Large Language Models (LLMs) are capable of producing content that perpetuates stereotypes, discrimination, and toxicity.The recently proposed moral self-correction is a computationally efficient method for reducing harmful content in the responses of LLMs. However, the process of how injecting self-correction instructions can modify the behavior of LLMs remains under-explored. In this paper, we explore the effectiveness of moral self-correction by answering three research questions: (1) In what scenarios does moral self-correction work? (2) What are the internal mechanisms of LLMs, e.g., hidden states, that are influenced by moral self-correction instructions? (3) Is intrinsic moral self-correction actually superficial in terms of reduced immorality in hidden states? We argue that self-correction can help LLMs find a shortcut to more morally correct output, rather than truly reducing the immorality stored in hidden states.Through empirical investigation with tasks of language generation and multi-choice question answering, we conclude: (i) LLMs exhibit good performance across both tasks, and self-correction instructions are particularly beneficial when the correct answer is already top-ranked; (ii) The morality levels in intermediate hidden states are strong indicators as to whether one instruction would be more effective than another; (iii) Based on our analysis of intermediate hidden states and task case studies of self-correction behaviors, we are first to propose the hypothesis that intrinsic moral self-correction is in fact superficial.

While task-agnostic debiasing provides notable generalizability and reduced reliance on downstream data, its impact on language modeling ability and the risk of relearning social biases from downstream task-specific data remain as the two most significant challenges when debiasing Pretrained Language Models (PLMs). The impact on language modeling ability can be alleviated given a high-quality and long-contextualized debiasing corpus, but there remains a deficiency in understanding the specifics of relearning biases. We empirically ascertain that the effectiveness of task-agnostic debiasing hinges on the quantitative bias level of both the task-specific data used for downstream applications and the debiased model. We empirically show that the lower bound of the bias level of the downstream fine-tuned model can be approximated by the bias level of the debiased model, in most practical cases. To gain more in-depth understanding about how the parameters of PLMs change during fine-tuning due to the forgetting issue of PLMs, we propose a novel framework which can Propagate Socially-fair Debiasing to Downstream Fine-tuning, ProSocialTuning. Our proposed framework can push the fine-tuned model to approach the bias lower bound during downstream fine-tuning, indicating that the ineffectiveness of debiasing can be alleviated by overcoming the forgetting issue through regularizing successfully debiased attention heads based on the PLMs’ bias levels from stages of pretraining and debiasing.

Natural Language Processing (NLP) models tend to inherit and amplify stereotypical biases present in their training data, leading to harmful societal consequences. Current efforts to rectify these biases typically revolve around making models oblivious to bias, which is at odds with the idea that humans require increased awareness to tackle these biases better. This prompts a fundamental research question: are bias-oblivious models the only viable solution to combat stereotypical biases? This paper answers this question by proposing the Agency-BeLiefs Embedding (ABLE) model, a novel approach that actively encodes stereotypical biases into the embedding space. ABLE draws upon social psychological theory to acquire and represent stereotypical biases in the form of agency and belief scores rather than directly representing stereotyped groups. Our experimental results showcase ABLE’s effectiveness in learning agency and belief stereotypes while preserving the language model’s proficiency. Furthermore, we underscore the practical significance of incorporating stereotypes within the ABLE model by demonstrating its utility in various downstream tasks. Our approach exemplifies the potential benefits of addressing bias through awareness, as opposed to the prevailing approach of mitigating bias through obliviousness.

Biases cause discrepancies in healthcare services. Race, gender, and age of a patient affect interactions with physicians and the medical treatments one receives. These biases in clinical practices can be amplified following the release of pre-trained language models trained on biomedical corpora. To bring awareness to such repercussions, we examine social biases present in the biomedical masked language models. We curate prompts based on evidence-based practice and compare generated diagnoses based on biases. For a case study, we measure bias in diagnosing coronary artery disease and using cardiovascular procedures based on bias. Our study demonstrates that biomedical models are less biased than BERT in gender, while the opposite is true for race and age.

Fine-tuning pretrained language models (PLMs) for downstream tasks is a large-scale optimization problem, in which the choice of the training algorithm critically determines how well the trained model can generalize to unseen test data, especially in the context of few-shot learning. To achieve good generalization performance and avoid overfitting, techniques such as data augmentation and pruning are often applied. However, adding these regularizations necessitates heavy tuning of the hyperparameters of optimization algorithms, such as the popular Adam optimizer. In this paper, we propose a two-stage fine-tuning method, PAC-tuning, to address this optimization challenge. First, based on PAC-Bayes training, PAC-tuning directly minimizes the PAC-Bayes generalization bound to learn proper parameter distribution. Second, PAC-tuning modifies the gradient by injecting noise with the variance learned in the first stage into the model parameters during training, resulting in a variant of perturbed gradient descent (PGD). In the past, the few-shot scenario posed difficulties for PAC-Bayes training because the PAC-Bayes bound, when applied to large models with limited training data, might not be stringent. Our experimental results across 5 GLUE benchmark tasks demonstrate that PAC-tuning successfully handles the challenges of fine-tuning tasks and outperforms strong baseline methods by a visible margin, further confirming the potential to apply PAC training for any other settings where the Adam optimizer is currently used for training.

Framing is a political strategy in which journalists and politicians emphasize certain aspects of a societal issue in order to influence and sway public opinion. Frameworks for detecting framing in news articles or social media posts are critical in understanding the spread of biased information in our society. In this paper, we propose CLoSE, a multi-task BERT-based model which uses contrastive learning to embed indicators of frames from news articles in order to predict political bias. We evaluate the performance of our proposed model on subframes and political bias classification tasks. We also demonstrate the model’s classification accuracy on zero-shot and few-shot learning tasks, providing a promising avenue for framing detection in unlabeled data.

With 56 million people actively trading and investing in cryptocurrency online and globally in 2020, there is an increasing need for automatic social media analysis tools to help understand trading discourse and behavior. In this work, we present a dual natural language modeling pipeline which leverages language and social network behaviors for the prediction of cryptocurrency day trading actions and their associated framing patterns. This pipeline first predicts if tweets can be used to guide day trading behavior, specifically if a cryptocurrency investor should buy, sell, or hold their cryptocurrencies in order to make a profit. Next, tweets are input to an unsupervised deep clustering approach to automatically detect trading framing patterns. Our contributions include the modeling pipeline for this novel task, a new Cryptocurrency Tweets Dataset compiled from influential accounts, and a Historical Price Dataset. Our experiments show that our approach achieves an 88.78% accuracy for day trading behavior prediction and reveals framing fluctuations prior to and during the COVID-19 pandemic that could be used to guide investment actions.

This abstract presents preliminary work in the application of natural language processing techniques and social network modeling for the prediction of cryptocurrency trading and investment behavior. Specifically, we are building models to use language and social network behaviors to predict if the tweets of a 24-hour period can be used to buy or sell cryptocurrency to make a profit. In this paper we present our novel task and initial language modeling studies.

Political discourse on social media microblogs, specifically Twitter, has become an undeniable part of mainstream U.S. politics. Given the length constraint of tweets, politicians must carefully word their statements to ensure their message is understood by their intended audience. This constraint often eliminates the context of the tweet, making automatic analysis of social media political discourse a difficult task. To overcome this challenge, we propose simultaneous modeling of high-level abstractions of political language, such as political slogans and framing strategies, with abstractions of how politicians behave on Twitter. These behavioral abstractions can be further leveraged as forms of supervision in order to increase prediction accuracy, while reducing the burden of annotation. In this work, we use Probabilistic Soft Logic (PSL) to build relational models to capture the similarities in language and behavior that obfuscate political messages on Twitter. When combined, these descriptors reveal the moral foundations underlying the discourse of U.S. politicians online, across differing governing administrations, showing how party talking points remain cohesive or change over time.

Previous works in computer science, as well as political and social science, have shown correlation in text between political ideologies and the moral foundations expressed within that text. Additional work has shown that policy frames, which are used by politicians to bias the public towards their stance on an issue, are also correlated with political ideology. Based on these associations, this work takes a first step towards modeling both the language and how politicians frame issues on Twitter, in order to predict the moral foundations that are used by politicians to express their stances on issues. The contributions of this work includes a dataset annotated for the moral foundations, annotation guidelines, and probabilistic graphical models which show the usefulness of jointly modeling abstract political slogans, as opposed to the unigrams of previous works, with policy frames for the prediction of the morality underlying political tweets.

This paper describes our proposed solution for SemEval 2017 Task 1: Semantic Textual Similarity (Daniel Cer and Specia, 2017). The task aims at measuring the degree of equivalence between sentences given in English. Performance is evaluated by computing Pearson Correlation scores between the predicted scores and human judgements. Our proposed system consists of two subsystems and one regression model for predicting STS scores. The two subsystems are designed to learn Paraphrase and Event Embeddings that can take the consideration of paraphrasing characteristics and sentence structures into our system. The regression model associates these embeddings to make the final predictions. The experimental result shows that our system acquires 0.8 of Pearson Correlation Scores in this task.

Politicians carefully word their statements in order to influence how others view an issue, a political strategy called framing. Simultaneously, these frames may also reveal the beliefs or positions on an issue of the politician. Simple language features such as unigrams, bigrams, and trigrams are important indicators for identifying the general frame of a text, for both longer congressional speeches and shorter tweets of politicians. However, tweets may contain multiple unigrams across different frames which limits the effectiveness of this approach. In this paper, we present a joint model which uses both linguistic features of tweets and ideological phrase indicators extracted from a state-of-the-art embedding-based model to predict the general frame of political tweets.

Framing is a political strategy in which politicians carefully word their statements in order to control public perception of issues. Previous works exploring political framing typically analyze frame usage in longer texts, such as congressional speeches. We present a collection of weakly supervised models which harness collective classification to predict the frames used in political discourse on the microblogging platform, Twitter. Our global probabilistic models show that by combining both lexical features of tweets and network-based behavioral features of Twitter, we are able to increase the average, unsupervised F1 score by 21.52 points over a lexical baseline alone.

During the 2016 United States presidential election, politicians have increasingly used Twitter to express their beliefs, stances on current political issues, and reactions concerning national and international events. Given the limited length of tweets and the scrutiny politicians face for what they choose or neglect to say, they must craft and time their tweets carefully. The content and delivery of these tweets is therefore highly indicative of a politician’s stances. We present a weakly supervised method for extracting how issues are framed and temporal activity patterns on Twitter for popular politicians and issues of the 2016 election. These behavioral components are combined into a global model which collectively infers the most likely stance and agreement patterns among politicians, with respective accuracies of 86.44% and 84.6% on average.