Croatia

Recent work has focused on compressing pre-trained language models (PLMs) like BERT where the major focus has been to improve the in-distribution performance for downstream tasks. However, very few of these studies have analyzed the impact of compression on the generalizability and robustness of compressed models for out-of-distribution (OOD) data. Towards this end, we study two popular model compression techniques including knowledge distillation and pruning and show that the compressed models are significantly less robust than their PLM counterparts on OOD test sets although they obtain similar performance on in-distribution development sets for a task. Further analysis indicates that the compressed models overfit on the shortcut samples and generalize poorly on the hard ones. We further leverage this observation to develop a regularization strategy for robust model compression based on sample uncertainty.

EACL 2023

Robustness Challenges in Model Distillation and Pruning for Natural Language Understanding

llms

ood generalization

model compression

technical paper

### Welcome to the 17th 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 had selected the Ukrainian community for the 16th EACL, which took place online due to the COVID pandemic. Unfortunately, the ongoing war made the organisation in Kyiv impossible. Considering the importance of physical interaction among researchers, especially after the restrictions imposed by the COVID pandemic, in addition to an online mode, the conference will be held in Dubrovnik, Croatia, from 2 to 6 of May, 2023. The original aim of strengthening the connection with the Ukrainian community will not change as our program will feature a dedicated session and a workshop to highlight work on Ukrainian language technologies. 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.

You need to register for the conference in order to access this site. Please visit https://2023.eacl.org/registration for more information.

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.

Argument mining seeks to extract arguments and their structure from unstructured texts. Identifying relations between arguments (such as attack, support, and neutral) is a challenging task because two arguments may be related to each other via implicit inferences. 
This task often requires external commonsense knowledge to discover how one argument relates to another. State-of-the-art methods, however, rely on pre-defined knowledge graphs, and thus might not cover target argument pairs well. We introduce a new generative neuro-symbolic approach to finding inference chains that connect the argument pairs by making use of the Commonsense Transformer (COMET). We evaluate our approach on three datasets for both the two-label (attack/support) and three-label (attack/support/neutral) tasks. Our approach significantly outperforms the state-of-the-art, by 2-5\% in F1 score, on all three datasets.

Uncovering Implicit Inferences for Improved Relational Argument Mining

Recent breakthroughs in self-supervised training have led to a new class of pretrained vision–language models. While there have been investigations of bias in multimodal models, they have mostly focused on gender and racial bias, giving much less attention to other relevant groups, such as minorities with regard to religion, nationality, sexual orientation, or disabilities. This is mainly due to lack of suitable benchmarks for such groups. We seek to address this gap by providing a visual and textual bias benchmark called MMBias, consisting of around 3,800 images and phrases covering 14 population subgroups. We utilize this dataset to assess bias in several prominent self-supervised multimodal models, including CLIP, ALBEF, and ViLT. Our results show that these models demonstrate meaningful bias favoring certain groups. Finally, we introduce a debiasing method designed specifically for such large pretrained models that can be applied as a post-processing step to mitigate bias, while preserving the remaining accuracy of the model.

Multi-Modal Bias: Introducing a Framework for Stereotypical Bias Assessment beyond Gender and Race in Vision–Language Models

Many measures of societal bias in language models have been proposed in recent years. A popular approach is to use a set of word filling prompts to evaluate the behavior of the language models. In this work, we analyze the validity of two such measures -- StereoSet and CrowS-Pairs. We show that these measures produce unexpected and illogical results when appropriate control group samples are constructed. Based on this, we believe that they are problematic and using them in the future should be reconsidered. We propose a way forward with an improved testing protocol. Finally, we also introduce a new gender bias dataset for Slovak.

In-Depth Look at Word Filling Societal Bias Measures

While human evaluation remains best practice for accurately judging the faithfulness of automatically-generated summaries, few solutions exist to address the increased difficulty and workload when evaluating long-form summaries. Through a survey of 162 papers on long-form summarization, we first shed light on current human evaluation practices surrounding long-form summaries. We find that 73\% of these papers do not perform any human evaluation on model-generated summaries, while other works face new difficulties that manifest when dealing with long documents (e.g., low inter-annotator agreement). Motivated by our survey, we present LongEval, a set of guidelines for human evaluation of faithfulness in long-form summaries that addresses the following challenges: (1) How can we achieve high inter-annotator agreement on faithfulness scores? (2) How can we minimize annotator workload while maintaining accurate faithfulness scores? and (3) Do humans benefit from automated alignment between summary and source snippets? We deploy LongEval in annotation studies on two long-form summarization datasets in different domains (SQuALITY and PubMed), and we find that switching to a finer granularity of judgment (e.g., clause-level) reduces inter-annotator variance in faithfulness scores (e.g., std-dev from 18.5 to 6.8). We also show that scores from a partial annotation of fine-grained units highly correlates with scores from a full annotation workload (0.89 Kendall's tau using 50\% judgements). We release our human judgments, annotation templates, and software as a Python library for future research.

LongEval: Guidelines for Human Evaluation of Faithfulness in Long-form Summarization

Large language models (LLMs) are subject to sociocultural and other biases previously identified using intrinsic evaluations. However, when and how these intrinsic biases in pre-trained LM representations propagate to downstream, fine-tuned NLP tasks like summarization is not well understood. In this work, we investigate one type of bias—name-nationality bias—and trace it from the pre-training stage to a downstream summarization task across multiple summarization modeling choices. We show that these biases manifest themselves as hallucinations in summarization, leading to factually incorrect summaries. We also find that this propagation of biases is algorithm-dependent: more abstractive models allow biases to propagate more directly to downstream tasks as hallucinated facts. Building on these observations, we further analyze how changes to the adaptation method and fine-tuning data set affect name nationality biases and show that while they can reduce the overall rate of hallucinations, they do not change the types of biases that do appear.

When Do Pre-Training Biases Propagate to Downstream Tasks? A Case Study in Text Summarization

Timely generation of radiology reports and diagnoses is a challenge worldwide due to the enormous number of cases and shortage of radiology specialists. In this paper, we propose a Knowledge Graph Augmented Vision Language BART (KGVL-BART) model that takes as input two chest X-ray images- one frontal and the other lateral- along with tags which are diagnostic keywords, and outputs a report with the patient-specific findings. Our system development effort is divided into 3 stages: i) construction of the Chest X-ray KG (referred to as chestX-KG), ii) image feature extraction, and iii) training a KGVL-BART model using the visual, text, and KG data. The dataset we use is the well-known Indiana University Chest X-ray reports with the train, validation, and test split of 3025 instances, 300 instances, and 500 instances respectively. We construct a Chest X-Ray knowledge graph from these reports by extracting entity1-relation-entity2 triples; the triples get extracted by a rule-based tool of our own. Constructed KG is verified by two experienced radiologists (with experience of 30 years and 8 years, respectively). We demonstrate that our model- KGVL-BART- outperforms State-of-the-Art transformer-based models on standard NLG scoring metrics. We also include a qualitative evaluation of our system by experienced radiologist (with experience of 30 years) on the test data, which showed that 73\% of the reports generated were fully correct, only 5.5\% are completely wrong and 21.5\% have important missing details though overall correct. To the best of our knowledge, ours is the first system to make use of multi-modality and domain knowledge to generate X-ray reports automatically.

KGVL-BART: Knowledge Graph Augmented Visual Language BART for Radiology Report Generation

This work empirically investigates punctuation insertions as adversarial attacks on NLP systems. Data from experiments on three tasks, five datasets, and six models with four attacks show that punctuation insertions, when limited to a few symbols (apostrophes and hyphens), are a superior attack vector compared to character insertions due to 1) a lower after-attack accuracy ($A_{aft-atk}$) than alphabetical character insertions; 2) higher semantic similarity between the resulting and original texts; and 3) a resulting text that is easier and faster to read as assessed with the Test of Word Reading Efficiency (TOWRE)). The tests also indicate that 4) grammar checking does not mitigate punctuation insertions and 5) punctuation insertions outperform word-level attacks in settings with a limited number of word synonyms and queries to the victim's model. Our findings indicate that inserting a few punctuation types that result in easy-to-read samples is a general attack mechanism. In light of this threat, we assess the impact of punctuation insertions, potential mitigations, the mitigation's tradeoffs, punctuation insertion's worst-case scenarios and summarize our findings in a qualitative casual map, so that developers can design safer, more secure systems.

Using Punctuation as an Adversarial Attack on Deep Learning-Based NLP Systems: An Empirical Study

While multimodal sentiment analysis (MSA) has gained much attention over the last few years, the main focus of most work on MSA has been limited to constructing multimodal representations that capture interactions between different modalities in a single task. This was largely due to a lack of unimodal annotations in MSA benchmark datasets. However, training a model using only multimodal representations can lead to suboptimal performance due to insufficient learning of each uni-modal representation. In this work, to fully optimize learning representations from multimodal data, we propose SUGRM which jointly trains multimodal and unimodal tasks using recalibrated features. The features are recalibrated such that the model learns to weight the features differently based on the features of other modalities. Further, to leverage unimodal tasks, we auto-generate unimodal annotations via a unimodal label generation module (ULGM). The experiment results on two benchmark datasets demonstrate the efficacy of our framework.

Self-Supervised Unimodal Label Generation Strategy Using Recalibrated Modality Representations for Multimodal Sentiment Analysis

Searching troves of videos with textual descriptions is a core multimodal retrieval task. Owing to the lack of a purpose-built dataset for text-to-video retrieval, video captioning datasets have been re-purposed to evaluate models by (1) treating captions as positive matches to their respective videos and (2) assuming all other videos to be negatives. However, this methodology leads to a fundamental flaw during evaluation: since captions are marked as relevant only to their original video, many alternate videos also match the caption, which introduces false-negative caption-video pairs.
We show that when these false negatives are corrected, a recent state-of-the-art model gains 25% recall points---a difference that threatens the validity of the benchmark itself. To diagnose and mitigate this issue, we annotate and release 683K additional caption-video pairs. Using these, we recompute effectiveness scores for three models on two standard benchmarks (MSR-VTT and MSVD). We find that (1) the recomputed metrics are up to 25% recall points higher for the best models, (2) these benchmarks are nearing saturation for Recall@10, (3) caption length (generality) is related to the number of positives, and (4) annotation costs can be mitigated through sampling. We recommend retiring these benchmarks in their current form, and we make recommendations for future text-to-video retrieval benchmarks.

Fighting FIRe with FIRE: Assessing the Validity of Text-to-Video Retrieval Benchmarks

Numbers have unique characteristics to words. Teaching models to understand numbers in text is an open-ended research question. Instead of discussing the required calculation skills, this paper focuses on a more fundamental topic: understanding numerals. We point out that innumeracy—the inability to handle basic numeral concepts—exists in most pretrained language models (LMs), and we propose a method to solve this issue by exploring the notation of numbers. Further, we discuss whether changing notation and pre-finetuning along with the comparing-number task can improve performance in three benchmark datasets containing quantitative-related tasks. The results of this study indicate that input reframing and the proposed pre-finetuning task is useful for RoBERTa.

Downloads

Next from EACL 2023

Uncovering Implicit Inferences for Improved Relational Argument Mining