual errors. Existing factuality-oriented abstractive summarization models only consider the integration of factual information and ignore the causes of factual errors. To address this issue, we propose a factuality-oriented abstractive summarization model DASum, which is based on a new task factual relation discrimination that is able to identify the causes of factual errors. First, we use data augmentation methods to construct counterfactual summaries (i. e., negative samples), and build a factual summarization dataset. Then, we propose the factual relation discrimination task, which determines the factuality of the dependency relations in summaries during summary generation and guides our DASum to generate factual relations, thereby improving the factuality of summaries. Experimental results on the CNN/DM and XSUM datasets show that our DASum outperforms several state-of-the-art benchmarks in terms of the factual metrics.
In this paper, we present HuatuoGPT, a Large Language Model (LLM) for medical consultation. The core recipe of HuatuoGPT is to leverage both distilled data from **ChatGPT** and real-world data from **doctors** in the supervised fine-tuning stage. This is not only because purely using **ChatGPT**-distilled data might cause ‘model collapse’, but also because real-world data from **doctors** would be complementary to **ChatGPT**-distilled data. The responses from ChatGPT are usually detailed, well-presented, fluent, and instruction-followed, but it cannot perform like a doctor in many aspects, e.g. for interactive diagnosis. Therefore, the extra doctors’ data could tame a distilled language model to perform like doctors. To synergize the strengths of both data sources, we introduce RLMF (Reinforcement Learning from Mixed Feedback) where a reward model is trained to align the language model with the merits that both sources (ChatGPT and doctors) bring. Experimental results (in GPT-4 evaluation, human evaluation, and medical benchmark datasets) demonstrate that HuatuoGPT achieves state-of-the-art results in performing medical consultation among open-source LLMs. It is worth noting that by using additional real-world data and RLMF, the distilled language model (i.e., HuatuoGPT) outperforms its teacher model (i.e., ChatGPT) in most cases.
Dialogue discourse parsing aims to reflect the relation-based structure of dialogue by establishing discourse links according to discourse relations. To alleviate data sparsity, previous studies have adopted multitasking approaches to jointly learn dialogue discourse parsing with related tasks (e.g., reading comprehension) that require additional human annotation, thus limiting their generality. In this paper, we propose a multitasking framework that integrates dialogue discourse parsing with its neighboring task addressee recognition. Addressee recognition reveals the reply-to structure that partially overlaps with the relation-based structure, which can be exploited to facilitate relation-based structure learning. To this end, we first proposed a reinforcement learning agent to identify training examples from addressee recognition that are most helpful for dialog discourse parsing. Then, a task-aware structure transformer is designed to capture the shared and private dialogue structure of different tasks, thereby further promoting dialogue discourse parsing. Experimental results on both the Molweni and STAC datasets show that our proposed method can outperform the SOTA baselines. The code will be available at https://github.com/yxfanSuda/RLTST.
Document-level Relation Extraction (DocRE) aims at extracting relations between entities in a given document. Since different mention pairs may express different relations or even no relation, it is crucial to identify key mention pairs responsible for the entity-level relation labels. However, most recent studies treat different mentions equally while predicting the relations between entities, leading to sub-optimal performance. To this end, we propose a novel DocRE model called Key Mention pairs Guided Relation Extractor (KMGRE) to directly model mention-level relations, containing two modules: a mention-level relation extractor and a key instance classifier. These two modules could be iteratively optimized with an EM-based algorithm to enhance each other. We also propose a new method to solve the multi-label problem in optimizing the mention-level relation extractor. Experimental results on two public DocRE datasets demonstrate that the proposed model is effective and outperforms previous state-of-the-art models.
Automatic Essay Scoring (AES) is the task of using the computer to evaluate the quality of essays automatically. Current research on AES focuses on scoring the overall quality or single trait of prompt-specific essays. However, the users not only expect to obtain the overall score but also the instant feedback from different traits to help their writing in the real world. Therefore, we first annotate a mutli-trait dataset ACEA including 1220 argumentative essays from four traits, i.e., essay organization, topic, logic, and language. And then we design a hierarchical multi-task trait scorer HMTS to evaluate the quality of writing by modeling these four traits. Moreover, we propose an inter-sequence attention mechanism to enhance information interaction between different tasks and design the trait-specific features for various tasks in AES. The experimental results on ACEA show that our HMTS can effectively score essays from multiple traits, outperforming several strong models.
We present Retriever-Transducer-Checker (ReTraCk), a neural semantic parsing framework for large scale knowledge base question answering (KBQA). ReTraCk is designed as a modular framework to maintain high flexibility. It includes a retriever to retrieve relevant KB items efficiently, a transducer to generate logical form with syntax correctness guarantees and a checker to improve transduction procedure. ReTraCk is ranked at top1 overall performance on the GrailQA leaderboard and obtains highly competitive performance on the typical WebQuestionsSP benchmark. Our system can interact with users timely, demonstrating the efficiency of the proposed framework.
Implicit discourse relation recognition (IDRR) is a critical task in discourse analysis. Previous studies only regard it as a classification task and lack an in-depth understanding of the semantics of different relations. Therefore, we first view IDRR as a generation task and further propose a method joint modeling of the classification and generation. Specifically, we propose a joint model, CG-T5, to recognize the relation label and generate the target sentence containing the meaning of relations simultaneously. Furthermore, we design three target sentence forms, including the question form, for the generation model to incorporate prior knowledge. To address the issue that large discourse units are hardly embedded into the target sentence, we also propose a target sentence construction mechanism that automatically extracts core sentences from those large discourse units. Experimental results both on Chinese MCDTB and English PDTB datasets show that our model CG-T5 achieves the best performance against several state-of-the-art systems.
Discourse structure tree construction is the fundamental task of discourse parsing and most previous work focused on English. Due to the cultural and linguistic differences, existing successful methods on English discourse parsing cannot be transformed into Chinese directly, especially in paragraph level suffering from longer discourse units and fewer explicit connectives. To alleviate the above issues, we propose two reading modes, i.e., the global backward reading and the local reverse reading, to construct Chinese paragraph level discourse trees. The former processes discourse units from the end to the beginning in a document to utilize the left-branching bias of discourse structure in Chinese, while the latter reverses the position of paragraphs in a discourse unit to enhance the differentiation of coherence between adjacent discourse units. The experimental results on Chinese MCDTB demonstrate that our model outperforms all strong baselines.
Recent neural models for data-to-text generation rely on massive parallel pairs of data and text to learn the writing knowledge. They often assume that writing knowledge can be acquired from the training data alone. However, when people are writing, they not only rely on the data but also consider related knowledge. In this paper, we enhance neural data-to-text models with external knowledge in a simple but effective way to improve the fidelity of generated text. Besides relying on parallel data and text as in previous work, our model attends to relevant external knowledge, encoded as a temporary memory, and combines this knowledge with the context representation of data before generating words. This allows the model to infer relevant facts which are not explicitly stated in the data table from an external knowledge source. Experimental results on twenty-one Wikipedia infobox-to-text datasets show our model, KBAtt, consistently improves a state-of-the-art model on most of the datasets. In addition, to quantify when and why external knowledge is effective, we design a metric, KBGain, which shows a strong correlation with the observed performance boost. This result demonstrates the relevance of external knowledge and sparseness of original data are the main factors affecting system performance.
Discourse parsing is a challenging task and plays a critical role in discourse analysis. This paper focus on the macro level discourse structure analysis, which has been less studied in the previous researches. We explore a macro discourse structure presentation schema to present the macro level discourse structure, and propose a corresponding corpus, named Macro Chinese Discourse Treebank. On these bases, we concentrate on two tasks of macro discourse structure analysis, including structure identification and nuclearity recognition. In order to reduce the error transmission between the associated tasks, we adopt a joint model of the two tasks, and an Integer Linear Programming approach is proposed to achieve global optimization with various kinds of constraints.
In view of the differences between the annotations of micro and macro discourse rela-tionships, this paper describes the relevant experiments on the construction of the Macro Chinese Discourse Treebank (MCDTB), a higher-level Chinese discourse corpus. Fol-lowing RST (Rhetorical Structure Theory), we annotate the macro discourse information, including discourse structure, nuclearity and relationship, and the additional discourse information, including topic sentences, lead and abstract, to make the macro discourse annotation more objective and accurate. Finally, we annotated 720 articles with a Kappa value greater than 0.6. Preliminary experiments on this corpus verify the computability of MCDTB.