Malta

Red teaming is a common strategy for identifying weaknesses in generative language models (LMs) by producing adversarial prompts that trigger models to generate unsafe responses. Red teaming is instrumental for both model alignment and evaluation, but is labor-intensive and difficult to scale when done by humans. In this paper, we present Gradient-Based Red Teaming (GBRT), a novel red teaming method for automatically generating diverse prompts that are likely to cause an LM to output unsafe responses. GBRT is a form of prompt learning, trained by scoring an LM response with a safety classifier and then backpropagating through the frozen safety classifier and LM to update the prompt. To improve the coherence of input prompts, we introduce two variants that add a realism loss and fine-tune a pretrained model to generate the prompts instead of learning the prompts directly. Our experiments show that GBRT is more effective at finding prompts that trigger an LM to generate unsafe responses than a strong reinforcement learning-based red teaming approach and works even when the LM has been fine-tuned to produce safer outputs.

EACL 2024

Gradient-Based Language Model Red Teaming

red teaming

safety

language model

adversarial

technical paper

Welcome to the 18th Conference of the European Chapter of the Association for Computational Linguistics (EACL). Continuing its mission of expanding and involving the science community of all European countries, EACL has selected the Malta community for the 18th EACL. Considering the importance of physical interaction among researchers, the conference will be held at the Hotel Radisson Blu, St. Julians, in Malta, from 17 to 22 of March, 2024. The conference will also feature the possibility to participate virtually. As the flagship European conference in the field of computational linguistics, EACL welcomes European and international researchers covering a broad spectrum of research areas that are concerned with computational approaches to natural language. 
#### [Conference Handbook](https://drive.google.com/file/d/13Wn38q0yev6U3RTgnf4Tn5T9GuAuZ-xW/view?usp=sharing) *downloadable*

To access the EACL 2024 event page you need to register [**here**](https://acl.swoogo.com/EACL2024). 

Please register!

As the flagship European conference in the field of computational linguistics, EACL welcomes European and international researchers covering a broad spectrum of research areas that are concerned with computational approaches to natural language.

Abstract: The single shortest path algorithm is undefined for weighted finite-state automata over non-idempotent semirings because such semirings do not guarantee the existence of a shortest path. However, in non-idempotent semirings admitting an order satisfying a monotonicity condition (such as the plus-times or log semirings), the shortest string is well-defined. We describe an algorithm which finds the shortest string for a weighted non-deterministic automaton over such semirings using the backwards shortest distance of an equivalent deterministic automaton (DFA) as a heuristic for A* search performed over a companion idempotent semiring, which is proven to return the shortest string. There may be exponentially more states in the DFA, but the proposed algorithm needs to visit only a small fraction of them if determinization is performed "on the fly".

A* shortest string decoding for non-idempotent semirings

Recent advances in NLP suggest that some tasks, such as argument detection and relation classification, are better framed in a multimodal perspective. We propose multimodal argument mining for argumentative fallacy classification in political debates. To this end, we release the first corpus for multimodal fallacy classification. Our experiments show that the integration of the audio modality leads to superior classification performance. Our findings confirm that framing fallacy classification as a multimodal task is essential to capturing paralinguistic aspects of fallacious arguments.

Multimodal Fallacy Classification in Political Debates

Misinformation and disinformation phenomena existed long before the advent of digital technologies. The exponential use of social media platforms, whose information feeds have created the conditions for many to many communication and instant amplification of the news has accelerated the diffusion of inaccurate and misleading information. As a result, the identification of claims have emerged as a pivotal technology for combating the influence of misinformation and disinformation within news media. Most existing work has concentrated on claim analysis at the sentence level, neglecting the crucial exploration of supplementary attributes such as the claimer and the claim object of the claim or confining it by limiting its scope to a predefined list of topics. Furthermore, previous research has been mostly centered around political debates, Wikipedia articles, and COVID-19 related content. By leveraging the advanced capabilities of Large Language Models (LLMs) in Natural Language Understanding (NLU) and text generation, we propose a novel architecture utilizing LLMs finetuned with LoRA to transform the claim, claimer and claim object detection task into a Question Answering (QA) setting. We evaluate our approach in a dataset of 867 scientific news articles of 3 domains (Health, Climate Change, Nutrition) (HCN), which are human annotated with the major claim, the claimer and the object of the major claim. We also evaluate our proposed model in the benchmark dataset of NEWSCLAIMS. Experimental and qualitative results showcase the effectiveness of the proposed approach. We make our dataset publicly available to encourage further research.

Leveraging fine-tuned Large Language Models with LoRA for Effective Claim, Claimer, and Claim Object Detection

While demographic factors like age and gender change the way people talk, and in particular, the way people talk to machines, there is little investigation into how large pre-trained language models (LMs) can adapt to these changes. To remedy this gap, we consider how demographic factors in LM language skills can be measured to determine compatibility with a target demographic. We suggest clinical techniques from Speech Language Pathology, which has norms for acquisition of language skills in humans. We conduct evaluation with a domain expert (i.e., a clinically licensed speech language pathologist), and also propose automated techniques to complement clinical evaluation at scale. Empirically, we focus on age, finding LM capability varies widely depending on task: GPT-3.5 mimics the ability of humans ranging from age 6-15 at tasks requiring inference, and simultaneously, outperforms a typical 21 year old at memorization. GPT-3.5 also has trouble with social language use, exhibiting less than 50% of the tested pragmatic skills. Findings affirm the importance of considering demographic alignment and conversational goals when using LMs as public-facing tools. Code, data, and a package will be available.

HumBEL: A Human-in-the-Loop Approach for Evaluating Demographic Factors of Language Models in Human-Machine Conversations

In-context learning with Large Language Models (LLMs) has emerged as a promising avenue of research in Dialog State Tracking (DST). However, the best-performing in-context learning methods involve retrieving and adding similar examples to the prompt, requiring access to labeled training data. Procuring such training data for a wide range of domains and applications is time-consuming, expensive, and, at times, infeasible. While zero-shot learning requires no training data, it significantly lags behind the few-shot setup. Thus, `\textit{Can we efficiently generate synthetic data for any dialogue schema to enable few-shot prompting?}' Addressing this question, we propose \method, a data generation framework tailored for DST, utilizing LLMs. Our approach only requires the dialogue schema and a few hand-crafted dialogue templates to synthesize natural, coherent, and free-flowing dialogues with DST annotations. Few-shot learning using data from {\method} results in $4-5\%$ improvement in Joint Goal Accuracy over the zero-shot baseline on MultiWOZ 2.1 and 2.4. Remarkably, our few-shot learning approach recovers nearly $98\%$ of the performance compared to the few-shot setup using human-annotated training data\footnote{Our synthetic data and code can be accessed at \\ \url{ https://github.com/apple/ml-synthdst}.}.

SynthDST: Synthetic Data is All You Need for Few-Shot Dialog State Tracking

Adversarial attacks against Language models (LMs) are a significant concern. In particular, adversarial samples exploit the model's sensitivity to small input changes. While these changes appear insignificant on the semantics of the input sample, they result in significant decay in model performance. In this paper, we propose Targeted Paraphrasing via RL (TPRL), an approach to automatically learn a policy to generate challenging samples that improve the model’s performance. TPRL leverages FLAN-T5, a language model, as a generator and employs a self-learned policy using a proximal policy optimization to generate the adversarial examples automatically. TPRL's reward is based on the confusion induced in the classifier, preserving the original text meaning through a Mutual Implication score. We demonstrate \& evaluate TPRL's effectiveness in discovering natural adversarial attacks and improving model performance through extensive experiments on four diverse NLP classification tasks via Automatic \& Human evaluation. TPRL outperforms strong baselines, exhibits generalizability across classifiers and datasets, and combines the strengths of language modeling and reinforcement learning to generate diverse and influential adversarial examples.

Finding a Needle in the Adversarial Haystack: A Targeted Paraphrasing Approach For Uncovering Edge Cases with Minimal Distribution Distortion

The prevalence of information manipulation online has created a need for propaganda detection systems. Such systems have typically focused on the surface words, ignoring the linguistic structure. Here we aim to bridge this gap. In particular, we present the first attempt at using discourse analysis for the task. We consider both paragraph-level and token-level classification and we propose a discourse-aware Transformer architecture. Our experiments on English and Russian demonstrate sizeable performance gains compared to a number of baselines. Moreover, our ablation study emphasizes the importance of specific types of discourse features, and our in-depth analysis reveals a strong correlation between propaganda instances and discourse spans.

Unleashing the Power of Discourse-Enhanced Transformers for Propaganda Detection

Image captioning models are typically trained by treating all samples equally, neglecting to account for mismatched or otherwise difficult data points. In contrast, recent work has shown the effectiveness of training models by scheduling the data using curriculum learning strategies. This paper contributes to this direction by actively curating difficult samples in datasets without increasing the total number of samples. We explore the effect of using three data curation methods within the training process: complete removal of an sample, caption replacement, or image replacement via a text-to-image generation model. Experiments on the Flickr30K and COCO datasets with the BLIP and BEiT-3 models demonstrate that these curation methods do indeed yield improved image captioning models, underscoring their efficacy.

The Role of Data Curation in Image Captioning

We have deployed an LLM-based spoken dialogue system in a real hospital. The ARI social robot embodies our system, which patients and their companions can have multi-party conversations with together. In order to enable this multi-party ability, multimodality is critical. Our system, therefore, receives speech and video as input, and generates both speech and gestures (arm, head, and eye movements). In this paper, we describe our complex setting and the architecture of our dialogue system. Each component is detailed, and a video of the full system is available with the appropriate components highlighted in real-time. Our system decides when it should take its turn, generates human-like clarification requests when the patient pauses mid-utterance, answers in-domain questions (grounding to the in-prompt knowledge), and responds appropriately to out-of-domain requests (like generating jokes or quizzes). This latter feature is particularly remarkable as real patients often utter unexpected sentences that could not be handled previously.

Multi-party Multimodal Conversations Between Patients, Their Companions, and a Social Robot in a Hospital Memory Clinic

That the recent LLM breakthroughs are solely due to scaling is a myth. Many difficult research problems had to be solved to make models like GPT4 and Mixtral possible. One of those difficult research problems is data quality. Data quality is a great challenge for NLP researchers with many opportunities for innovation and impact on current generative AI developments. I will focus on two examples in my talk: quality data for training a highly multilingual language model and quality data for instruction tuning.

Premium content

Gradient-Based Language Model Red Teaming

Downloads

Next from EACL 2024

A* shortest string decoding for non-idempotent semirings

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES