Canada

Large-scale Pre-Trained Language Models (PTLMs) capture knowledge from massive human-written data which contains latent societal biases and toxic contents. In this paper, we leverage the primary task of PTLMs, i.e., language modeling, and propose a new metric to quantify manifested implicit representational harms in PTLMs towards 13 marginalized demographics. Using this metric, we conducted an empirical analysis of 24 widely used PTLMs. Our analysis provides insights into the correlation between the proposed metric in this work and other related metrics for representational harm. We observe that our metric correlates with most of the gender-specific metrics in the literature. Through extensive experiments, we explore the connections between PTLMs architectures and representational harms across two dimensions: depth and width of the networks. We found that prioritizing depth over width, mitigates representational harms in some PTLMs. Our code and data can be found at https://github.com/microsoft/SafeNLP.

ACL 2023

An Empirical Study of Metrics to Measure Representational Harms in Pre-Trained Language Models

responsible nlp

llms

workshop paper

### Welcome to ACL 2023, the 61st Annual Meeting of the Association for Computational Linguistics! 
 The conference will be held in Toronto, Canada, July 9-14, 2023. 
Following the succession of the recent conferences in our field, ACL 2023 will adopt a hybrid format.
While the impact of Covid has considerably diminished in terms of traveling, obtaining visas to Canada
entails a very long process. Moreover, the global economic conditions pose challenges for many individuals to travel to conferences. Recognizing these circumstances, we know many participants may not be
able to attend the conference in person. Therefore, we are committed to providing a great virtual platform
so everyone has the opportunity to interact with other participants and enjoy the conference. Based on the
current registered participants, approxiately 30% have chosen to attend the conference virtually. Whether
you join us in person or virtually, we sincerely hope everyone has a remarkable conference experience. 
This General Chair’s message is where I express my gratitude to the many individuals who have made
enormous contributions to the conference over the past year.

Read [**ACL 2023 General Chair's message**](https://docs.google.com/document/d/1WobYM7norbG4dI48s75HfJoD89qgX5a_F-6U8AteLSA/edit?usp=sharing/) in full.

##### **[Conference Handbook](https://2023.aclweb.org/downloads/acl2023-handbook.pdf)**

The Association for Computational Linguistics (ACL) is the premier international scientific and professional society for people working on computational problems involving human language, a field often referred to as either computational linguistics or natural language processing.

We investigate the phenomenon of an LLM's untruthful response using a large set of 220 handcrafted linguistic features. We focus on GPT-3 models and find that the linguistic profiles of responses are similar across model sizes. That is, how varying-sized LLMs respond to given prompts stays similar on the linguistic properties level. We expand upon this finding by training support vector machines that rely only upon the stylistic components of model responses to classify the truthfulness of statements. Though the dataset size limits our current findings, we show the possibility that truthfulness detection is possible without evaluating the content itself. But at the same time, the limited scope of our experiments must be taken into account in interpreting the results.

Linguistic Properties of Truthful Response

Despite the remarkable performances in various applications, machine learning (ML) models could potentially discriminate. They may result in biasness in decision-making, leading to an impact negatively on individuals and society. Recently, various methods have been developed to mitigate biasness and achieve significant performance. Attention mechanisms are a fundamental component of many state-of-the-art ML models and may potentially impact the fairness of ML models. However, how they explicitly influence fairness has yet to be thoroughly explored. In this paper, we investigate how different attention mechanisms affect the fairness of ML models, focusing on models used in Natural Language Processing (NLP) models. We evaluate the performance of fairness of several models with and without different attention mechanisms on widely used benchmark datasets. Our results indicate that the majority of attention mechanisms that have been assessed can improve the fairness performance of Bidirectional Gated Recurrent Unit (BiGRU) and Bidirectional Long Short Term Memory (BiLSTM) in all three datasets regarding religious and gender-sensitive groups, however, with varying degrees of trade-offs in accuracy measures. Our findings highlight the possibility of fairness being affected by adopting specific attention mechanisms in machine learning models for certain datasets.

Debunking Biases in Attention

By accounting for the part-of-speech of words during text-to-text privatization, we find a significant performance boost for downstream tasks of language models.

Guiding Text-to-Text Privatization by Syntax

This paper presents novel experiments shedding light on the shortcomings of current metrics for assessing biases of gender discrimination made by machine learning algorithms on textual data. We focus on the Bios dataset, and our learning task is to predict the occupation of individuals, based on their biography. Such prediction tasks are common in commercial Natural Language Processing (NLP) applications such as automatic job recommendations. We address an important limitation of theoretical discussions dealing with group-wise fairness metrics: they focus on large datasets, although the norm in many industrial NLP applications is to use small to reasonably large linguistic datasets for which the main practical constraint is to get a good prediction accuracy. We then question how reliable are different popular measures of bias when the size of the training set is simply sufficient to learn reasonably accurate predictions. Our experiments sample the Bios dataset and learn more than 200 models on different sample sizes. This allows us to statistically study our results and to confirm that common gender bias indices provide diverging and sometimes unreliable results when applied to relatively small training and test samples. This highlights the crucial importance of variance calculations for providing sound results in this field.

Are fairness metric scores enough to assess discrimination biases in machine learning?

Wikipedia articles are a common source of training data for Natural Language Processing (NLP) research, especially as a source for corpora in languages other than English. However, research has shown that not all Wikipedia editions are produced organically by native speakers, and there are substantial levels of automation and translation activities in the Wikipedia project that could negatively impact the degree to which they truly represent the language and the culture of native speakers. To encourage transparency in the Wikipedia project, Wikimedia Foundation introduced the depth metric as an indication of the degree of collaboration or how frequently users edit a Wikipedia edition’s articles. While a promising start, this depth metric suffers from a few serious problems, like a lack of adequate handling of inflation of edits metric and a lack of full utilization of users-related metrics. In this paper, we propose the DEPTH+ metric, provide its mathematical definitions, and describe how it reflects a better representation of the depth of human collaborativeness. We also quantify the bot activities in Wikipedia and offer a bot-free depth metric after the removal of the bot-created articles and the bot-made edits on the Wikipedia articles.

DEPTH+: An Enhanced Depth Metric for Wikipedia Corpora Quality

As generative NLP can now produce content nearly indistinguishable from human writing, it becomes difficult to identify genuine research contributions in academic writing and scientific publications. Moreover, information in NLP-generated text can potentially be factually wrong or even entirely fabricated. This study introduces a novel benchmark dataset, containing human-written and machine-generated scientific papers from SCIgen, GPT-2, GPT-3, ChatGPT, and Galactica. After describing the generation and extraction pipelines, we also experiment with four distinct classifiers as a baseline for detecting the authorship of scientific text. A strong focus is put on generalization capabilities and explainability to highlight the strengths and weaknesses of detectors. We believe our work serves as an important step towards creating more robust methods for distinguishing between human-written and machine-generated scientific papers, ultimately ensuring the integrity of scientific literature.

Distinguishing Fact from Fiction: A Benchmark Dataset for Identifying Machine-Generated Scientific Papers in the LLM Era.

Textual counterfactual examples explain a prediction by modifying the tokens of an initial instance in order to flip the outcome of a classifier. Even under sparsity constraint, counterfactual generation can lead to numerous changes from the initial text, making the explanation hard to understand. We propose Counterfactual Feature Importance, a method to make non-sparse counterfactual explanations more intelligible. Counterfactual Feature Importance assesses token change importance between an instance to explain and its counterfactual example. We develop two ways of computing Counterfactual Feature Importance, respectively based on classifier gradient computation and counterfactual generator loss evolution during counterfactual search. Then we design a global version of Counterfactual Feature Importance, providing rich information about semantic fields globally impacting classifier predictions. Counterfactual Feature Importance enables to focus on impacting parts of counterfactual explanations, making counterfactual explanations involving numerous changes more understandable.

Enhancing textual counterfactual explanation intelligibility through Counterfactual Feature Importance

This work investigates the effectiveness of different pseudonymization techniques, ranging from rule-based substitutions to using pre-trained Large Language Models (LLMs), on a variety of datasets and models used for two widely used NLP tasks: text classification and summarization. Our work provides crucial insights into the gaps between original and anonymized data (focusing on the pseudonymization technique) and model quality and fosters future research into higher-quality anonymization techniques better to balance the trade-offs between data protection and utility preservation. We make our code, pseudonymized datasets, and downstream models publicly available.

Privacy- And Utility-Preserving NLP with Anonymized data: A case study of Pseudonymization

Though state-of-the-art (SOTA) NLP systems have achieved remarkable performance on a variety of language understanding tasks, they primarily focus on questions that have a correct and a definitive answer. However, in real-world applications, users often ask questions that don't have a definitive answer such as questions about future events, questions lacking necessary details to find the answer, and questions that are ambiguous. Incorrectly answering such questions certainly hampers a system's reliability and trustworthiness. Can SOTA models accurately identify such questions and provide a reasonable response?

To investigate the above question, we introduce QnotA, a dataset consisting of five different categories of questions that don't have definitive answers. Furthermore, for each QnotA instance, we also provide a corresponding 'QA' instance i.e. an alternate question that "can be" answered. With this data, we formulate three evaluation tasks that test a system's ability to 'identify', 'distinguish', and 'justify' QnotA questions. Through comprehensive experiments, we show that even SOTA models including GPT-3 and Flan T5 do not fare well on these tasks and lack considerably behind the human performance baseline. We conduct a thorough analysis which further leads to several interesting findings such as, despite not being able to accurately identify a QnotA question, GPT-3 on being prompted to output a justification of why the given QnotA question doesn't have a definitive answer is able to provide a reasonable justification. Finally, we believe our work and findings will encourage and facilitate development of more robust NLP systems that can also reasonably respond to questions that don't have a definitive answer.

Can NLP Models 'Identify', 'Distinguish', and 'Justify' Questions that Don't have a Definitive Answer?

This work analyzes backdoor watermarks in an autoregressive transformer fine-tuned to perform a generative sequence-to-sequence task, specifically summarization. We propose and demonstrate an attack to identify trigger words or phrases by analyzing open ended generations from autoregressive models that have backdoor watermarks inserted. It is shown in our work that triggers based on random common words are easier to identify than those based on single, rare tokens. The attack proposed is easy to implement and only requires access to the model weights. Code used to create the backdoor watermarked models and analyze their outputs is shared at \url{https://github.com/evan-person/findingBackdoorWatermarks}.

Premium content

An Empirical Study of Metrics to Measure Representational Harms in Pre-Trained Language Models

Downloads

Next from ACL 2023

Linguistic Properties of Truthful Response

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES