Location

South Ballroom, Hemmingson Center

Start Date

3-4-2025 2:15 PM

End Date

3-4-2025 3:00 PM

Description

Chaired by Gina Sprint, Ph.D. (Gonzaga University)

Large Language Models (LLMs) can impressively generate human-like text, translate languages, write different kinds of creative content, and answer your questions in an informative way. They are rapidly becoming integrated into every facet of our digital lives, powering our search engines, social media platforms, and even customer service interactions. But amidst the excitement and hype surrounding LLMs, we must pause and ask a critical question: do they fairly represent the diversity of human language, culture, and experiences?

This talk focuses on the intricate landscape of building truly inclusive LLMs. It’s not just about removing harmful stereotypes or ensuring equal representation across demographic categories; it’s about creating language technology that genuinely understands and respects the multifaceted nature of human communication. One of the key challenges in building inclusive LLMs lies in addressing the inherent biases that can creep into these systems. LLMs are trained on massive datasets of text and code, which often reflect the biases and prejudices prevalent in society. As a result, LLMs can inadvertently perpetuate and even amplify these biases, leading to unfair or discriminatory outcomes. Tackling this challenge requires a multi-pronged approach, encompassing technical innovations, ethical considerations, and a commitment to engaging with diverse communities and stakeholders.

Our journey towards inclusive LLMs begins with a critical examination of social media, a vibrant and dynamic space where a kaleidoscope of voices converge. Social media platforms have become fertile ground for the expression of diverse perspectives, representing a multitude of languages, dialects, and cultural backgrounds. However, this abundance of information can also be overwhelming, making it difficult to extract meaningful insights and understand the nuances of different communities. Automatic text summarization offers a potential solution, enabling us to condense large volumes of social media content while preserving salient information. But how can we ensure that these summaries truly capture the diversity of voices and perspectives present in the data? To address this question, we introduce DivSumm, a carefully crafted dataset comprising dialect-diverse tweets from African American English, Hispanic-aligned Language, and White-aligned Language communities paired with corresponding human-written summaries. An empirical analysis of these reference summaries reveals that human annotators are able to construct well-balanced and inclusive representations, whereas system-generated summaries tend to exhibit dialect bias, disproportionately amplifying certain linguistic patterns while diminishing others. A promising technique for mitigating this issue is cluster-based preprocessing, which enhances dialectal fairness without adversely impacting summarization quality, thereby presenting a scalable intervention for fairness-aware summarization.

Our exploration of dialect diversity leads us further, questioning whether simply including a variety of dialects is sufficient to ensure fair representation. We find the subtle yet pervasive influence of position bias, a phenomenon where the order in which input documents are presented to the LLM can significantly impact the fairness of summarization outputs. Even when the textual quality of the summary remains consistent, position bias can lead to disproportionate representation or omission of certain groups, depending on their position in the input sequence. This finding motivates the development of innovative methods that can effectively mitigate the influence of position bias, ensuring that all voices are heard and valued, regardless of their position in the input.

Recognizing the challenges of achieving fair and balanced representation in LLM-generated summaries, we introduce two novel methods specifically designed for fair summarization: FairExtract and FairGPT. FairExtract employs a clustering-based approach, grouping similar tweets together before extracting sentences for the summary. This method ensures that diverse perspectives are included by selecting sentences from different clusters, promoting a more balanced representation of the input data. FairGPT, on the other hand, leverages the power of LLM with carefully designed fairness constraints. Through rigorous evaluation on the DivSumm dataset, we demonstrate that both FairExtract and FairGPT achieve superior fairness compared to existing approaches while maintaining competitive summarization quality.

While fairness in text summarization is a critical concern, bias in LLM-generated narratives presents broader challenges. To investigate this, we analyse gender and ethnicity representation in AI-generated occupational narratives across 25 professional fields. Using Llama, Claude, and GPT-4.0, we evaluate how LLMs associate demographic groups with specific professions, revealing systematic disparities in representation. For instance, LLMs may disproportionately associate women with caregiving roles while attributing technical and leadership positions to men, thus perpetuating existing biases. To mitigate these biases, we propose explainability-driven fairness constraints in LLMs, aligning model transparency with demographic fairness. By integrating feedback mechanisms that adjust model outputs, we reduce demographic disparities in occupational narratives by 2%–20% across different job categories.

As AI systems increasingly mediate human interactions, professional opportunities, and knowledge dissemination, fairness in NLP applications is no longer an abstract concern – it is an ethical imperative. Throughout this journey, we have encountered complex challenges and difficult questions. How can we ensure that AI respects linguistic and cultural diversity? How do we uncover and address hidden biases in LLMs? How do we balance powerful AI capabilities with ethical responsibility? These are questions that demand a collective effort. Researchers, developers, policymakers, and everyday users must collaborate to build AI technologies that truly benefit everyone.

In conclusion, building truly inclusive AI is not just a technical challenge but a social and ethical imperative, requiring a concerted effort from all stakeholders. By addressing biases in training data, developing fairness-aware algorithms, and engaging in open and transparent dialogue about the ethical implications of LLMs, we can create language technology that empowers and benefits everyone.

Download

Share

COinS
 
Apr 3rd, 2:15 PM Apr 3rd, 3:00 PM

Building Fair and Inclusive AI for a Diverse World

South Ballroom, Hemmingson Center

Chaired by Gina Sprint, Ph.D. (Gonzaga University)

Large Language Models (LLMs) can impressively generate human-like text, translate languages, write different kinds of creative content, and answer your questions in an informative way. They are rapidly becoming integrated into every facet of our digital lives, powering our search engines, social media platforms, and even customer service interactions. But amidst the excitement and hype surrounding LLMs, we must pause and ask a critical question: do they fairly represent the diversity of human language, culture, and experiences?

This talk focuses on the intricate landscape of building truly inclusive LLMs. It’s not just about removing harmful stereotypes or ensuring equal representation across demographic categories; it’s about creating language technology that genuinely understands and respects the multifaceted nature of human communication. One of the key challenges in building inclusive LLMs lies in addressing the inherent biases that can creep into these systems. LLMs are trained on massive datasets of text and code, which often reflect the biases and prejudices prevalent in society. As a result, LLMs can inadvertently perpetuate and even amplify these biases, leading to unfair or discriminatory outcomes. Tackling this challenge requires a multi-pronged approach, encompassing technical innovations, ethical considerations, and a commitment to engaging with diverse communities and stakeholders.

Our journey towards inclusive LLMs begins with a critical examination of social media, a vibrant and dynamic space where a kaleidoscope of voices converge. Social media platforms have become fertile ground for the expression of diverse perspectives, representing a multitude of languages, dialects, and cultural backgrounds. However, this abundance of information can also be overwhelming, making it difficult to extract meaningful insights and understand the nuances of different communities. Automatic text summarization offers a potential solution, enabling us to condense large volumes of social media content while preserving salient information. But how can we ensure that these summaries truly capture the diversity of voices and perspectives present in the data? To address this question, we introduce DivSumm, a carefully crafted dataset comprising dialect-diverse tweets from African American English, Hispanic-aligned Language, and White-aligned Language communities paired with corresponding human-written summaries. An empirical analysis of these reference summaries reveals that human annotators are able to construct well-balanced and inclusive representations, whereas system-generated summaries tend to exhibit dialect bias, disproportionately amplifying certain linguistic patterns while diminishing others. A promising technique for mitigating this issue is cluster-based preprocessing, which enhances dialectal fairness without adversely impacting summarization quality, thereby presenting a scalable intervention for fairness-aware summarization.

Our exploration of dialect diversity leads us further, questioning whether simply including a variety of dialects is sufficient to ensure fair representation. We find the subtle yet pervasive influence of position bias, a phenomenon where the order in which input documents are presented to the LLM can significantly impact the fairness of summarization outputs. Even when the textual quality of the summary remains consistent, position bias can lead to disproportionate representation or omission of certain groups, depending on their position in the input sequence. This finding motivates the development of innovative methods that can effectively mitigate the influence of position bias, ensuring that all voices are heard and valued, regardless of their position in the input.

Recognizing the challenges of achieving fair and balanced representation in LLM-generated summaries, we introduce two novel methods specifically designed for fair summarization: FairExtract and FairGPT. FairExtract employs a clustering-based approach, grouping similar tweets together before extracting sentences for the summary. This method ensures that diverse perspectives are included by selecting sentences from different clusters, promoting a more balanced representation of the input data. FairGPT, on the other hand, leverages the power of LLM with carefully designed fairness constraints. Through rigorous evaluation on the DivSumm dataset, we demonstrate that both FairExtract and FairGPT achieve superior fairness compared to existing approaches while maintaining competitive summarization quality.

While fairness in text summarization is a critical concern, bias in LLM-generated narratives presents broader challenges. To investigate this, we analyse gender and ethnicity representation in AI-generated occupational narratives across 25 professional fields. Using Llama, Claude, and GPT-4.0, we evaluate how LLMs associate demographic groups with specific professions, revealing systematic disparities in representation. For instance, LLMs may disproportionately associate women with caregiving roles while attributing technical and leadership positions to men, thus perpetuating existing biases. To mitigate these biases, we propose explainability-driven fairness constraints in LLMs, aligning model transparency with demographic fairness. By integrating feedback mechanisms that adjust model outputs, we reduce demographic disparities in occupational narratives by 2%–20% across different job categories.

As AI systems increasingly mediate human interactions, professional opportunities, and knowledge dissemination, fairness in NLP applications is no longer an abstract concern – it is an ethical imperative. Throughout this journey, we have encountered complex challenges and difficult questions. How can we ensure that AI respects linguistic and cultural diversity? How do we uncover and address hidden biases in LLMs? How do we balance powerful AI capabilities with ethical responsibility? These are questions that demand a collective effort. Researchers, developers, policymakers, and everyday users must collaborate to build AI technologies that truly benefit everyone.

In conclusion, building truly inclusive AI is not just a technical challenge but a social and ethical imperative, requiring a concerted effort from all stakeholders. By addressing biases in training data, developing fairness-aware algorithms, and engaging in open and transparent dialogue about the ethical implications of LLMs, we can create language technology that empowers and benefits everyone.