The presence of specific linguistic signals particular to a certain sub-group of people can be picked up by language models during training. This may lead to discrimination if the model has learnt to pick up on a certain group's language. If the model begins to associate specific language with a distinct group, any decisions made based upon this language would hold a strong correlation to a decision based on their protected characteristic. We explore a possible technique for bias mitigation in the form of simplification of text. The driving force of this idea is that simplifying text should standardise language to one way of speaking while keeping the same meaning. The experiment shows promising results as the classifier accuracy for predicting the sensitive attribute drops by up to 17% for the simplified data.
相關內容
This paper presents a novel framework for structured argumentation, named extend argumentative decision graph ($xADG$). It is an extension of argumentative decision graphs built upon Dung's abstract argumentation graphs. The $xADG$ framework allows for arguments to use boolean logic operators and multiple premises (supports) within their internal structure, resulting in more concise argumentation graphs that may be easier for users to understand. The study presents a methodology for construction of $xADGs$ and evaluates their size and predictive capacity for classification tasks of varying magnitudes. Resulting $xADGs$ achieved strong (balanced) accuracy, which was accomplished through an input decision tree, while also reducing the average number of supports needed to reach a conclusion. The results further indicated that it is possible to construct plausibly understandable $xADGs$ that outperform other techniques for building $ADGs$ in terms of predictive capacity and overall size. In summary, the study suggests that $xADG$ represents a promising framework to developing more concise argumentative models that can be used for classification tasks and knowledge discovery, acquisition, and refinement.
We present an extensible, mathematically-structured algebraic simplification library design. We structure the library using universal algebraic concepts: a free algebra -- fral -- and a free extension -- frex -- of an algebra by a set of variables. The library's dependently-typed API guarantees simplification modules, even user-defined ones, are terminating, sound, and complete with respect to a well-specified class of equations. Completeness offers intangible benefits in practice -- our main contribution is the novel design. Cleanly separating between the interface and implementation of simplification modules provides two new modularity axes. First, simplification modules share thousands of lines of infrastructure code dealing with term-representation, pretty-printing, certification, and macros/reflection. Second, new simplification modules can reuse existing ones. We demonstrate this design by developing simplification modules for monoid varieties: ordinary, commutative, and involutive. We implemented this design in the new Idris2 dependently-typed programming language, and in Agda.
Pretrained language models are publicly available and constantly finetuned for various real-life applications. As they become capable of grasping complex contextual information, harmful biases are likely increasingly intertwined with those models. This paper analyses gender bias in BERT models with two main contributions: First, a novel bias measure is introduced, defining biases as the difference in sentiment valuation of female and male sample versions. Second, we comprehensively analyse BERT's biases on the example of a realistic IMDB movie classifier. By systematically varying elements of the training pipeline, we can conclude regarding their impact on the final model bias. Seven different public BERT models in nine training conditions, i.e. 63 models in total, are compared. Almost all conditions yield significant gender biases. Results indicate that reflected biases stem from public BERT models rather than task-specific data, emphasising the weight of responsible usage.
Recommendation systems have witnessed significant advancements and have been widely used over the past decades. However, most traditional recommendation methods are task-specific and therefore lack efficient generalization ability. Recently, the emergence of ChatGPT has significantly advanced NLP tasks by enhancing the capabilities of conversational models. Nonetheless, the application of ChatGPT in the recommendation domain has not been thoroughly investigated. In this paper, we employ ChatGPT as a general-purpose recommendation model to explore its potential for transferring extensive linguistic and world knowledge acquired from large-scale corpora to recommendation scenarios. Specifically, we design a set of prompts and evaluate ChatGPT's performance on five recommendation scenarios. Unlike traditional recommendation methods, we do not fine-tune ChatGPT during the entire evaluation process, relying only on the prompts themselves to convert recommendation tasks into natural language tasks. Further, we explore the use of few-shot prompting to inject interaction information that contains user potential interest to help ChatGPT better understand user needs and interests. Comprehensive experimental results on Amazon Beauty dataset show that ChatGPT has achieved promising results in certain tasks and is capable of reaching the baseline level in others. We conduct human evaluations on two explainability-oriented tasks to more accurately evaluate the quality of contents generated by different models. And the human evaluations show ChatGPT can truly understand the provided information and generate clearer and more reasonable results. We hope that our study can inspire researchers to further explore the potential of language models like ChatGPT to improve recommendation performance and contribute to the advancement of the recommendation systems field.
Despite the tremendous success in text-to-image generative models, localized text-to-image generation (that is, generating objects or features at specific locations in an image while maintaining a consistent overall generation) still requires either explicit training or substantial additional inference time. In this work, we show that localized generation can be achieved by simply controlling cross attention maps during inference. With no additional training, model architecture modification or inference time, our proposed cross attention control (CAC) provides new open-vocabulary localization abilities to standard text-to-image models. CAC also enhances models that are already trained for localized generation when deployed at inference time. Furthermore, to assess localized text-to-image generation performance automatically, we develop a standardized suite of evaluations using large pretrained recognition models. Our experiments show that CAC improves localized generation performance with various types of location information ranging from bounding boxes to semantic segmentation maps, and enhances the compositional capability of state-of-the-art text-to-image generative models.
In a world increasingly reliant on artificial intelligence, it is more important than ever to consider the ethical implications of artificial intelligence on humanity. One key under-explored challenge is labeler bias, which can create inherently biased datasets for training and subsequently lead to inaccurate or unfair decisions in healthcare, employment, education, and law enforcement. Hence, we conducted a study to investigate and measure the existence of labeler bias using images of people from different ethnicities and sexes in a labeling task. Our results show that participants possess stereotypes that influence their decision-making process and that labeler demographics impact assigned labels. We also discuss how labeler bias influences datasets and, subsequently, the models trained on them. Overall, a high degree of transparency must be maintained throughout the entire artificial intelligence training process to identify and correct biases in the data as early as possible.
Model specification searches and modifications are commonly employed in covariance structure analysis (CSA) or structural equation modeling (SEM) to improve the goodness-of-fit. However, these practices can be susceptible to capitalizing on chance, as a model that fits one sample may not generalize to another sample from the same population. This paper introduces the improved Lagrange Multipliers (LM) test, which provides a reliable method for conducting a thorough model specification search and effectively identifying missing parameters. By leveraging the stepwise bootstrap method in the standard LM and Wald tests, our data-driven approach enhances the accuracy of parameter identification. The results from Monte Carlo simulations and two empirical applications in political science demonstrate the effectiveness of the improved LM test, particularly when dealing with small sample sizes and models with large degrees of freedom. This approach contributes to better statistical fit and addresses the issue of capitalization on chance in model specification.
The security of computer systems typically relies on a hardware root of trust. As vulnerabilities in hardware can have severe implications on a system, there is a need for techniques to support security verification activities. Assertion-based verification is a popular verification technique that involves capturing design intent in a set of assertions that can be used in formal verification or testing-based checking. However, writing security-centric assertions is a challenging task. In this work, we investigate the use of emerging large language models (LLMs) for code generation in hardware assertion generation for security, where primarily natural language prompts, such as those one would see as code comments in assertion files, are used to produce SystemVerilog assertions. We focus our attention on a popular LLM and characterize its ability to write assertions out of the box, given varying levels of detail in the prompt. We design an evaluation framework that generates a variety of prompts, and we create a benchmark suite comprising real-world hardware designs and corresponding golden reference assertions that we want to generate with the LLM.
Existing machine learning models have proven to fail when it comes to their performance for minority groups, mainly due to biases in data. In particular, datasets, especially social data, are often not representative of minorities. In this paper, we consider the problem of representation bias identification on image datasets without explicit attribute values. Using the notion of data coverage for detecting a lack of representation, we develop multiple crowdsourcing approaches. Our core approach, at a high level, is a divide and conquer algorithm that applies a search space pruning strategy to efficiently identify if a dataset misses proper coverage for a given group. We provide a different theoretical analysis of our algorithm, including a tight upper bound on its performance which guarantees its near-optimality. Using this algorithm as the core, we propose multiple heuristics to reduce the coverage detection cost across different cases with multiple intersectional/non-intersectional groups. We demonstrate how the pre-trained predictors are not reliable and hence not sufficient for detecting representation bias in the data. Finally, we adjust our core algorithm to utilize existing models for predicting image group(s) to minimize the coverage identification cost. We conduct extensive experiments, including live experiments on Amazon Mechanical Turk to validate our problem and evaluate our algorithms' performance.
This research delves into the construction and utilization of synthetic datasets, specifically within the telematics sphere, leveraging OpenAI's powerful language model, ChatGPT. Synthetic datasets present an effective solution to challenges pertaining to data privacy, scarcity, and control over variables - characteristics that make them particularly valuable for research pursuits. The utility of these datasets, however, largely depends on their quality, measured through the lenses of diversity, relevance, and coherence. To illustrate this data creation process, a hands-on case study is conducted, focusing on the generation of a synthetic telematics dataset. The experiment involved an iterative guidance of ChatGPT, progressively refining prompts and culminating in the creation of a comprehensive dataset for a hypothetical urban planning scenario in Columbus, Ohio. Upon generation, the synthetic dataset was subjected to an evaluation, focusing on the previously identified quality parameters and employing descriptive statistics and visualization techniques for a thorough analysis. Despite synthetic datasets not serving as perfect replacements for actual world data, their potential in specific use-cases, when executed with precision, is significant. This research underscores the potential of AI models like ChatGPT in enhancing data availability for complex sectors like telematics, thus paving the way for a myriad of new research opportunities.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191