Computation and Language (cs.CL)

Abstract: This thesis explores challenges in semantic parsing, specifically focusing on scenarios with limited data and computational resources. It offers solutions using techniques like automatic data curation, knowledge transfer, active learning, and continual learning. For tasks with no parallel training data, the thesis proposes generating synthetic training examples from structured database schemas. When there is abundant data in a source domain but limited parallel data in a target domain, knowledge from the source is leveraged to improve parsing in the target domain. For multilingual situations with limited data in the target languages, the thesis introduces a method to adapt parsers using a limited human translation budget. Active learning is applied to select source-language samples for manual translation, maximizing parser performance in the target language. In addition, an alternative method is also proposed to utilize machine translation services, supplemented by human-translated data, to train a more effective parser. When computational resources are limited, a continual learning approach is introduced to minimize training time and computational memory. This maintains the parser's efficiency in previously learned tasks while adapting it to new tasks, mitigating the problem of catastrophic forgetting. Overall, the thesis provides a comprehensive set of methods to improve semantic parsing in resource-constrained conditions.

Abstract: Sifting through vast textual data and summarizing key information imposes a substantial burden on how clinicians allocate their time. Although large language models (LLMs) have shown immense promise in natural language processing (NLP) tasks, their efficacy across diverse clinical summarization tasks has not yet been rigorously examined. In this work, we employ domain adaptation methods on eight LLMs, spanning six datasets and four distinct summarization tasks: radiology reports, patient questions, progress notes, and doctor-patient dialogue. Our thorough quantitative assessment reveals trade-offs between models and adaptation methods in addition to instances where recent advances in LLMs may not lead to improved results. Further, in a clinical reader study with six physicians, we depict that summaries from the best adapted LLM are preferable to human summaries in terms of completeness and correctness. Our ensuing qualitative analysis delineates mutual challenges faced by both LLMs and human experts. Lastly, we correlate traditional quantitative NLP metrics with reader study scores to enhance our understanding of how these metrics align with physician preferences. Our research marks the first evidence of LLMs outperforming human experts in clinical text summarization across multiple tasks. This implies that integrating LLMs into clinical workflows could alleviate documentation burden, empowering clinicians to focus more on personalized patient care and other irreplaceable human aspects of medicine.

Abstract: Despite the progress we have recorded in the last few years in multilingual natural language processing, evaluation is typically limited to a small set of languages with available datasets which excludes a large number of low-resource languages. In this paper, we created SIB-200 -- a large-scale open-sourced benchmark dataset for topic classification in 200 languages and dialects to address the lack of evaluation dataset for Natural Language Understanding (NLU). For many of the languages covered in SIB-200, this is the first publicly available evaluation dataset for NLU. The dataset is based on Flores-200 machine translation corpus. We annotated the English portion of the dataset and extended the sentence-level annotation to the remaining 203 languages covered in the corpus. Despite the simplicity of this task, our evaluation in full-supervised setting, cross-lingual transfer setting and prompting of large language model setting show that there is still a large gap between the performance of high-resource and low-resource languages when multilingual evaluation is scaled to numerous world languages. We found that languages unseen during the pre-training of multilingual language models, under-represented language families (like Nilotic and Altantic-Congo), and languages from the regions of Africa, Americas, Oceania and South East Asia, often have the lowest performance on our topic classification dataset. We hope our dataset will encourage a more inclusive evaluation of multilingual language models on a more diverse set of languages. https://github.com/dadelani/sib-200

Abstract: Large Language Models (LLMs) have demonstrated impressive performance on Natural Language Processing (NLP) tasks, such as Question Answering, Summarization, and Classification. The use of LLMs as evaluators, that can rank or score the output of other models (usually LLMs) has become increasingly popular, due to the limitations of current evaluation techniques including the lack of appropriate benchmarks, metrics, cost, and access to human annotators. While LLMs are capable of handling approximately 100 languages, the majority of languages beyond the top 20 lack systematic evaluation across various tasks, metrics, and benchmarks. This creates an urgent need to scale up multilingual evaluation to ensure a precise understanding of LLM performance across diverse languages. LLM-based evaluators seem like the perfect solution to this problem, as they do not require human annotators, human-created references, or benchmarks and can theoretically be used to evaluate any language covered by the LLM. In this paper, we investigate whether LLM-based evaluators can help scale up multilingual evaluation. Specifically, we calibrate LLM-based evaluation against 20k human judgments of five metrics across three text-generation tasks in eight languages. Our findings indicate that LLM-based evaluators may exhibit bias towards higher scores and should be used with caution and should always be calibrated with a dataset of native speaker judgments, particularly in low-resource and non-Latin script languages.

Abstract: This paper proposes a direct text to speech translation system using discrete acoustic units. This framework employs text in different source languages as input to generate speech in the target language without the need for text transcriptions in this language. Motivated by the success of acoustic units in previous works for direct speech to speech translation systems, we use the same pipeline to extract the acoustic units using a speech encoder combined with a clustering algorithm. Once units are obtained, an encoder-decoder architecture is trained to predict them. Then a vocoder generates speech from units. Our approach for direct text to speech translation was tested on the new CVSS corpus with two different text mBART models employed as initialisation. The systems presented report competitive performance for most of the language pairs evaluated. Besides, results show a remarkable improvement when initialising our proposed architecture with a model pre-trained with more languages.

Abstract: In this work, we present a web application named DBLPLink, which performs entity linking over the DBLP scholarly knowledge graph. DBLPLink uses text-to-text pre-trained language models, such as T5, to produce entity label spans from an input text question. Entity candidates are fetched from a database based on the labels, and an entity re-ranker sorts them based on entity embeddings, such as TransE, DistMult and ComplEx. The results are displayed so that users may compare and contrast the results between T5-small, T5-base and the different KG embeddings used. The demo can be accessed at https://ltdemos.informatik.uni-hamburg.de/dblplink/.

Abstract: Implicit discourse relation recognition (IDRR) aims at recognizing the discourse relation between two text segments without an explicit connective. Recently, the prompt learning has just been applied to the IDRR task with great performance improvements over various neural network-based approaches. However, the discrete nature of the state-art-of-art prompting approach requires manual design of templates and answers, a big hurdle for its practical applications. In this paper, we propose a continuous version of prompt learning together with connective knowledge distillation, called AdaptPrompt, to reduce manual design efforts via continuous prompting while further improving performance via knowledge transfer. In particular, we design and train a few virtual tokens to form continuous templates and automatically select the most suitable one by gradient search in the embedding space. We also design an answer-relation mapping rule to generate a few virtual answers as the answer space. Furthermore, we notice the importance of annotated connectives in the training dataset and design a teacher-student architecture for knowledge transfer. Experiments on the up-to-date PDTB Corpus V3.0 validate our design objectives in terms of the better relation recognition performance over the state-of-the-art competitors.

Abstract: We present new supertaggers trained on HPSG-based treebanks. These treebanks feature high-quality annotation based on a well-developed linguistic theory and include diverse and challenging test datasets, beyond the usual WSJ section 23 and Wikipedia data. HPSG supertagging has previously relied on MaxEnt-based models. We use SVM and neural CRF- and BERT-based methods and show that both SVM and neural supertaggers achieve considerably higher accuracy compared to the baseline. Our fine-tuned BERT-based tagger achieves 97.26% accuracy on 1000 sentences from WSJ23 and 93.88% on the completely out-of-domain The Cathedral and the Bazaar (cb)). We conclude that it therefore makes sense to integrate these new supertaggers into modern HPSG parsers, and we also hope that the diverse and difficult datasets we used here will gain more popularity in the field. We contribute the complete dataset reformatted for token classification.

Abstract: We introduce C-Pack, a package of resources that significantly advance the field of general Chinese embeddings. C-Pack includes three critical resources. 1) C-MTEB is a comprehensive benchmark for Chinese text embeddings covering 6 tasks and 35 datasets. 2) C-MTP is a massive text embedding dataset curated from labeled and unlabeled Chinese corpora for training embedding models. 3) C-TEM is a family of embedding models covering multiple sizes. Our models outperform all prior Chinese text embeddings on C-MTEB by up to +10% upon the time of the release. We also integrate and optimize the entire suite of training methods for C-TEM. Along with our resources on general Chinese embedding, we release our data and models for English text embeddings. The English models achieve state-of-the-art performance on MTEB benchmark; meanwhile, our released English data is 2 times larger than the Chinese data. All these resources are made publicly available at https://github.com/FlagOpen/FlagEmbedding.

Abstract: Credibility signals represent a wide range of heuristics that are typically used by journalists and fact-checkers to assess the veracity of online content. Automating the task of credibility signal extraction, however, is very challenging as it requires high-accuracy signal-specific extractors to be trained, while there are currently no sufficiently large datasets annotated with all credibility signals. This paper investigates whether large language models (LLMs) can be prompted effectively with a set of 18 credibility signals to produce weak labels for each signal. We then aggregate these potentially noisy labels using weak supervision in order to predict content veracity. We demonstrate that our approach, which combines zero-shot LLM credibility signal labeling and weak supervision, outperforms state-of-the-art classifiers on two misinformation datasets without using any ground-truth labels for training. We also analyse the contribution of the individual credibility signals towards predicting content veracity, which provides new valuable insights into their role in misinformation detection.

Abstract: The Multimodal Video Search by Examples (MVSE) project investigates using video clips as the query term for information retrieval, rather than the more traditional text query. This enables far richer search modalities such as images, speaker, content, topic, and emotion. A key element for this process is highly rapid, flexible, search to support large archives, which in MVSE is facilitated by representing video attributes by embeddings. This work aims to mitigate any performance loss from this rapid archive search by examining reranking approaches. In particular, zero-shot reranking methods using large language models are investigated as these are applicable to any video archive audio content. Performance is evaluated for topic-based retrieval on a publicly available video archive, the BBC Rewind corpus. Results demonstrate that reranking can achieve improved retrieval ranking without the need for any task-specific training data.

Abstract: Despite the success of neural models in solving reasoning tasks, their compositional generalization capabilities remain unclear. In this work, we propose a new setting of the structured explanation generation task to facilitate compositional reasoning research. Previous works found that symbolic methods achieve superior compositionality by using pre-defined inference rules for iterative reasoning. But these approaches rely on brittle symbolic transfers and are restricted to well-defined tasks. Hence, we propose a dynamic modularized reasoning model, MORSE, to improve the compositional generalization of neural models. MORSE factorizes the inference process into a combination of modules, where each module represents a functional unit. Specifically, we adopt modularized self-attention to dynamically select and route inputs to dedicated heads, which specializes them to specific functions. We conduct experiments for increasing lengths and shapes of reasoning trees on two benchmarks to test MORSE's compositional generalization abilities, and find it outperforms competitive baselines. Model ablation and deeper analyses show the effectiveness of dynamic reasoning modules and their generalization abilities.

Abstract: Data visualization has emerged as an effective tool for getting insights from massive datasets. Due to the hardness of manipulating the programming languages of data visualization, automatic data visualization generation from natural languages (Text-to-Vis) is becoming increasingly popular. Despite the plethora of research effort on the English Text-to-Vis, studies have yet to be conducted on data visualization generation from questions in Chinese. Motivated by this, we propose a Chinese Text-to-Vis dataset in the paper and demonstrate our first attempt to tackle this problem. Our model integrates multilingual BERT as the encoder, boosts the cross-lingual ability, and infuses the $n$-gram information into our word representation learning. Our experimental results show that our dataset is challenging and deserves further research.

Abstract: This paper presents a novel evaluation approach to text-based speaker diarization (SD), tackling the limitations of traditional metrics that do not account for any contextual information in text. Two new metrics are proposed, Text-based Diarization Error Rate and Diarization F1, which perform utterance- and word-level evaluations by aligning tokens in reference and hypothesis transcripts. Our metrics encompass more types of errors compared to existing ones, allowing us to make a more comprehensive analysis in SD. To align tokens, a multiple sequence alignment algorithm is introduced that supports multiple sequences in the reference while handling high-dimensional alignment to the hypothesis using dynamic programming. Our work is packaged into two tools, align4d providing an API for our alignment algorithm and TranscribeView for visualizing and evaluating SD errors, which can greatly aid in the creation of high-quality data, fostering the advancement of dialogue systems.

Abstract: Conversational recommender systems (CRS) generate recommendations through an interactive process. However, not all CRS approaches use human conversations as their source of interaction data; the majority of prior CRS work simulates interactions by exchanging entity-level information. As a result, claims of prior CRS work do not generalise to real-world settings where conversations take unexpected turns, or where conversational and intent understanding is not perfect. To tackle this challenge, the research community has started to examine holistic CRS, which are trained using conversational data collected from real-world scenarios. Despite their emergence, such holistic approaches are under-explored. We present a comprehensive survey of holistic CRS methods by summarizing the literature in a structured manner. Our survey recognises holistic CRS approaches as having three components: 1) a backbone language model, the optional use of 2) external knowledge, and/or 3) external guidance. We also give a detailed analysis of CRS datasets and evaluation methods in real application scenarios. We offer our insight as to the current challenges of holistic CRS and possible future trends.

Abstract: While recent advancements in the capabilities and widespread accessibility of generative language models, such as ChatGPT (OpenAI, 2022), have brought about various benefits by generating fluent human-like text, the task of distinguishing between human- and large language model (LLM) generated text has emerged as a crucial problem. These models can potentially deceive by generating artificial text that appears to be human-generated. This issue is particularly significant in domains such as law, education, and science, where ensuring the integrity of text is of the utmost importance. This survey provides an overview of the current approaches employed to differentiate between texts generated by humans and ChatGPT. We present an account of the different datasets constructed for detecting ChatGPT-generated text, the various methods utilized, what qualitative analyses into the characteristics of human versus ChatGPT-generated text have been performed, and finally, summarize our findings into general insights

Abstract: While the recently introduced Tree of Thoughts (ToT) has heralded advancements in allowing Large Language Models (LLMs) to reason through foresight and backtracking for global decision-making, it has overlooked the inherent local uncertainties in intermediate decision points or "thoughts". These local uncertainties, intrinsic to LLMs given their potential for diverse responses, remain a significant concern in the reasoning process. Addressing this pivotal gap, we introduce the Tree of Uncertain Thoughts (TouT) - a reasoning framework tailored for LLMs. Our TouT effectively leverages Monte Carlo Dropout to quantify uncertainty scores associated with LLMs' diverse local responses at these intermediate steps. By marrying this local uncertainty quantification with global search algorithms, TouT enhances the model's precision in response generation. We substantiate our approach with rigorous experiments on two demanding planning tasks: Game of 24 and Mini Crosswords. The empirical evidence underscores TouT's superiority over both ToT and chain-of-thought prompting methods.

Abstract: Large-scale self-supervised pre-trained speech encoders outperform conventional approaches in speech recognition and translation tasks. Due to the high cost of developing these large models, building new encoders for new tasks and deploying them to on-device applications are infeasible. Prior studies propose model compression methods to address this issue, but those works focus on smaller models and less realistic tasks. Thus, we propose Contrastive Layer-to-layer Distillation (CoLLD), a novel knowledge distillation method to compress pre-trained speech encoders by leveraging masked prediction and contrastive learning to train student models to copy the behavior of a large teacher model. CoLLD outperforms prior methods and closes the gap between small and large models on multilingual speech-to-text translation and recognition benchmarks.

Abstract: The phonological discrepancies between a speaker's native (L1) and the non-native language (L2) serves as a major factor for mispronunciation. This paper introduces a novel multilingual MDD architecture, L1-MultiMDD, enriched with L1-aware speech representation. An end-to-end speech encoder is trained on the input signal and its corresponding reference phoneme sequence. First, an attention mechanism is deployed to align the input audio with the reference phoneme sequence. Afterwards, the L1-L2-speech embedding are extracted from an auxiliary model, pretrained in a multi-task setup identifying L1 and L2 language, and are infused with the primary network. Finally, the L1-MultiMDD is then optimized for a unified multilingual phoneme recognition task using connectionist temporal classification (CTC) loss for the target languages: English, Arabic, and Mandarin. Our experiments demonstrate the effectiveness of the proposed L1-MultiMDD framework on both seen -- L2-ARTIC, LATIC, and AraVoiceL2v2; and unseen -- EpaDB and Speechocean762 datasets. The consistent gains in PER, and false rejection rate (FRR) across all target languages confirm our approach's robustness, efficacy, and generalizability.

Abstract: The meanings of words and phrases depend not only on where they are used (contexts) but also on who use them (writers). Pretrained language models (PLMs) are powerful tools for capturing context, but they are typically pretrained and fine-tuned for universal use across different writers. This study aims to improve the accuracy of text understanding tasks by personalizing the fine-tuning of PLMs for specific writers. We focus on a general setting where only the plain text from target writers are available for personalization. To avoid the cost of fine-tuning and storing multiple copies of PLMs for different users, we exhaustively explore using writer-specific prompts to personalize a unified PLM. Since the design and evaluation of these prompts is an underdeveloped area, we introduce and compare different types of prompts that are possible in our setting. To maximize the potential of prompt-based personalized fine-tuning, we propose a personalized intermediate learning based on masked language modeling to extract task-independent traits of writers' text. Our experiments, using multiple tasks, datasets, and PLMs, reveal the nature of different prompts and the effectiveness of our intermediate learning approach.

Abstract: Recent advances in eXplainable AI (XAI) have provided new insights into how models for vision, language, and tabular data operate. However, few approaches exist for understanding speech models. Existing work focuses on a few spoken language understanding (SLU) tasks, and explanations are difficult to interpret for most users. We introduce a new approach to explain speech classification models. We generate easy-to-interpret explanations via input perturbation on two information levels. 1) Word-level explanations reveal how each word-related audio segment impacts the outcome. 2) Paralinguistic features (e.g., prosody and background noise) answer the counterfactual: ``What would the model prediction be if we edited the audio signal in this way?'' We validate our approach by explaining two state-of-the-art SLU models on two speech classification tasks in English and Italian. Our findings demonstrate that the explanations are faithful to the model's inner workings and plausible to humans. Our method and findings pave the way for future research on interpreting speech models.

Abstract: Research on pronunciation assessment systems focuses on utilizing phonetic and phonological aspects of non-native (L2) speech, often neglecting the rich layer of information hidden within the non-verbal cues. In this study, we proposed a novel pronunciation assessment framework, IntraVerbalPA. % The framework innovatively incorporates both fine-grained frame- and abstract utterance-level non-verbal cues, alongside the conventional speech and phoneme representations. Additionally, we introduce ''Goodness of phonemic-duration'' metric to effectively model duration distribution within the framework. Our results validate the effectiveness of the proposed IntraVerbalPA framework and its individual components, yielding performance that either matches or outperforms existing research works.

Abstract: Large Language Models (LLMs) have shown impressive performance across a variety of Artificial Intelligence (AI) and natural language processing tasks, such as content creation, report generation, etc. However, unregulated malign application of these models can create undesirable consequences such as generation of fake news, plagiarism, etc. As a result, accurate detection of AI-generated language can be crucial in responsible usage of LLMs. In this work, we explore 1) whether a certain body of text is AI generated or written by human, and 2) attribution of a specific language model in generating a body of text. Texts in both English and Spanish are considered. The datasets used in this study are provided as part of the Automated Text Identification (AuTexTification) shared task. For each of the research objectives stated above, we propose an ensemble neural model that generates probabilities from different pre-trained LLMs which are used as features to a Traditional Machine Learning (TML) classifier following it. For the first task of distinguishing between AI and human generated text, our model ranked in fifth and thirteenth place (with macro $F1$ scores of 0.733 and 0.649) for English and Spanish texts, respectively. For the second task on model attribution, our model ranked in first place with macro $F1$ scores of 0.625 and 0.653 for English and Spanish texts, respectively.

Abstract: Interactive Object Grasping (IOG) is the task of identifying and grasping the desired object via human-robot natural language interaction. Current IOG systems assume that a human user initially specifies the target object's category (e.g., bottle). Inspired by pragmatics, where humans often convey their intentions by relying on context to achieve goals, we introduce a new IOG task, Pragmatic-IOG, and the corresponding dataset, Intention-oriented Multi-modal Dialogue (IM-Dial). In our proposed task scenario, an intention-oriented utterance (e.g., "I am thirsty") is initially given to the robot. The robot should then identify the target object by interacting with a human user. Based on the task setup, we propose a new robotic system that can interpret the user's intention and pick up the target object, Pragmatic Object Grasping (PROGrasp). PROGrasp performs Pragmatic-IOG by incorporating modules for visual grounding, question asking, object grasping, and most importantly, answer interpretation for pragmatic inference. Experimental results show that PROGrasp is effective in offline (i.e., target object discovery) and online (i.e., IOG with a physical robot arm) settings.

Abstract: Effectively leveraging multimodal information from social media posts is essential to various downstream tasks such as sentiment analysis, sarcasm detection and hate speech classification. However, combining text and image information is challenging because of the idiosyncratic cross-modal semantics with hidden or complementary information present in matching image-text pairs. In this work, we aim to directly model this by proposing the use of two auxiliary losses jointly with the main task when fine-tuning any pre-trained multimodal model. Image-Text Contrastive (ITC) brings image-text representations of a post closer together and separates them from different posts, capturing underlying dependencies. Image-Text Matching (ITM) facilitates the understanding of semantic correspondence between images and text by penalizing unrelated pairs. We combine these objectives with five multimodal models, demonstrating consistent improvements across four popular social media datasets. Furthermore, through detailed analysis, we shed light on the specific scenarios and cases where each auxiliary task proves to be most effective.

Abstract: When considering the opening part of 1800 short stories, we find that the first dozen paragraphs of the average narrative follow an action principle as defined in arXiv:2309.06600. When the order of the paragraphs is shuffled, the average no longer exhibits this property. The findings show that there is a preferential direction we take in semantic space when starting a story, possibly related to a common Western storytelling tradition as implied by Aristotle in Poetics.

Abstract: Automatic identification of clinical trials for which a patient is eligible is complicated by the fact that trial eligibility is stated in natural language. A potential solution to this problem is to employ text classification methods for common types of eligibility criteria. In this study, we focus on seven common exclusion criteria in cancer trials: prior malignancy, human immunodeficiency virus, hepatitis B, hepatitis C, psychiatric illness, drug/substance abuse, and autoimmune illness. Our dataset consists of 764 phase III cancer trials with these exclusions annotated at the trial level. We experiment with common transformer models as well as a new pre-trained clinical trial BERT model. Our results demonstrate the feasibility of automatically classifying common exclusion criteria. Additionally, we demonstrate the value of a pre-trained language model specifically for clinical trials, which yields the highest average performance across all criteria.

Abstract: In recent years, large language models (LLMs) have shown remarkable capabilities at scale, particularly at generating text conditioned on a prompt. In our work, we investigate the use of LLMs to augment training data of small language models~(SLMs) with automatically generated counterfactual~(CF) instances -- i.e. minimally altered inputs -- in order to improve out-of-domain~(OOD) performance of SLMs in the extractive question answering~(QA) setup. We show that, across various LLM generators, such data augmentation consistently enhances OOD performance and improves model calibration for both confidence-based and rationale-augmented calibrator models. Furthermore, these performance improvements correlate with higher diversity of CF instances in terms of their surface form and semantic content. Finally, we show that CF augmented models which are easier to calibrate also exhibit much lower entropy when assigning importance, indicating that rationale-augmented calibrators prefer concise explanations.

Abstract: As language models are adapted by a more sophisticated and diverse set of users, the importance of guaranteeing that they provide factually correct information supported by verifiable sources is critical across fields of study & professions. This is especially the case for high-stakes fields, such as medicine and law, where the risk of propagating false information is high and can lead to undesirable societal consequences. Previous work studying factuality and attribution has not focused on analyzing these characteristics of language model outputs in domain-specific scenarios. In this work, we present an evaluation study analyzing various axes of factuality and attribution provided in responses from a few systems, by bringing domain experts in the loop. Specifically, we first collect expert-curated questions from 484 participants across 32 fields of study, and then ask the same experts to evaluate generated responses to their own questions. We also ask experts to revise answers produced by language models, which leads to ExpertQA, a high-quality long-form QA dataset with 2177 questions spanning 32 fields, along with verified answers and attributions for claims in the answers.

Abstract: Recent advances on large language models (LLMs) enable researchers and developers to build autonomous language agents that can automatically solve various tasks and interact with environments, humans, and other agents using natural language interfaces. We consider language agents as a promising direction towards artificial general intelligence and release Agents, an open-source library with the goal of opening up these advances to a wider non-specialist audience. Agents is carefully engineered to support important features including planning, memory, tool usage, multi-agent communication, and fine-grained symbolic control. Agents is user-friendly as it enables non-specialists to build, customize, test, tune, and deploy state-of-the-art autonomous language agents without much coding. The library is also research-friendly as its modularized design makes it easily extensible for researchers. Agents is available at https://github.com/aiwaves-cn/agents.

Abstract: Training large language models to follow instructions makes them perform better on a wide range of tasks, generally becoming more helpful. However, a perfectly helpful model will follow even the most malicious instructions and readily generate harmful content. In this paper, we raise concerns over the safety of models that only emphasize helpfulness, not safety, in their instruction-tuning. We show that several popular instruction-tuned models are highly unsafe. Moreover, we show that adding just 3% safety examples (a few hundred demonstrations) in the training set when fine-tuning a model like LLaMA can substantially improve their safety. Our safety-tuning does not make models significantly less capable or helpful as measured by standard benchmarks. However, we do find a behavior of exaggerated safety, where too much safety-tuning makes models refuse to respond to reasonable prompts that superficially resemble unsafe ones. Our study sheds light on trade-offs in training LLMs to follow instructions and exhibit safe behavior.

Abstract: In-context learning (ICL) i.e. showing LLMs only a few task-specific demonstrations has led to downstream gains with no task-specific fine-tuning required. However, LLMs are sensitive to the choice of prompts, and therefore a crucial research question is how to select good demonstrations for ICL. One effective strategy is leveraging semantic similarity between the ICL demonstrations and test inputs by using a text retriever, which however is sub-optimal as that does not consider the LLM's existing knowledge about that task. From prior work (Min et al., 2022), we already know that labels paired with the demonstrations bias the model predictions. This leads us to our hypothesis whether considering LLM's existing knowledge about the task, especially with respect to the output label space can help in a better demonstration selection strategy. Through extensive experimentation on three text classification tasks, we find that it is beneficial to not only choose semantically similar ICL demonstrations but also to choose those demonstrations that help resolve the inherent label ambiguity surrounding the test example. Interestingly, we find that including demonstrations that the LLM previously mis-classified and also fall on the test example's decision boundary, brings the most performance gain.

Abstract: Starting from the resurgence of deep learning, vision-language models (VLMs) benefiting from large language models (LLMs) have never been so popular. However, while LLMs can utilize extensive background knowledge and task information with in-context learning, most VLMs still struggle with understanding complex multi-modal prompts with multiple images. The issue can traced back to the architectural design of VLMs or pre-training data. Specifically, the current VLMs primarily emphasize utilizing multi-modal data with a single image some, rather than multi-modal prompts with interleaved multiple images and text. Even though some newly proposed VLMs could handle user prompts with multiple images, pre-training data does not provide more sophisticated multi-modal prompts than interleaved image and text crawled from the web. We propose MMICL to address the issue by considering both the model and data perspectives. We introduce a well-designed architecture capable of seamlessly integrating visual and textual context in an interleaved manner and MIC dataset to reduce the gap between the training data and the complex user prompts in real-world applications, including: 1) multi-modal context with interleaved images and text, 2) textual references for each image, and 3) multi-image data with spatial, logical, or temporal relationships. Our experiments confirm that MMICL achieves new stat-of-the-art zero-shot and few-shot performance on a wide range of general vision-language tasks, especially for complex reasoning benchmarks including MME and MMBench. Our analysis demonstrates that MMICL effectively deals with the challenge of complex multi-modal prompt understanding. The experiments on ScienceQA-IMG also show that MMICL successfully alleviates the issue of language bias in VLMs, which we believe is the reason behind the advanced performance of MMICL.