Computation and Language (cs.CL)

Abstract: Semantic feature norms, lists of features that concepts do and do not possess, have played a central role in characterizing human conceptual knowledge, but require extensive human labor. Large language models (LLMs) offer a novel avenue for the automatic generation of such feature lists, but are prone to significant error. Here, we present a new method for combining a learned model of human lexical-semantics from limited data with LLM-generated data to efficiently generate high-quality feature norms.

Abstract: Large language models (LLMs) have achieved impressive performance on code generation. However, for complex programming tasks, generating the correct solution in one go becomes challenging, thus some prior works have designed program repair approaches to improve code generation performance. In this work, we propose Self-Debugging, which teaches a large language model to debug its predicted program via few-shot demonstrations. In particular, we demonstrate that Self-Debugging can teach the large language model to perform rubber duck debugging; i.e., without any feedback on the code correctness or error messages, the model is able to identify its mistakes by explaining the generated code in natural language. Self-Debugging achieves the state-of-the-art performance on several code generation benchmarks, including the Spider dataset for text-to-SQL generation, TransCoder for C++-to-Python translation, and MBPP for text-to-Python generation. On the Spider benchmark where there are no unit tests to verify the correctness of predictions, Self-Debugging with code explanation consistently improves the baseline by 2-3%, and improves the prediction accuracy on problems of the hardest label by 9%. On TransCoder and MBPP where unit tests are available, Self-Debugging improves the baseline accuracy by up to 12%. Meanwhile, by leveraging feedback messages and reusing failed predictions, Self-Debugging notably improves sample efficiency, and can match or outperform baseline models that generate more than 10x candidate programs.

Abstract: With the rapid progress of large language models (LLMs), many downstream NLP tasks can be well solved given good prompts. Though model developers and researchers work hard on dialog safety to avoid generating harmful content from LLMs, it is still challenging to steer AI-generated content (AIGC) for the human good. As powerful LLMs are devouring existing text data from various domains (e.g., GPT-3 is trained on 45TB texts), it is natural to doubt whether the private information is included in the training data and what privacy threats can these LLMs and their downstream applications bring. In this paper, we study the privacy threats from OpenAI's model APIs and New Bing enhanced by ChatGPT and show that application-integrated LLMs may cause more severe privacy threats ever than before. To this end, we conduct extensive experiments to support our claims and discuss LLMs' privacy implications.

Abstract: Multi-document summarization is challenging because the summaries should not only describe the most important information from all documents but also provide a coherent interpretation of the documents. This paper proposes a method for multi-document summarization based on cluster similarity. In the extractive method we use hybrid model based on a modified version of the PageRank algorithm and a text correlation considerations mechanism. After generating summaries by selecting the most important sentences from each cluster, we apply BARTpho and ViT5 to construct the abstractive models. Both extractive and abstractive approaches were considered in this study. The proposed method achieves competitive results in VLSP 2022 competition.

Abstract: Large pre-trained language models such as BERT have been widely used as a framework for natural language understanding (NLU) tasks. However, recent findings have revealed that pre-trained language models are insensitive to word order. The performance on NLU tasks remains unchanged even after randomly permuting the word of a sentence, where crucial syntactic information is destroyed. To help preserve the importance of word order, we propose a simple approach called Forced Invalidation (FI): forcing the model to identify permuted sequences as invalid samples. We perform an extensive evaluation of our approach on various English NLU and QA based tasks over BERT-based and attention-based models over word embeddings. Our experiments demonstrate that Forced Invalidation significantly improves the sensitivity of the models to word order.

Abstract: Once powerful conversational models have become available for a wide audience, users started actively engaging in social interactions with this technology. Such unprecedented interaction experiences may pose considerable social and psychological risks to the users unless the technology is properly controlled. This creates an urgent need for scalable and robust evaluation metrics for conversational chatbots. Existing automatic evaluation metrics usually focus on objective quality measures and disregard subjective perceptions of social dimensions. Moreover, most of these approaches operate on pre-produced dialogs from available benchmark corpora, which implies human involvement for preparing the material for evaluation and, thus, impeded scalability of the metrics. To address this limitation, we propose to make use of the emerging large language models (LLMs) from the GPT-family and describe a new framework allowing to conduct dialog system evaluation with prompting. With this framework, we are able to achieve full automation of the evaluation pipeline and reach impressive correlation with the human judgement (up to Pearson r=0.95 on system level). The underlying concept is to collect synthetic chat logs of evaluated bots with a LLM in the other-play setting, where LLM is carefully conditioned to follow a specific scenario. We further explore different prompting approaches to produce evaluation scores with the same LLM. The best-performing prompts, containing few-show demonstrations and instructions, show outstanding performance on the tested dataset and demonstrate the ability to generalize to other dialog corpora.

Abstract: Existing fact-checking models for biomedical claims are typically trained on synthetic or well-worded data and hardly transfer to social media content. This mismatch can be mitigated by adapting the social media input to mimic the focused nature of common training claims. To do so, Wuehrl & Klinger (2022) propose to extract concise claims based on medical entities in the text. However, their study has two limitations: First, it relies on gold-annotated entities. Therefore, its feasibility for a real-world application cannot be assessed since this requires detecting relevant entities automatically. Second, they represent claim entities with the original tokens. This constitutes a terminology mismatch which potentially limits the fact-checking performance. To understand both challenges, we propose a claim extraction pipeline for medical tweets that incorporates named entity recognition and terminology normalization via entity linking. We show that automatic NER does lead to a performance drop in comparison to using gold annotations but the fact-checking performance still improves considerably over inputting the unchanged tweets. Normalizing entities to their canonical forms does, however, not improve the performance.

Abstract: Reinforcement Learning from Human Feedback (RLHF) facilitates the alignment of large language models with human preferences, significantly enhancing the quality of interactions between humans and these models. InstructGPT implements RLHF through several stages, including Supervised Fine-Tuning (SFT), reward model training, and Proximal Policy Optimization (PPO). PPO, however, is sensitive to hyperparameters and requires a minimum of four models in its standard implementation, which makes it hard to train. In contrast, we propose a novel learning paradigm called RRHF, which scores responses generated by different sampling policies and learns to align them with human preferences through ranking loss. RRHF can efficiently align language model output probabilities with human preferences as robust as fine-tuning and it only needs 1 to 2 models during tuning. In addition, RRHF can be considered an extension of SFT and reward models while being simpler than PPO in terms of coding, model counts, and hyperparameters. The entire alignment process can be accomplished within a single RRHF training session. We evaluate RRHF using LLaMA and Alpaca on Helpful and Harmless data, demonstrating performance comparable to PPO.

Abstract: Large language models (LLMs) have shown incredible capabilities and transcended the natural language processing (NLP) community, with adoption throughout many services like healthcare, therapy, education, and customer service. Since users include people with critical information needs like students or patients engaging with chatbots, the safety of these systems is of prime importance. Therefore, a clear understanding of the capabilities and limitations of LLMs is necessary. To this end, we systematically evaluate toxicity in over half a million generations of ChatGPT, a popular dialogue-based LLM. We find that setting the system parameter of ChatGPT by assigning it a persona, say that of the boxer Muhammad Ali, significantly increases the toxicity of generations. Depending on the persona assigned to ChatGPT, its toxicity can increase up to 6x, with outputs engaging in incorrect stereotypes, harmful dialogue, and hurtful opinions. This may be potentially defamatory to the persona and harmful to an unsuspecting user. Furthermore, we find concerning patterns where specific entities (e.g., certain races) are targeted more than others (3x more) irrespective of the assigned persona, that reflect inherent discriminatory biases in the model. We hope that our findings inspire the broader AI community to rethink the efficacy of current safety guardrails and develop better techniques that lead to robust, safe, and trustworthy AI systems.

Abstract: This paper describes Adam Mickiewicz University's (AMU) solution for the 4th Shared Task on SlavNER. The task involves the identification, categorization, and lemmatization of named entities in Slavic languages. Our approach involved exploring the use of foundation models for these tasks. In particular, we used models based on the popular BERT and T5 model architectures. Additionally, we used external datasets to further improve the quality of our models. Our solution obtained promising results, achieving high metrics scores in both tasks. We describe our approach and the results of our experiments in detail, showing that the method is effective for NER and lemmatization in Slavic languages. Additionally, our models for lemmatization will be available at: https://huggingface.co/amu-cai.