AMAva Martinez

Ava Martinez

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Ava Martinez has a total of 17 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Adaptive Latent Interaction Reasoning for Multimodal Misinformation Analysis

    by Tyler Anderson, Madeline Brooks, Ava Martinez, Jordan Williams

    The rapid growth of online social platforms has fundamentally transformed the way information is produced, disseminated, and consumed, while simultaneously amplifying the societal impact of misleading and fabricated content. In response to this challenge, multimodal fake news detection has emerged as a critical research problem, aiming to jointly leverage textual and visual signals embedded in social media posts. Existing methods predominantly rely on direct fusion of unimodal representations or shallow cross-modal interactions, which often fail to explicitly model the semantic alignment and latent inconsistencies across modalities. In particular, the potential of contrastive learning paradigms for learning robust and semantically grounded multimodal representations in fake news scenarios remains underexplored. In this work, we introduce ALIGNER, an Adaptive Latent Interaction Guided coNtrastivE Reasoning framework designed for multimodal fake news detection. ALIGNER adopts a dual-encoder architecture to learn modality-specific semantic representations and employs cross-modal ontrastive learning to explicitly align visual and textual semantics. To address the inherent noise and ambiguity of image–text associations in real-world fake news data, wefurther propose a latent consistency objective that relaxes the rigid one-hot supervision imposed by conventional contrastive losses. This auxiliary learning signal enables the model to capture fine grained semantic relatedness among unpaired or weakly related multimodal samples. Building upon the aligned unimodal features, ALIGNER incorporates a dedicated cross-modal interaction module to capture higher-order correlations between visual and linguistic representations. Moreover, wedesign an attention-based aggregation mechanism equipped with an explicit guidance signal to adaptively weigh the contributions of different modalities during decision making, thereby enhancing both effectiveness and interpretability. Extensive experiments conducted on two widely adopted benchmarks, Twitter and Weibo, demonstrate that ALIGNER consistently surpasses existing state of-the-art approaches by a substantial margin, highlighting the advantages of adaptive contrastive reasoning for multimodal fake news detection.


  • Open Access

    Beyond References: Human-Aligned Caption Reliability Assessment

    by Jaxon Carter, Caleb Turner, Ava Martinez, Hailey Peterson

    Despite the rapid progress of modern image captioning systems, the reliability of generated captions in practical deployments often lags behind expectations. In critical scenarios—such as assistive technologies or human-AI interaction platforms—unreliable descriptions may undermine user trust and lead to serious usability issues. To mitigate this gap, we investigate the task of Caption Quality Estimation (CQE) without references, where the objective is to directly judge the appropriateness of a caption based only on its paired image. This paradigm allows untrustworthy outputs to be filtered during inference, offering a proactive safeguard for real-world captioning applications. We introduce VQAR, a novel reference-free framework explicitly crafted to approximate human perception of caption adequacy. Central to this framework is a large-scale dataset we collected, containing over 600,000 binary human judgments across roughly 55,000 $\langle image, caption \rangle$ pairs from 16,000 diverse images. Each annotation acts as a binary signal of visual-semantic compatibility, capturing whether humans deem a caption acceptable for its associated image. To demonstrate both reliability and scalability, we validate the dataset through consistency analyses and benchmark several CQE models. Moreover, we supplement the coarse binary annotations with a fine-grained subset of expert evaluations, enabling us to assess the generalization of learned models. Experimental results show that models trained exclusively on coarse judgments can nonetheless approximate nuanced human preferences, underlining the practicality of VQAR for large-scale deployment. Our contributions are threefold: (i) establishing a reference-independent framework for caption validation; (ii) curating a high-coverage dataset with over 600k human annotations; (iii) providing empirical benchmarks that highlight the difficulty and distinctiveness of CQE compared with conventional captioning or retrieval tasks. By eliminating the dependency on reference captions, VQAR offers a robust, human-centric path toward improving the trustworthiness of captioning systems in interactive and mission-critical environments.

  • Open Access

    Caption-Grounded Structural Parsing for Compound Scientific Visuals

    by Omar Al-Mansoori, Aiden Johnson, Ava Martinez, Oliver Smith

    The unprecedented growth of scholarly literature has triggered a parallel explosion in visual artifacts, particularly figures that encapsulate experimental findings. Strikingly, more than 30% of these figures are compound in nature—comprising multiple heterogeneous subfigures—thus presenting formidable obstacles to automated parsing and comprehension. Conventional retrieval and analysis pipelines are typically designed under the assumption that each figure embodies a single, coherent semantic theme. This assumption breaks down when applied to compound figures, where diverse and semantically independent components coexist. To overcome this limitation, we propose SEMCLIP, a layoutsensitive, semantics-driven framework tailored for figure decomposition. Instead of merely segmenting visual regions based on low-level appearance, SEMCLIP introduces the notion of master images: semantically aligned units constructed through explicit modeling of symbolic labels embedded within figures. The system employs a cascaded two-stage design. First, a label localization network identifies references, which encode both structural layout and semantic grouping. These anchors are then fused with learned descriptors of regional visual features, producing coherent segments aligned with caption semantics. To address difficulties posed by uneven annotation distributions and sparse symbolic cues, we develop a bifurcated training paradigm that independently refines detection sensitivity and classification robustness. Experimental results on a large-scale annotated benchmark confirm that SEMCLIP significantly outperforms heuristic- and detection-based baselines, achieving superior segmentation fidelity and improved alignment between visual segments and textual captions. This work establishes a new pathway toward semantically grounded interpretation of visual evidence in scholarly communication.

  • Open Access

    Fusion of Visual and Textual Data for Enhanced Semantic Representations

    by Lyra Sterling, Kairos Vale, Ava Martinez

    Generic text embeddings have demonstrated considerable success across a multitude of applications. However, these embeddings are typically derived by modeling the co-occurrence patterns within text-only corpora, which can limit their ability to generalize effectively across diverse contexts. In this study, we investigate methodologies that incorporate visual information into textual representations to overcome these limitations. Through extensive ablation studies, we introduce a novel and straightforward architecture named VisualText Fusion Network (VTFN). This architecture not only surpasses existing multimodal approaches on a range of well-established benchmark datasets but also achieves state-of-the-art performance on image-related textual datasets while utilizing significantly less training data. Our findings underscore the potential of integrating visual modalities to substantially enhance the robustness and applicability of text embeddings, paving the way for more nuanced and contextually rich semantic representations.

  • Open Access

    SensorySync: Multimodal Integration Framework for Unified Perceptual Understanding

    by Zephyr Lawson, Ava Martinez, Seraphina Quinn

    Generic text embeddings have demonstrated considerable success across a multitude of applications. However, these embeddings are typically derived by modeling the co-occurrence patterns within text-only corpora, which can limit their ability to generalize effectively across diverse contexts. In this study, we investigate methodologies that incorporate visual information into textual representations to overcome these limitations. Through extensive ablation studies, we introduce a novel and straightforward architecture named VisualText Fusion Network (VTFN). This architecture not only surpasses existing multimodal approaches on a range of well-established benchmark datasets but also achieves state-of-the-art performance on image-related textual datasets while utilizing significantly less training data. Our findings underscore the potential of integrating visual modalities to substantially enhance the robustness and applicability of text embeddings, paving the way for more nuanced and contextually rich semantic representations.

  • Open Access

    Multimodal Information Fusion with Neural Gating

    by Olivia Smith, Ava Martinez, Noah Brown

    In this study, we introduce an innovative framework for multimodal learning that leverages enhanced fusion gate units within gated neural network architectures. The proposed Fusion Gate Unit (FGU) serves as a pivotal component in neural network designs, aiming to derive a comprehensive intermediate representation by amalgamating data from diverse modalities. The FGU is adept at determining the extent to which each modality influences the unit's activation through the utilization of multiplicative gating mechanisms. We conducted evaluations on a multilabel genre classification task for movies, utilizing both plot summaries and poster images as input modalities. The results demonstrate that the FGU significantly elevates the macro F-score compared to single-modality approaches and surpasses existing fusion techniques, including mixture of experts models. Additionally, we present the MM-IMDb dataset alongside this publication, which, to our knowledge, represents the most extensive publicly accessible multimodal dataset for movie genre prediction to date. This dataset is expected to facilitate further research and development in the field of multimodal information processing.

  • Open Access

    Enhanced Modal Fusion Learning for Multimodal Sentiment Interpretation

    by Kayla Robinson, Ava Martinez, Ethan Turner

    Multimodal sentiment analysis is rapidly gaining traction due to its ability to comprehensively interpret opinions expressed in video content, which is ubiquitous across various digital platforms. Despite its promising potential, the field is hindered by the limited availability of high-quality, annotated datasets, which poses substantial challenges to the generalizability of predictive models. Models trained on such scarce data often inadvertently assign excessive importance to irrelevant features, such as personal attributes (e.g., eyewear), thereby diminishing their accuracy and robustness. To address this issue, we propose an Enhanced Modal Fusion Learning (EMFL) methodology aimed at significantly improving the generalization capabilities of neural networks. EMFL achieves this by optimizing the integration and interpretation processes of multimodal data, ensuring that sentiment-relevant features are prioritized over confounding attributes. Through extensive experiments conducted on multiple benchmark datasets, we demonstrate that EMFL consistently elevates the accuracy of sentiment predictions across verbal, acoustic, and visual modalities. These findings underscore EMFL's efficacy in mitigating the impact of non-relevant features and enhancing the overall performance of multimodal sentiment analysis models.