Text embedding models have significantly contributed to advancements in natural language processing by adeptly capturing semantic properties of textual data. However, the ability of these models to generalize across a wide range of syntactic contexts remains under-explored. In this paper, we first develop an evaluation set, named \textbf{SR}, to scrutinize the capability for syntax understanding of text embedding models from two crucial syntactic aspects: Structural heuristics, and Relational understanding among concepts, as revealed by the performance gaps in previous studies. Our findings reveal that existing text embedding models have not sufficiently addressed these syntactic understanding challenges, and such ineffectiveness becomes even more apparent when evaluated against existing benchmark datasets. Furthermore, we conduct rigorous analysis to unearth factors that lead to such limitations and examine why previous evaluations fail to detect such ineffectiveness. Lastly, we propose strategies to augment the generalization ability of text embedding models in diverse syntactic scenarios. This study serves to highlight the hurdles associated with syntactic generalization and provides pragmatic guidance for boosting model performance across varied syntactic contexts.
相關內容
Large Pre-Trained Language Models have demonstrated state-of-the-art performance in different downstream tasks, including dialogue state tracking and end-to-end response generation. Nevertheless, most of the publicly available datasets and benchmarks on task-oriented dialogues focus on written conversations. Consequently, the robustness of the developed models to spoken interactions is unknown. In this work, we have evaluated the performance of LLMs for spoken task-oriented dialogues on the DSTC11 test sets. Due to the lack of proper spoken dialogue datasets, we have automatically transcribed a development set of spoken dialogues with a state-of-the-art ASR engine. We have characterized the ASR-error types and their distributions and simulated these errors in a large dataset of dialogues. We report the intrinsic (perplexity) and extrinsic (human evaluation) performance of fine-tuned GPT-2 and T5 models in two subtasks of response generation and dialogue state tracking, respectively. The results show that LLMs are not robust to spoken noise by default, however, fine-tuning/training such models on a proper dataset of spoken TODs can result in a more robust performance.
The enduring inability of image generative models to recreate intricate geometric features, such as those present in human hands and fingers has been an ongoing problem in image generation for nearly a decade. While strides have been made by increasing model sizes and diversifying training datasets, this issue remains prevalent across all models, from denoising diffusion models to Generative Adversarial Networks (GAN), pointing to a fundamental shortcoming in the underlying architectures. In this paper, we demonstrate how this problem can be mitigated by augmenting convolution layers geometric capabilities through providing them with a single input channel incorporating the relative $n$-dimensional Cartesian coordinate system. We show that this drastically improves quality of hand and face images generated by GANs and Variational AutoEncoders (VAE).
Advances in model editing through neuron pruning hold promise for removing undesirable concepts from large language models. However, it remains unclear whether models have the capacity to reacquire pruned concepts after editing. To investigate this, we evaluate concept relearning in models by tracking concept saliency and similarity in pruned neurons during retraining. Our findings reveal that models can quickly regain performance post-pruning by relocating advanced concepts to earlier layers and reallocating pruned concepts to primed neurons with similar semantics. This demonstrates that models exhibit polysemantic capacities and can blend old and new concepts in individual neurons. While neuron pruning provides interpretability into model concepts, our results highlight the challenges of permanent concept removal for improved model \textit{safety}. Monitoring concept reemergence and developing techniques to mitigate relearning of unsafe concepts will be important directions for more robust model editing. Overall, our work strongly demonstrates the resilience and fluidity of concept representations in LLMs post concept removal.
Concept-based learning improves a deep learning model's interpretability by explaining its predictions via human-understandable concepts. Deep learning models trained under this paradigm heavily rely on the assumption that neural networks can learn to predict the presence or absence of a given concept independently of other concepts. Recent work, however, strongly suggests that this assumption may fail to hold in Concept Bottleneck Models (CBMs), a quintessential family of concept-based interpretable architectures. In this paper, we investigate whether CBMs correctly capture the degree of conditional independence across concepts when such concepts are localised both spatially, by having their values entirely defined by a fixed subset of features, and semantically, by having their values correlated with only a fixed subset of predefined concepts. To understand locality, we analyse how changes to features outside of a concept's spatial or semantic locality impact concept predictions. Our results suggest that even in well-defined scenarios where the presence of a concept is localised to a fixed feature subspace, or whose semantics are correlated to a small subset of other concepts, CBMs fail to learn this locality. These results cast doubt upon the quality of concept representations learnt by CBMs and strongly suggest that concept-based explanations may be fragile to changes outside their localities.
Prompt injection attacks exploit vulnerabilities in large language models (LLMs) to manipulate the model into unintended actions or generate malicious content. As LLM integrated applications gain wider adoption, they face growing susceptibility to such attacks. This study introduces a novel evaluation framework for quantifying the resilience of applications. The framework incorporates innovative techniques designed to ensure representativeness, interpretability, and robustness. To ensure the representativeness of simulated attacks on the application, a meticulous selection process was employed, resulting in 115 carefully chosen attacks based on coverage and relevance. For enhanced interpretability, a second LLM was utilized to evaluate the responses generated from these simulated attacks. Unlike conventional malicious content classifiers that provide only a confidence score, the LLM-based evaluation produces a score accompanied by an explanation, thereby enhancing interpretability. Subsequently, a resilience score is computed by assigning higher weights to attacks with greater impact, thus providing a robust measurement of the application resilience. To assess the framework's efficacy, it was applied on two LLMs, namely Llama2 and ChatGLM. Results revealed that Llama2, the newer model exhibited higher resilience compared to ChatGLM. This finding substantiates the effectiveness of the framework, aligning with the prevailing notion that newer models tend to possess greater resilience. Moreover, the framework exhibited exceptional versatility, requiring only minimal adjustments to accommodate emerging attack techniques and classifications, thereby establishing itself as an effective and practical solution. Overall, the framework offers valuable insights that empower organizations to make well-informed decisions to fortify their applications against potential threats from prompt injection.
This paper explores the causal reasoning of large language models (LLMs) to enhance their interpretability and reliability in advancing artificial intelligence. Despite the proficiency of LLMs in a range of tasks, their potential for understanding causality requires further exploration. We propose a novel causal attribution model that utilizes "do-operators" for constructing counterfactual scenarios, allowing us to systematically quantify the influence of input numerical data and LLMs' pre-existing knowledge on their causal reasoning processes. Our newly developed experimental setup assesses LLMs' reliance on contextual information and inherent knowledge across various domains. Our evaluation reveals that LLMs' causal reasoning ability depends on the context and domain-specific knowledge provided, and supports the argument that "knowledge is, indeed, what LLMs principally require for sound causal reasoning". On the contrary, in the absence of knowledge, LLMs still maintain a degree of causal reasoning using the available numerical data, albeit with limitations in the calculations.
Feature attribution methods are popular in interpretable machine learning. These methods compute the attribution of each input feature to represent its importance, but there is no consensus on the definition of "attribution", leading to many competing methods with little systematic evaluation, complicated in particular by the lack of ground truth attribution. To address this, we propose a dataset modification procedure to induce such ground truth. Using this procedure, we evaluate three common methods: saliency maps, rationales, and attentions. We identify several deficiencies and add new perspectives to the growing body of evidence questioning the correctness and reliability of these methods applied on datasets in the wild. We further discuss possible avenues for remedy and recommend new attribution methods to be tested against ground truth before deployment. The code is available at \url{//github.com/YilunZhou/feature-attribution-evaluation}.
Non-convex optimization is ubiquitous in modern machine learning. Researchers devise non-convex objective functions and optimize them using off-the-shelf optimizers such as stochastic gradient descent and its variants, which leverage the local geometry and update iteratively. Even though solving non-convex functions is NP-hard in the worst case, the optimization quality in practice is often not an issue -- optimizers are largely believed to find approximate global minima. Researchers hypothesize a unified explanation for this intriguing phenomenon: most of the local minima of the practically-used objectives are approximately global minima. We rigorously formalize it for concrete instances of machine learning problems.
This paper presents a new approach for assembling graph neural networks based on framelet transforms. The latter provides a multi-scale representation for graph-structured data. With the framelet system, we can decompose the graph feature into low-pass and high-pass frequencies as extracted features for network training, which then defines a framelet-based graph convolution. The framelet decomposition naturally induces a graph pooling strategy by aggregating the graph feature into low-pass and high-pass spectra, which considers both the feature values and geometry of the graph data and conserves the total information. The graph neural networks with the proposed framelet convolution and pooling achieve state-of-the-art performance in many types of node and graph prediction tasks. Moreover, we propose shrinkage as a new activation for the framelet convolution, which thresholds the high-frequency information at different scales. Compared to ReLU, shrinkage in framelet convolution improves the graph neural network model in terms of denoising and signal compression: noises in both node and structure can be significantly reduced by accurately cutting off the high-pass coefficients from framelet decomposition, and the signal can be compressed to less than half its original size with the prediction performance well preserved.
The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191