Previous research on expert advice-taking shows that humans exhibit two contradictory behaviors: on the one hand, people tend to overvalue their own opinions undervaluing the expert opinion, and on the other, people often defer to other people's advice even if the advice itself is rather obviously wrong. In our study, we conduct an exploratory evaluation of users' AI-advice accepting behavior when evaluating the truthfulness of a health-related statement in different "advice quality" settings. We find that even feedback that is confined to just stating that "the AI thinks that the statement is false/true" results in more than half of people moving their statement veracity assessment towards the AI suggestion. The different types of advice given influence the acceptance rates, but the sheer effect of getting a suggestion is often bigger than the suggestion-type effect.

Adult content detection still poses a great challenge for automation. Existing classifiers primarily focus on distinguishing between erotic and non-erotic texts. However, they often need more nuance in assessing the potential harm. Unfortunately, the content of this nature falls beyond the reach of generative models due to its potentially harmful nature. Ethical restrictions prohibit large language models (LLMs) from analyzing and classifying harmful erotics, let alone generating them to create synthetic datasets for other neural models. In such instances where data is scarce and challenging, a thorough analysis of the structure of such texts rather than a large model may offer a viable solution. Especially given that harmful erotic narratives, despite appearing similar to harmless ones, usually reveal their harmful nature first through contextual information hidden in the non-sexual parts of the narrative. This paper introduces a hybrid neural and rule-based context-aware system that leverages coreference resolution to identify harmful contextual cues in erotic content. Collaborating with professional moderators, we compiled a dataset and developed a classifier capable of distinguishing harmful from non-harmful erotic content. Our hybrid model, tested on Polish text, demonstrates a promising accuracy of 84% and a recall of 80%. Models based on RoBERTa and Longformer without explicit usage of coreference chains achieved significantly weaker results, underscoring the importance of coreference resolution in detecting such nuanced content as harmful erotics. This approach also offers the potential for enhanced visual explainability, supporting moderators in evaluating predictions and taking necessary actions to address harmful content.

Text sanitization is the task of redacting a document to mask all occurrences of (direct or indirect) personal identifiers, with the goal of concealing the identity of the individual(s) referred in it. In this paper, we consider a two-step approach to text sanitization and provide a detailed analysis of its empirical performance on two recently published datasets: the Text Anonymization Benchmark (Pil\'an et al., 2022) and a collection of Wikipedia biographies (Papadopoulou et al., 2022). The text sanitization process starts with a privacy-oriented entity recognizer that seeks to determine the text spans expressing identifiable personal information. This privacy-oriented entity recognizer is trained by combining a standard named entity recognition model with a gazetteer populated by person-related terms extracted from Wikidata. The second step of the text sanitization process consists in assessing the privacy risk associated with each detected text span, either isolated or in combination with other text spans. We present five distinct indicators of the re-identification risk, respectively based on language model probabilities, text span classification, sequence labelling, perturbations, and web search. We provide a contrastive analysis of each privacy indicator and highlight their benefits and limitations, notably in relation to the available labeled data.

The vast number of parameters in large language models (LLMs) endows them with remarkable capabilities, allowing them to excel in a variety of natural language processing tasks. However, this complexity also presents challenges, making LLMs difficult to train and inhibiting their ability to continuously assimilate new knowledge, which may lead to inaccuracies in their outputs. To mitigate these issues, this paper presents DynaMind, a novel continual learning framework designed for LLMs. DynaMind incorporates memory mechanisms to assimilate new knowledge and modular operators to enhance the model inference process with the newly assimilated knowledge, consequently improving the accuracies of LLMs' outputs. Benchmark experiments demonstrate DynaMind's effectiveness in overcoming these challenges. The code and demo of DynaMind are available on GitHub: https://github.com/Elfsong/DynaMind.

Over the past decade, Artificial Intelligence (AI) has had great success recently and is being used in a wide range of academic and industrial fields. More recently, LLMs have made rapid advancements that have propelled AI to a new level, enabling even more diverse applications and industrial domains with intelligence, particularly in areas like software engineering and natural language processing. Nevertheless, a number of emerging trustworthiness concerns and issues exhibited in LLMs have already recently received much attention, without properly solving which the widespread adoption of LLMs could be greatly hindered in practice. The distinctive characteristics of LLMs, such as the self-attention mechanism, extremely large model scale, and autoregressive generation schema, differ from classic AI software based on CNNs and RNNs and present new challenges for quality analysis. Up to the present, it still lacks universal and systematic analysis techniques for LLMs despite the urgent industrial demand. Towards bridging this gap, we initiate an early exploratory study and propose a universal analysis framework for LLMs, LUNA, designed to be general and extensible, to enable versatile analysis of LLMs from multiple quality perspectives in a human-interpretable manner. In particular, we first leverage the data from desired trustworthiness perspectives to construct an abstract model as an auxiliary analysis asset, which is empowered by various abstract model construction methods. To assess the quality of the abstract model, we collect and define a number of evaluation metrics, aiming at both abstract model level and the semantics level. Then, the semantics, which is the degree of satisfaction of the LLM w.r.t. the trustworthiness perspective, is bound to and enriches the abstract model with semantics, which enables more detailed analysis applications for diverse purposes.

Fact verification systems assess a claim's veracity based on evidence. An important consideration in designing them is faithfulness, i.e. generating explanations that accurately reflect the reasoning of the model. Recent works have focused on natural logic, which operates directly on natural language by capturing the semantic relation of spans between an aligned claim with its evidence via set-theoretic operators. However, these approaches rely on substantial resources for training, which are only available for high-resource languages. To this end, we propose to use question answering to predict natural logic operators, taking advantage of the generalization capabilities of instruction-tuned language models. Thus, we obviate the need for annotated training data while still relying on a deterministic inference system. In a few-shot setting on FEVER, our approach outperforms the best baseline by $4.3$ accuracy points, including a state-of-the-art pre-trained seq2seq natural logic system, as well as a state-of-the-art prompt-based classifier. Our system demonstrates its robustness and portability, achieving competitive performance on a counterfactual dataset and surpassing all approaches without further annotation on a Danish verification dataset. A human evaluation indicates that our approach produces more plausible proofs with fewer erroneous natural logic operators than previous natural logic-based systems.

Data augmentation is a widely used technique to address the problem of text classification when there is a limited amount of training data. Recent work often tackles this problem using large language models (LLMs) like GPT3 that can generate new examples given already available ones. In this work, we propose a method to generate more helpful augmented data by utilizing the LLM's abilities to follow instructions and perform few-shot classifications. Our specific PromptMix method consists of two steps: 1) generate challenging text augmentations near class boundaries; however, generating borderline examples increases the risk of false positives in the dataset, so we 2) relabel the text augmentations using a prompting-based LLM classifier to enhance the correctness of labels in the generated data. We evaluate the proposed method in challenging 2-shot and zero-shot settings on four text classification datasets: Banking77, TREC6, Subjectivity (SUBJ), and Twitter Complaints. Our experiments show that generating and, crucially, relabeling borderline examples facilitates the transfer of knowledge of a massive LLM like GPT3.5-turbo into smaller and cheaper classifiers like DistilBERT$_{base}$ and BERT$_{base}$. Furthermore, 2-shot PromptMix outperforms multiple 5-shot data augmentation methods on the four datasets. Our code is available at https://github.com/ServiceNow/PromptMix-EMNLP-2023.

This paper addresses the problem of generating questions from a given context and an answer, specifically focusing on questions that require multi-hop reasoning across an extended context. Previous studies have suggested that key phrase selection is essential for question generation (QG), yet it is still challenging to connect such disjointed phrases into meaningful questions, particularly for long context. To mitigate this issue, we propose MultiFactor, a novel QG framework based on multi-level content planning. Specifically, MultiFactor includes two components: FA-model, which simultaneously selects key phrases and generates full answers, and Q-model which takes the generated full answer as an additional input to generate questions. Here, full answer generation is introduced to connect the short answer with the selected key phrases, thus forming an answer-aware summary to facilitate QG. Both FA-model and Q-model are formalized as simple-yet-effective Phrase-Enhanced Transformers, our joint model for phrase selection and text generation. Experimental results show that our method outperforms strong baselines on two popular QG datasets. Our code is available at https://github.com/zeaver/MultiFactor.

Various types of social biases have been reported with pretrained Masked Language Models (MLMs) in prior work. However, multiple underlying factors are associated with an MLM such as its model size, size of the training data, training objectives, the domain from which pretraining data is sampled, tokenization, and languages present in the pretrained corpora, to name a few. It remains unclear as to which of those factors influence social biases that are learned by MLMs. To study the relationship between model factors and the social biases learned by an MLM, as well as the downstream task performance of the model, we conduct a comprehensive study over 39 pretrained MLMs covering different model sizes, training objectives, tokenization methods, training data domains and languages. Our results shed light on important factors often neglected in prior literature, such as tokenization or model objectives.

Open large language models (LLMs) with great performance in various tasks have significantly advanced the development of LLMs. However, they are far inferior to commercial models such as ChatGPT and GPT-4 when acting as agents to tackle complex tasks in the real world. These agent tasks employ LLMs as the central controller responsible for planning, memorization, and tool utilization, necessitating both fine-grained prompting methods and robust LLMs to achieve satisfactory performance. Though many prompting methods have been proposed to complete particular agent tasks, there is lack of research focusing on improving the agent capabilities of LLMs themselves without compromising their general abilities. In this work, we present AgentTuning, a simple and general method to enhance the agent abilities of LLMs while maintaining their general LLM capabilities. We construct AgentInstruct, a lightweight instruction-tuning dataset containing high-quality interaction trajectories. We employ a hybrid instruction-tuning strategy by combining AgentInstruct with open-source instructions from general domains. AgentTuning is used to instruction-tune the Llama 2 series, resulting in AgentLM. Our evaluations show that AgentTuning enables LLMs' agent capabilities without compromising general abilities. The AgentLM-70B is comparable to GPT-3.5-turbo on unseen agent tasks, demonstrating generalized agent capabilities. We open source the AgentInstruct and AgentLM-7B, 13B, and 70B models at https://github.com/THUDM/AgentTuning, serving open and powerful alternatives to commercial LLMs for agent tasks.