Canada

Building robust deep neural networks (DNNs) against adversarial attacks is an important but challenging task. Previous defense approaches mainly focus on developing new model structures or training algorithms, but they do little to tap the potential of training instances, especially instances with robust patterns carring innate robustness. In this paper, we show that robust and non-robust instances in the training dataset, though are both important for test performance, have contrary impacts on robustness, which makes it possible to build a highly robust model by leveraging the training dataset in a more effective way. We propose a new method that can distinguish between robust instances from non-robust ones according to the model&#39;s sensitivity to perturbations on individual instances during training. Surprisingly, we find that the model under standard training easily overfits the robust instances by relying on their simple patterns before the model completely learns their robust features. Finally, we propose a new mitigation algorithm to further release the potential of robust instances. Experimental results show that proper use of robust instances in the original dataset is a new line to achieve highly robust models.

&lt;iframe src=&quot;https://app.sli.do/event/kgQkh7cPzt9VEM52HG3MJu/embed/polls/9147730a-59c1-4c07-a203-d21cf0a71912&quot; width=&quot;300&quot; height=&quot;400&quot;&gt;&lt;/iframe&gt;

ACL 2023

Characterizing the Impacts of Instances on Robustness

robust

overfitting

adversarial training

### Welcome to ACL 2023, the 61st Annual Meeting of the Association for Computational Linguistics! 
<br>The conference will be held in Toronto, Canada, July 9-14, 2023.<br><br>
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.<br>
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)**
<br><br><br><br>

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.

Building robust deep neural networks (DNNs) against adversarial attacks is an important but challenging task. Previous defense approaches mainly focus on developing new model structures or training algorithms, but they do little to tap the potential of training instances, especially instances with robust patterns carring innate robustness. In this paper, we show that robust and non-robust instances in the training dataset, though are both important for test performance, have contrary impacts on robustness, which makes it possible to build a highly robust model by leveraging the training dataset in a more effective way. We propose a new method that can distinguish between robust instances from non-robust ones according to the model's sensitivity to perturbations on individual instances during training. Surprisingly, we find that the model under standard training easily overfits the robust instances by relying on their simple patterns before the model completely learns their robust features. Finally, we propose a new mitigation algorithm to further release the potential of robust instances. Experimental results show that proper use of robust instances in the original dataset is a new line to achieve highly robust models.

<iframe src="https://app.sli.do/event/kgQkh7cPzt9VEM52HG3MJu/embed/polls/9147730a-59c1-4c07-a203-d21cf0a71912" width="300" height="400"></iframe>

poster

Re-rankers, which order retrieved documents with respect to the relevance score on the given query, have gained attention for the information retrieval (IR) task. Rather than fine-tuning the pre-trained language model (PLM), the large-scale language model (LLM) is utilized as a zero-shot re-ranker with excellent results. While LLM is highly dependent on the prompts, the impact and the optimization of the prompts for the zero-shot re-ranker are not explored yet. Along with highlighting the impact of optimization on the zero-shot re-ranker, we propose a novel discrete prompt optimization method, Constrained Prompt generation (Co-Prompt), with the metric estimating the optimum for re-ranking. Co-Prompt guides the generated texts from PLM toward optimal prompts based on the metric without parameter update. The experimental results demonstrate that Co-Prompt leads to outstanding re-ranking performance against the baselines. Also, Co-Prompt generates more interpretable prompts for humans against other prompt optimization methods.

<iframe src="https://app.sli.do/event/eX4PwSmcCc1EvCqjxcYdKB/embed/polls/4ee0d92f-1052-4fb0-8f41-f7754d7a1983" width="300" height="400"></iframe>

Discrete Prompt Optimization via Constrained Generation for Zero-shot Re-ranker

When scoring argumentative essays in an educational context, not only the presence or absence of certain argumentative elements but also their quality is important. On the recently published student essay dataset PERSUADE, we first show that the automatic scoring of argument quality benefits from additional information about context, writing prompt and argument type. We then explore the different combinations of three tasks: automated span detection, type and quality prediction. Results show that a multi-task learning approach combining the three tasks outperforms sequential approaches that first learn to segment and then predict the quality/type of a segment.

<iframe src="https://app.sli.do/event/eX4PwSmcCc1EvCqjxcYdKB/embed/polls/a16f0f07-ffaf-48a9-9479-0d9a514926f5" width="300" height="400"></iframe>

Score It All Together: A Multi-Task Learning Study on Automatic Scoring of Argumentative Essays

Knowledge-based referring expression comprehension (KB-REC) aims to identify visual objects referred to by expressions that incorporate knowledge. Existing methods employ sentence-level retrieval and fusion methods, which may lead to issues of similarity bias and interference from irrelevant information in unstructured knowledge sentences. To address these limitations, we propose a segment-level and category-oriented network (SLCO). Our approach includes a segment-level and prompt-based knowledge retrieval method to mitigate the similarity bias problem and a category-based grounding method to alleviate interference from irrelevant information in knowledge sentences. Experimental results show that our SLCO can eliminate interference and improve the overall performance of the KB-REC task.

Segment-Level and Category-Oriented Network for Knowledge-Based Referring Expression Comprehension

Sentiment analysis (SA) systems are used in many products and hundreds of languages. Gender and racial biases are well-studied in English SA systems, but understudied in other languages, with few resources for such studies. To remedy this, we build a counterfactual evaluation corpus for gender and racial/migrant bias in four languages. We demonstrate its usefulness by answering a simple but important question that an engineer might need to answer when deploying a system: What biases do systems import from pre-trained models when compared to a baseline with no pre-training? Our evaluation corpus, by virtue of being counterfactual, not only reveals which models have less bias, but also pinpoints changes in model bias behaviour, which enables more targeted mitigation strategies. We release our code and evaluation corpora to facilitate future research.

Bias Beyond English: Counterfactual Tests for Bias in Sentiment Analysis in Four Languages

Cosine similarity between two words, computed using their contextualised token embeddings obtained from masked language models (MLMs) such as BERT has shown to underestimate the actual similarity between those words~\cite{zhou-etal-2022-problems}.This similarity underestimation problem is particularly severe for high frequent words.Although this problem has been noted in prior work, no solution has been proposed thus far.We observe that the $\ell_2$ norm of contextualised embeddings of a word correlates with its log-frequency in the pretraining corpus.Consequently, the larger $\ell_2$ norms associated with the high frequent words reduce the cosine similarity values measured between them, thus underestimating the similarity scores.To solve this issue, we propose a method to \emph{discount} the $\ell_2$ norm of a contextualised word embedding by the frequency of that word in a corpus when measuring the cosine similarities between words.We show that the so called \emph{stop} words behave differently from the rest of the words, which require special consideration during their discounting process.Experimental results on a contextualised word similarity dataset show that our proposed discounting method accurately solves the similarity underestimation problem.\footnote{An anonymized version of the source code of our proposed method is submitted to the reviewing system.

<iframe src="https://app.sli.do/event/eX4PwSmcCc1EvCqjxcYdKB/embed/polls/df1cb0a4-133c-44b5-ace7-8e4f9b57eddf" width="300" height="400"></iframe>

Solving Cosine Similarity Underestimation between High Frequency Words by $\ell_2$ Norm Discounting

In recent years, joint Vision-Language (VL) models have increased in popularity and capability. Very few studies have attempted to investigate bias in VL models, even though it is a well-known issue in both individual modalities.This paper presents the first multi-dimensional analysis of bias in English VL models, focusing on gender, ethnicity, and age as dimensions.When subjects are input as images, pre-trained VL models complete a neutral template with a hurtful word 5% of the time, with higher percentages for female and young subjects.Bias presence in downstream models has been tested on Visual Question Answering. We developed a novel bias metric called the Vision-Language Association Test based on questions designed to elicit biased associations between stereotypical concepts and targets. Our findings demonstrate that pre-trained VL models contain biases that are perpetuated in downstream tasks.

<iframe src="https://app.sli.do/event/eX4PwSmcCc1EvCqjxcYdKB/embed/polls/29b078b6-72f5-4f3a-b735-200a929b5b73" width="300" height="400"></iframe>

A Multi-dimensional study on Bias in Vision-Language models

Membership Inference attacks (MIAs) aim to predict whether a data sample was present in the training data of a machine learning model or not, and are widely used for assessing the privacy risks of language models. Most existing attacks rely on the observation that models tend toassign higher probabilities to their training samples than non-training points. However, simple thresholding of the model score in isolation tends to lead to high false-positive rates as it does not account for the intrinsic complexity of a sample. Recent work has demonstrated that reference-based attacks which compare model scores to those obtained from a reference model trained on similar data can substantially improve the performance of MIAs.However, in order to train reference models, attacks of this kind make the strong and arguably unrealistic assumption that an adversary has access to samples closely resembling the original training data. Therefore, we investigate their performance in more realistic scenarios and find that they are highly fragile in relation to the data distribution used to train reference models. To investigate whether this fragility provides a layer of safety, we propose and evaluate neighbourhood attacks, which compare model scores for a given sample to scores of synthetically generated neighbour texts and therefore eliminate the need for access to the training data distribution. We show that, in addition to being competitive with reference-based attacks that have perfect knowledge about the training data distribution, our attack clearly outperforms existing reference-free attacks as well as reference-based attacks with imperfect knowledge, which demonstrates the need for a reevaluation of the threat model of adversarial attacks.

<iframe src="https://app.sli.do/event/t38CPUax2NoHBQcXN6okdA/embed/polls/dc326db2-98fe-4af7-a74d-2fc13ddd8bc1" width="300" height="400"></iframe>

Membership Inference Attacks against Language Models via Neighbourhood Comparison

A simile is a figure of speech that compares two different things (called the tenor and the vehicle) via shared properties. The tenor and the vehicle are usually connected with comparator words such as "like" or "as". The simile phenomena are unique and complex in a real-life dialogue scene where the tenor and the vehicle can be verbal phrases or sentences, mentioned by different speakers, exist in different sentences, or occur in reversed order. However, the current simile research usually focuses on similes in a triplet tuple (tenor, property, vehicle) or a single sentence where the tenor and vehicle are usually entities or noun phrases, which could not reflect complex simile phenomena in real scenarios. In this paper, we propose a novel and high-quality multilingual simile dialogue (MSD) dataset to facilitate the study of complex simile phenomena. The MSD is the largest manually annotated simile data ($\sim$21K) and it contains both English and Chinese data. Meanwhile, the MSD data can also be used on dialogue tasks to test the ability of dialogue systems when using similes. We design 3 simile tasks (recognition, interpretation, and generation) and 2 dialogue tasks (retrieval and generation) with MSD. For each task, we provide experimental results from strong pre-trained or state-of-the-art models. The experiments demonstrate the challenge of MSD and we will release the data/code on GitHub.

<iframe src="https://app.sli.do/event/t38CPUax2NoHBQcXN6okdA/embed/polls/1224191d-85d3-485e-adfe-570618e27127" width="300" height="400"></iframe>

I run as fast as a rabbit, can you? A Multilingual Simile Dialogues Datasets

Despite recent advancements, NLP models continue to be vulnerable to bias. This bias often originates from the uneven distribution of real-world data and can propagate through the annotation process. Escalated integration of these models in our lives calls for methods to mitigate bias without overbearing annotation costs. While active learning (AL) has shown promise in training models with a small amount of annotated data, AL's reliance on the model's behavior for selective sampling can lead to an accumulation of unwanted bias rather than bias mitigation. However, infusing clustering with AL can overcome the bias issue of both AL and traditional annotation methods while exploiting AL's annotation efficiency. In this paper, we propose a novel adaptive clustering-based active learning algorithm, D-CALM, that dynamically adjusts clustering and annotation efforts in response to an estimated classifier error-rate. Experiments on eight datasets for a diverse set of text classification tasks, including emotion, hatespeech, dialog act, and book type detection, demonstrate that our proposed algorithm significantly outperforms baseline AL approaches with both pretrained transformers and traditional Support Vector Machines. D-CALM showcases robustness against different measures of information gain and, as evident from our analysis of label and error distribution, can significantly reduce unwanted model bias.

<iframe src="https://app.sli.do/event/t38CPUax2NoHBQcXN6okdA/embed/polls/661dcd1f-1ab3-441b-a68a-5d0c4cb7bb03" width="300" height="400"></iframe>

D-CALM: A Dynamic Clustering-based Active Learning Approach for Mitigating Bias

Quality Estimation (QE) is the task of evaluating the quality of a translation when reference translation is not available. The goal of QE aligns with the task of corpus filtering, where we assign the quality score to the sentence pairs present in the pseudo-parallel corpus. We propose a Quality Estimation based Filtering approach to extract high-quality parallel data from the pseudo-parallel corpus. To the best of our knowledge, this is a novel adaptation of QE framework to extracting quality parallel corpus from the pseudo-parallel corpus.. By training with this filtered corpus, we observe an improvement in the Machine Translation (MT) system's performance by up to 1.8 BLEU points, for English-Marathi, Chinese-English, and Hindi-Bengali language pairs, over the baseline model. The baseline model is the one that is trained on the whole pseudo-parallel corpus. Our Few-shot QE model transfer learned from the English-Marathi QE model and fine-tuned on only 500 Hindi-Bengali training instances, shows an improvement of up to 0.6 BLEU points for Hindi-Bengali language pair, compared to the baseline model. This demonstrates the promise of transfer learning in the setting under discussion. QE systems typically require in the order of (7K-25K) of training data. Our Hindi-Bengali QE is trained on only 500 instances of training that is 1/40th of the normal requirement and achieves comparable performance. All the scripts and datasets utilized in this study will be publicly available.

<iframe src="https://app.sli.do/event/t38CPUax2NoHBQcXN6okdA/embed/polls/652f0797-f04b-47eb-bd71-e61c1d154b82" width="300" height="400"></iframe>

Downloads

Next from ACL 2023

Discrete Prompt Optimization via Constrained Generation for Zero-shot Re-ranker

Similar lecture

Exploring Robust Overfitting for Pre-trained Language Models