Despite commendable achievements made by existing work, prevailing multimodal sarcasm detection studies rely more on textual content over visual information. It unavoidably induces spurious correlations between textual words and labels, thereby significantly hindering the models' generalization capability. To address this problem, we define the task of out-of-distribution (OOD) multimodal sarcasm detection, which aims to evaluate models' generalizability when the word distribution is different in training and testing settings. Moreover, we propose a novel debiasing multimodal sarcasm detection framework with contrastive learning, which aims to mitigate the harmful effect of biased textual factors for robust OOD generalization. In particular, we first design counterfactual data augmentation to construct the positive samples with dissimilar word biases and negative samples with similar word biases. Subsequently, we devise an adapted debiasing contrastive learning mechanism to empower the model to learn robust task-relevant features and alleviate the adverse effect of biased words. Extensive experiments show the superiority of the proposed framework.
相關內容
As the text generation capabilities of large language models become increasingly prominent, recent studies have focused on controlling particular aspects of the generated text to make it more personalized. However, most research on controllable text generation focuses on controlling the content or modeling specific high-level/coarse-grained attributes that reflect authors' writing styles, such as formality, domain, or sentiment. In this paper, we focus on controlling fine-grained attributes spanning multiple linguistic dimensions, such as lexical and syntactic attributes. We introduce a novel benchmark to train generative models and evaluate their ability to generate personalized text based on multiple fine-grained linguistic attributes. We systematically investigate the performance of various large language models on our benchmark and draw insights from the factors that impact their performance. We make our code, data, and pretrained models publicly available.
Direct alignment from preferences (DAP) methods, such as DPO, have recently emerged as efficient alternatives to reinforcement learning from human feedback (RLHF), that do not require a separate reward model. However, the preference datasets used in DAP methods are usually collected ahead of training and never updated, thus the feedback is purely offline. Moreover, responses in these datasets are often sampled from a language model distinct from the one being aligned, and since the model evolves over training, the alignment phase is inevitably off-policy. In this study, we posit that online feedback is key and improves DAP methods. Our method, online AI feedback (OAIF), uses an LLM as annotator: on each training iteration, we sample two responses from the current model and prompt the LLM annotator to choose which one is preferred, thus providing online feedback. Despite its simplicity, we demonstrate via human evaluation in several tasks that OAIF outperforms both offline DAP and RLHF methods. We further show that the feedback leveraged in OAIF is easily controllable, via instruction prompts to the LLM annotator.
Controllable layout generation refers to the process of creating a plausible visual arrangement of elements within a graphic design (e.g., document and web designs) with constraints representing design intentions. Although recent diffusion-based models have achieved state-of-the-art FID scores, they tend to exhibit more pronounced misalignment compared to earlier transformer-based models. In this work, we propose the $\textbf{LA}$yout $\textbf{C}$onstraint diffusion mod$\textbf{E}$l (LACE), a unified model to handle a broad range of layout generation tasks, such as arranging elements with specified attributes and refining or completing a coarse layout design. The model is based on continuous diffusion models. Compared with existing methods that use discrete diffusion models, continuous state-space design can enable the incorporation of differentiable aesthetic constraint functions in training. For conditional generation, we introduce conditions via masked input. Extensive experiment results show that LACE produces high-quality layouts and outperforms existing state-of-the-art baselines.
Recommender systems have made significant strides in various industries, primarily driven by extensive efforts to enhance recommendation accuracy. However, this pursuit of accuracy has inadvertently given rise to echo chamber/filter bubble effects. Especially in industry, it could impair user's experiences and prevent user from accessing a wider range of items. One of the solutions is to take diversity into account. However, most of existing works focus on user's explicit preferences, while rarely exploring user's non-interaction preferences. These neglected non-interaction preferences are especially important for broadening user's interests in alleviating echo chamber/filter bubble effects.Therefore, in this paper, we first define diversity as two distinct definitions, i.e., user-explicit diversity (U-diversity) and user-item non-interaction diversity (N-diversity) based on user historical behaviors. Then, we propose a succinct and effective method, named as Controllable Category Diversity Framework (CCDF) to achieve both high U-diversity and N-diversity simultaneously.Specifically, CCDF consists of two stages, User-Category Matching and Constrained Item Matching. The User-Category Matching utilizes the DeepU2C model and a combined loss to capture user's preferences in categories, and then selects the top-$K$ categories with a controllable parameter $K$.These top-$K$ categories will be used as trigger information in Constrained Item Matching. Offline experimental results show that our proposed DeepU2C outperforms state-of-the-art diversity-oriented methods, especially on N-diversity task. The whole framework is validated in a real-world production environment by conducting online A/B testing.
This work establishes new convergence guarantees for gradient descent in smooth convex optimization via a computer-assisted analysis technique. Our theory allows nonconstant stepsize policies with frequent long steps potentially violating descent by analyzing the overall effect of many iterations at once rather than the typical one-iteration inductions used in most first-order method analyses. We show that long steps, which may increase the objective value in the short term, lead to provably faster convergence in the long term. A conjecture towards proving a faster $O(1/T\log T)$ rate for gradient descent is also motivated along with simple numerical validation.
Motivated by developments in machine learning technologies, unsupervised learning (UL)-based solvers for CO problems have recently been proposed. These solvers train a neural network that outputs a solution by optimizing the CO objective directly. UL-based solvers have several advantages over traditional methods. However, various studies have shown that these solvers underperform compared to greedy algorithms for complex CO problems. In addition, these solvers employ a continuous relaxation strategy; thus, post-learning rounding from the continuous space back to the original discrete space is required, undermining the robustness of the results. To address these problems, we propose the continuous relaxation annealing (CRA) strategy. The CRA introduces a penalty term to control the continuity and discreteness of the relaxed variables and eliminate local optima. In addition, the CRA implements an annealing process for the penalty term that initially prioritizes continuous solutions and progressively transitions towards discreet solutions until the relaxed variables become nearly discrete, eliminating the artificial rounding. Experimental results demonstrate that the CRA significantly enhances the UL-based solvers, outperforming both existing UL-based solvers and greedy algorithms for complex CO problems.
The sequential recommendation problem has attracted considerable research attention in the past few years, leading to the rise of numerous recommendation models. In this work, we explore how Large Language Models (LLMs), which are nowadays introducing disruptive effects in many AI-based applications, can be used to build or improve sequential recommendation approaches. Specifically, we design three orthogonal approaches and hybrids of those to leverage the power of LLMs in different ways. In addition, we investigate the potential of each approach by focusing on its comprising technical aspects and determining an array of alternative choices for each one. We conduct extensive experiments on three datasets and explore a large variety of configurations, including different language models and baseline recommendation models, to obtain a comprehensive picture of the performance of each approach. Among other observations, we highlight that initializing state-of-the-art sequential recommendation models such as BERT4Rec or SASRec with embeddings obtained from an LLM can lead to substantial performance gains in terms of accuracy. Furthermore, we find that fine-tuning an LLM for recommendation tasks enables it to learn not only the tasks, but also concepts of a domain to some extent. We also show that fine-tuning OpenAI GPT leads to considerably better performance than fine-tuning Google PaLM 2. Overall, our extensive experiments indicate a huge potential value of leveraging LLMs in future recommendation approaches. We publicly share the code and data of our experiments to ensure reproducibility.
Recent contrastive representation learning methods rely on estimating mutual information (MI) between multiple views of an underlying context. E.g., we can derive multiple views of a given image by applying data augmentation, or we can split a sequence into views comprising the past and future of some step in the sequence. Contrastive lower bounds on MI are easy to optimize, but have a strong underestimation bias when estimating large amounts of MI. We propose decomposing the full MI estimation problem into a sum of smaller estimation problems by splitting one of the views into progressively more informed subviews and by applying the chain rule on MI between the decomposed views. This expression contains a sum of unconditional and conditional MI terms, each measuring modest chunks of the total MI, which facilitates approximation via contrastive bounds. To maximize the sum, we formulate a contrastive lower bound on the conditional MI which can be approximated efficiently. We refer to our general approach as Decomposed Estimation of Mutual Information (DEMI). We show that DEMI can capture a larger amount of MI than standard non-decomposed contrastive bounds in a synthetic setting, and learns better representations in a vision domain and for dialogue generation.
Triple extraction is an essential task in information extraction for natural language processing and knowledge graph construction. In this paper, we revisit the end-to-end triple extraction task for sequence generation. Since generative triple extraction may struggle to capture long-term dependencies and generate unfaithful triples, we introduce a novel model, contrastive triple extraction with a generative transformer. Specifically, we introduce a single shared transformer module for encoder-decoder-based generation. To generate faithful results, we propose a novel triplet contrastive training object. Moreover, we introduce two mechanisms to further improve model performance (i.e., batch-wise dynamic attention-masking and triple-wise calibration). Experimental results on three datasets (i.e., NYT, WebNLG, and MIE) show that our approach achieves better performance than that of baselines.
We propose a new method for event extraction (EE) task based on an imitation learning framework, specifically, inverse reinforcement learning (IRL) via generative adversarial network (GAN). The GAN estimates proper rewards according to the difference between the actions committed by the expert (or ground truth) and the agent among complicated states in the environment. EE task benefits from these dynamic rewards because instances and labels yield to various extents of difficulty and the gains are expected to be diverse -- e.g., an ambiguous but correctly detected trigger or argument should receive high gains -- while the traditional RL models usually neglect such differences and pay equal attention on all instances. Moreover, our experiments also demonstrate that the proposed framework outperforms state-of-the-art methods, without explicit feature engineering.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191