Despite the remarkable performance of foundation vision-language models, the shared representation space for text and vision can also encode harmful label associations detrimental to fairness. While prior work has uncovered bias in vision-language models' (VLMs) classification performance across geography, work has been limited along the important axis of harmful label associations due to a lack of rich, labeled data. In this work, we investigate harmful label associations in the recently released Casual Conversations datasets containing more than 70,000 videos. We study bias in the frequency of harmful label associations across self-provided labels for age, gender, apparent skin tone, and physical adornments across several leading VLMs. We find that VLMs are $4-13$x more likely to harmfully classify individuals with darker skin tones. We also find scaling transformer encoder model size leads to higher confidence in harmful predictions. Finally, we find improvements on standard vision tasks across VLMs does not address disparities in harmful label associations.
Referring expression segmentation (RES), a task that involves localizing specific instance-level objects based on free-form linguistic descriptions, has emerged as a crucial frontier in human-AI interaction. It demands an intricate understanding of both visual and textual contexts and often requires extensive training data. This paper introduces RESMatch, the first semi-supervised learning (SSL) approach for RES, aimed at reducing reliance on exhaustive data annotation. Extensive validation on multiple RES datasets demonstrates that RESMatch significantly outperforms baseline approaches, establishing a new state-of-the-art. Although existing SSL techniques are effective in image segmentation, we find that they fall short in RES. Facing the challenges including the comprehension of free-form linguistic descriptions and the variability in object attributes, RESMatch introduces a trifecta of adaptations: revised strong perturbation, text augmentation, and adjustments for pseudo-label quality and strong-weak supervision. This pioneering work lays the groundwork for future research in semi-supervised learning for referring expression segmentation.
Code-switching entails mixing multiple languages. It is an increasingly occurring phenomenon in social media texts. Usually, code-mixed texts are written in a single script, even though the languages involved have different scripts. Pre-trained multilingual models primarily utilize the data in the native script of the language. In existing studies, the code-switched texts are utilized as they are. However, using the native script for each language can generate better representations of the text owing to the pre-trained knowledge. Therefore, a cross-language-script knowledge sharing architecture utilizing the cross attention and alignment of the representations of text in individual language scripts was proposed in this study. Experimental results on two different datasets containing Nepali-English and Hindi-English code-switched texts, demonstrate the effectiveness of the proposed method. The interpretation of the model using model explainability technique illustrates the sharing of language-specific knowledge between language-specific representations.
Rapid progress in multimodal large language models (MLLMs) highlights the need to introduce challenging yet realistic benchmarks to the academic community. Existing benchmarks primarily focus on simple natural image understanding, but Multi emerges as a cutting-edge benchmark for MLLMs, offering a comprehensive dataset for evaluating MLLMs against understanding complex figures and tables, and scientific questions. This benchmark, reflecting current realistic examination styles, provides multimodal inputs and requires responses that are either precise or open-ended, similar to real-life school tests. It challenges MLLMs with a variety of tasks, ranging from formula derivation to image detail analysis, and cross-modality reasoning. Multi includes over 18,000 questions, with a focus on science-based QA in diverse formats. We also introduce Multi-Elite, a 500-question subset for testing the extremities of MLLMs, and Multi-Extend, which enhances In-Context Learning research with more than 4,500 knowledge pieces. Our evaluation indicates significant potential for MLLM advancement, with GPT-4V achieving a 63.7% accuracy rate on Multi, in contrast to other MLLMs scoring between 31.3% and 53.7%. Multi serves not only as a robust evaluation platform but also paves the way for the development of expert-level AI.
Background: The semantics of entities extracted from a clinical text can be dramatically altered by modifiers, including entity negation, uncertainty, conditionality, severity, and subject. Existing models for determining modifiers of clinical entities involve regular expression or features weights that are trained independently for each modifier. Methods: We develop and evaluate a multi-task transformer architecture design where modifiers are learned and predicted jointly using the publicly available SemEval 2015 Task 14 corpus and a new Opioid Use Disorder (OUD) data set that contains modifiers shared with SemEval as well as novel modifiers specific for OUD. We evaluate the effectiveness of our multi-task learning approach versus previously published systems and assess the feasibility of transfer learning for clinical entity modifiers when only a portion of clinical modifiers are shared. Results: Our approach achieved state-of-the-art results on the ShARe corpus from SemEval 2015 Task 14, showing an increase of 1.1% on weighted accuracy, 1.7% on unweighted accuracy, and 10% on micro F1 scores. Conclusions: We show that learned weights from our shared model can be effectively transferred to a new partially matched data set, validating the use of transfer learning for clinical text modifiers
Several popular language models represent local contexts in an input text as bags of words. Such representations are naturally encoded by a sequence graph whose vertices are the distinct words occurring in x, with edges representing the (ordered) co-occurrence of two words within a sliding window of size w. However, this compressed representation is not generally bijective, and may introduce some degree of ambiguity. Some sequence graphs may admit several realizations as a sequence, while others may not admit any realization. In this paper, we study the realizability and ambiguity of sequence graphs from a combinatorial and computational point of view. We consider the existence and enumeration of realizations of a sequence graph under multiple settings: window size w, presence/absence of graph orientation, and presence/absence of weights (multiplicities). When w = 2, we provide polynomial time algorithms for realizability and enumeration in all cases except the undirected/weighted setting, where we show the #P-hardness of enumeration. For a window of size at least 3, we prove hardness of all variants, even when w is considered as a constant, with the notable exception of the undirected/unweighted case for which we propose an XP algorithms for both (realizability and enumeration) problems, tight due to a corresponding W[1]-hardness result. We conclude with an integer program formulation to solve the realizability problem, and with dynamic programming to solve the enumeration problem. This work leaves open the membership to NP for both problems, a non-trivial question due to the existence of minimum realizations having exponential size on the instance encoding.
We study the problem of generating plans for given natural language planning task requests. On one hand, LLMs excel at natural language processing but do not perform well on planning. On the other hand, classical planning tools excel at planning tasks but require input in a structured language such as the Planning Domain Definition Language (PDDL). We leverage the strengths of both the techniques by using an LLM for generating the PDDL representation (task PDDL) of planning task requests followed by using a classical planner for computing a plan. Unlike previous approaches that use LLMs for generating task PDDLs directly, our approach comprises of (a) translate: using an LLM only for generating a logically interpretable intermediate representation of natural language task descriptions, (b) infer: deriving additional logically dependent information from the intermediate representation using a logic reasoner (currently, Answer Set Programming solver), and (c) compile: generating the target task PDDL from the base and inferred information. We observe that using an LLM to only output the intermediate representation significantly reduces LLM errors. Consequently, TIC approach achieves, for at least one LLM, high accuracy on task PDDL generation for all seven domains of our evaluation dataset.
SemEval-2024 Task 8 introduces the challenge of identifying machine-generated texts from diverse Large Language Models (LLMs) in various languages and domains. The task comprises three subtasks: binary classification in monolingual and multilingual (Subtask A), multi-class classification (Subtask B), and mixed text detection (Subtask C). This paper focuses on Subtask A & B. Each subtask is supported by three datasets for training, development, and testing. To tackle this task, two methods: 1) using traditional machine learning (ML) with natural language preprocessing (NLP) for feature extraction, and 2) fine-tuning LLMs for text classification. The results show that transformer models, particularly LoRA-RoBERTa, exceed traditional ML methods in effectiveness, with majority voting being particularly effective in multilingual contexts for identifying machine-generated texts.
Modelling learning objects (LO) within their context enables the learner to advance from a basic, remembering-level, learning objective to a higher-order one, i.e., a level with an application- and analysis objective. While hierarchical data models are commonly used in digital learning platforms, using graph-based models enables representing the context of LOs in those platforms. This leads to a foundation for personalized recommendations of learning paths. In this paper, the transformation of hierarchical data models into knowledge graph (KG) models of LOs using text mining is introduced and evaluated. We utilize custom text mining pipelines to mine semantic relations between elements of an expert-curated hierarchical model. We evaluate the KG structure and relation extraction using graph quality-control metrics and the comparison of algorithmic semantic-similarities to expert-defined ones. The results show that the relations in the KG are semantically comparable to those defined by domain experts, and that the proposed KG improves representing and linking the contexts of LOs through increasing graph communities and betweenness centrality.
We explore the mechanism of in-context learning and propose a hypothesis using locate-and-project method. In shallow layers, the features of demonstrations are merged into their corresponding labels, and the features of the input text are aggregated into the last token. In deep layers, in-context heads make great contributions. In each in-context head, the value-output matrix extracts the labels' features. Query and key matrices compute the attention weights between the input text and each demonstration. The larger the attention weight is, the more label information is transferred into the last token for predicting the next word. Query and key matrices can be regarded as two towers for learning the similarity metric between the input text and each demonstration. Based on this hypothesis, we explain why imbalanced labels and demonstration order affect predictions. We conduct experiments on GPT2 large, Llama 7B, 13B and 30B. The results can support our analysis. Overall, our study provides a new method and a reasonable hypothesis for understanding the mechanism of in-context learning. Our code will be released on github.