Thailand

Traditional reference-based metrics, such as BLEU and ROUGE, are less effective for assessing outputs from Large Language Models (LLMs) that produce highly creative or superior-quality text, or in situations where reference outputs are unavailable. While human evaluation remains an option, it is costly and difficult to scale. Recent work using LLMs as evaluators (LLM-as-a-judge) is promising, but trust and reliability remain a significant concern. Integrating human input is crucial to ensure criteria used to evaluate are aligned with the human&#39;s intent, and evaluations are robust and consistent. This paper presents a user study of a design exploration called EvaluLLM, that enables users to leverage LLMs as customizable judges, promoting human involvement to balance trust and cost-saving potential with caution. Through interviews with eight domain experts, we identified the need for assistance in developing effective evaluation criteria aligning the LLM-as-a-judge with practitioners&#39; preferences and expectations. We offer findings and design recommendations to optimize human-assisted LLM-as-judge systems.

ACL 2024

Human-Centered Design Recommendations for LLM-as-a-judge

llm-as-a-judge

human-centered design

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

Traditional reference-based metrics, such as BLEU and ROUGE, are less effective for assessing outputs from Large Language Models (LLMs) that produce highly creative or superior-quality text, or in situations where reference outputs are unavailable. While human evaluation remains an option, it is costly and difficult to scale. Recent work using LLMs as evaluators (LLM-as-a-judge) is promising, but trust and reliability remain a significant concern. Integrating human input is crucial to ensure criteria used to evaluate are aligned with the human's intent, and evaluations are robust and consistent. This paper presents a user study of a design exploration called EvaluLLM, that enables users to leverage LLMs as customizable judges, promoting human involvement to balance trust and cost-saving potential with caution. Through interviews with eight domain experts, we identified the need for assistance in developing effective evaluation criteria aligning the LLM-as-a-judge with practitioners' preferences and expectations. We offer findings and design recommendations to optimize human-assisted LLM-as-judge systems.

workshop paper

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

The field of natural language generation has witnessed significant advancements in recent years, including the development of controllable text generation techniques. However, controlling the attributes of the generated text remains a challenge, especially when aiming to avoid undesirable behavior such as toxicity. In this work, we introduce Detoxification Generator (DETOXIGEN), an inference-time algorithm that steers the generation away from unwanted styles. DETOXIGEN is an ensemble of a pre-trained language model (generator) and a detoxifier. The detoxifier is trained intentionally on the toxic data representative of the undesirable attribute, encouraging it to generate text in that style exclusively. During the actual generation, we use the trained detoxifier to produce undesirable tokens for the generator to contrast against at each decoding step. This approach directly informs the generator to avoid generating tokens that the detoxifier considers highly likely. We evaluate DETOXIGEN on the commonly used REALTOXICITYPROMPTS benchmark (Gehman et al., 2020) with various language models as generators. We find that it significantly outperforms previous approaches in detoxification metrics while not compromising on the generation quality. Moreover, the detoxifier is obtained by soft prompt-tuning using the same backbone language model as the generator. Hence, DETOXIGEN requires only a tiny amount of extra weights from the virtual tokens of the detoxifier to be loaded into GPU memory while decoding, making it a promising lightweight, practical, and parameter-efficient detoxification strategy.

Parameter-Efficient Detoxification with Contrastive Decoding

Motivational Interviewing is a counselling style that requires skillful usage of reflective listening and engaging in conversations about sensitive and personal subjects. In this paper, we investigate to what extent we can use generative large language models in motivational interviewing chatbots to generate precise and variable reflections on user responses. We conduct a two-step human evaluation where we first independently assess the generated reflections based on four criteria essential to health counseling; appropriateness, specificity, naturalness, and engagement. In the second step, we compare the overall quality of generated and human-authored reflections via a ranking evaluation. We use GPT-4, BLOOM, and FLAN-T5 models to generate motivational interviewing reflections, based on real conversational data collected via chatbots designed to provide support for smoking cessation and sexual health. We discover that GPT-4 can produce reflections of a quality comparable to human-authored reflections. Finally, we conclude that large language models have the potential to enhance and expand reflections in predetermined health counseling chatbots, but a comprehensive manual review is advised.

To What Extent Are Large Language Models Capable of Generating Substantial Reflections for Motivational Interviewing Counseling Chatbots? A Human Evaluation

As Large Language Models (LLMs) and multi-agent systems are increasingly employed in intricate social environments, the research on how human interact with LLMs/agents become promising. This study aims to explore the dynamics of Human-AI interaction through a case study of the Diplomacy game, a strategic board game renowned for its emphasis on negotiation and alliance formation. The research investigates how agents and human influence each other on decision-making, trust, and strategy within interactive settings, specifically focusing on whether agents can mimic human-like strategic behaviors including deception, and how this affects human player perceptions and actions. To this end, we propose to utilize LLMs to play the game with integrated natural language processing capabilities for negotiating and strategizing. The experimental design\footnote{As this extended abstract is a position paper, we haven't yet provide experimental results but some tentative design.} will involve mixed groups of human and AI players, allowing for the analysis of game dynamics across different team configurations.

Exploring Human-AI Interaction: A Case Study on the Diplomacy Game

Active formulation of subgoals in problem-solving is an effective learning strategy for programming learners, allowing the transfer of knowledge across similar problems. Although proper guidance and feedback for learners are crucial to correct mistakes and misconceptions during subgoal formulation, providing them at scale is challenging and costly. With recent advances in generative AI, we investigate the practicality of using generative AI as TAs in programming education by examining their effectiveness in a subgoal learning environment. We explore whether programming learners can distinguish AI TAs from humans and whether the subgoal learning workflow with AI TAs yields learning gains and assess their capability to assist in coding the subgoals into executable programs. Our study shows that learners can distinguish AI TAs from human TAs based on response length and accuracy. Learners show learning gains over sessions with AI TAs in formulating subgoals and can produce code solutions faster with comparable satisfaction scores with AI TAs as human TAs.

Learning from Teaching Assistants to Formulate Subgoals for Programming Tasks: Exploring the Potential for AI Teaching Assistants

Large Vision Language Models can be used to assist visually impaired individuals by describing images they capture in their daily lives. Current evaluation datasets may not reflect the diverse cultural user backgrounds nor the situational context of this use case. To address this problem, we create a survey to determine caption preferences and propose a culture-centric evaluation benchmark by filtering VizWiz, an existing dataset with images taken by people who are blind. We then evaluate different models and prompts, investigating their reliability as visual assistants. While the evaluation results for state-of-the-art models seem promising, we identified some weak spots such as hallucinations and problems with conventional evaluation metrics. Our survey, data, code, and model outputs will be publicly available.

Vision-Language Models under Cultural and Inclusive Considerations

Despite the advancement of automatic summarization methods based on pre-trained language models, evaluating their effectiveness remains a challenge, particularly in the absence of a medical document reference-free summary evaluation metric. This paper proposes a novel reference-free evaluation metric for medical document summaries by employing contrastive learning using medical text-tailored data augmentation techniques. Our research showcases the metric’s superior performance in assessing the quality of generated summaries without the need for comparison texts. Through extensive experimentation and analysis, this work makes significant strides in improving the reliability and usability of automatic medical document evaluation tools in medical document settings.

Reference-free Medical Multi-document Summary Evaluation Metric via Contrastive Learning

We present a framework to assess the sensitivity of Large Language Models (LLMs) to textually embedded social signals using an Appraisal Theory perspective. We report on an experiment that uses prompts encoding three dimensions of social signals: Affect, Judgment, and Appreciation. In response to the prompt, an LLM generates both an analysis (Insight) and a conversational Response, which are analyzed in terms of sensitivity to the signals. We quantitatively evaluate the output text through topical analysis of the Insight and predicted social intelligence scores of the Response in terms of empathy and emotional polarity. Key findings show that LLMs are more sensitive to positive signals. The personas impact Responses but not the Insight. We discuss how our framework can be extended to a broader set of social signals, personas, and scenarios to evaluate LLM behaviors under various conditions.

Evaluating Large Language Models on Social Signal Sensitivity: An Appraisal Theory Approach

As Large Language Models (LLMs) reach wider application, aligning them with human values like helpfulness and harmlessness becomes increasingly crucial. Although Reinforcement Learning with Human Feedback (RLHF) proves effective, it is complicated and highly resource intensive. As such, offline RLHF methods have emerged as simpler alternatives, directly optimizing LLMs with ranking losses on a fixed preference dataset. Despite their promise, these methods only capture the ordering relationship between responses, overlooking the crucial aspect of ``how much'' one is preferred over the other. To address this issue, we propose a simple yet effective solution based on reward difference prediction. Specifically, we introduce reward difference coefficients to reweigh sample pairs in offline RLHF methods. We then present a difference model that considers rich interactions between a pair of responses for predicting these difference coefficients. Experiments with Alpaca-7b and the HH dataset demonstrate the effectiveness of our method in both automatic metrics and human evaluation, highlighting its potential for aligning LLMs with human values.

Offline RLHF Methods Need More Accurate Supervision Signals

During conversations, people align to one another over time, by using similar words, concepts, and syntax. This helps form a shared understanding of the conversational content and is associated with increased engagement and satisfaction. It also affects conversation outcomes: e.g., when talking to language learners, an above normal level of linguistic alignment of parents or language teachers is correlated with faster language acquisition. These benefits make human-like alignment an important property of dialogue systems, which has often been overlooked by the NLP community. In order to fill this gap, we ask: (RQ1) Due to the importance for engagement and satisfaction, to what degree do state-of-the-art dialogue systems align to adult users? (RQ2) With a potential application to child language acquisition in mind, do systems, similar to parents, show high levels of alignment during conversations with children? Our experiments show that ChatGPT aligns to adults at roughly human levels, while Llama2 shows elevated alignment. However, when responding to a child, both systems' alignment is below human levels.

Aligning to Adults Is Easy, Aligning to Children Is Hard: A Study of Linguistic Alignment in Dialogue Systems

Generative models are increasingly being used in various applications, such as text generation, commonsense reasoning, and question-answering. To be effective globally, these models must be aware of and account for local socio-cultural contexts, making it necessary to have benchmarks to evaluate the models for their cultural familiarity. Since the training data for LLMs is web-based and the Web is limited in its representation of information, it does not capture knowledge present within communities that are not on the Web. Thus, these models exacerbate the inequities, semantic misalignment, and stereotypes from the Web. There has been a growing call for community-centered participatory research methods in NLP. In this work, we respond to this call by using participatory research methods to introduce DOSA, the first community-generated Dataset of 615 Social Artifacts, by engaging with 260 participants from 19 different Indian geographic subcultures. We use a gamified framework that relies on collective sensemaking to collect the names and descriptions of these artifacts such that the descriptions semantically align with the shared sensibilities of the individuals from those cultures. Next, we benchmark four popular LLMs and find that they show significant variation across regional sub-cultures in their ability to infer the artifacts.

Downloads

Next from ACL 2024

Parameter-Efficient Detoxification with Contrastive Decoding

Similar lecture

From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2024

Parameter-Efficient Detoxification with Contrastive Decoding

Similar lecture

From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge

Downloads