Digital Simulations (DS) provide safe environments where users interact with an agent through conversational prompts, providing engaging learning experiences that can be used to train teacher candidates in realistic classroom scenarios. These simulations usually include open-ended questions, allowing teacher candidates to express their thoughts but complicating an automatic response analysis. To address this issue, we have evaluated Large Language Models (LLMs) to identify characteristics (user behaviors) in the responses of DS for teacher education. We evaluated the performance of DeBERTaV3 and Llama 3, combined with zero-shot, few-shot, and fine-tuning. Our experiments discovered a significant variation in the LLMs' performance depending on the characteristic to identify. Additionally, we noted that DeBERTaV3 significantly reduced its performance when it had to identify new characteristics. In contrast, Llama 3 performed better than DeBERTaV3 in detecting new characteristics and showing more stable performance. Therefore, in DS where teacher educators need to introduce new characteristics because they change depending on the simulation or the educational objectives, it is more recommended to use Llama 3. These results can guide other researchers in introducing LLMs to provide the highly demanded automatic evaluations in DS.

Offensive language detection is an important task for filtering out abusive expressions and improving online user experiences. However, malicious users often attempt to avoid filtering systems through the involvement of textual noises. In this paper, we propose these evasions as user-intended adversarial attacks that insert special symbols or leverage the distinctive features of the Korean language. Furthermore, we introduce simple yet effective pooling strategies in a layer-wise manner to defend against the proposed attacks, focusing on the preceding layers not just the last layer to capture both offensiveness and token embeddings. We demonstrate that these pooling strategies are more robust to performance degradation even when the attack rate is increased, without directly training of such patterns. Notably, we found that models pre-trained on clean texts could achieve a comparable performance in detecting attacked offensive language, to models pre-trained on noisy texts by employing these pooling strategies.

The choice of hyperparameters greatly impacts performance in natural language processing. Often, it is hard to tell if a method is better than another or just better tuned. Tuning curves fix this ambiguity by accounting for tuning effort. Specifically, they plot validation performance as a function of the number of hyperparameter choices tried so far. While several estimators exist for these curves, it is common to use point estimates, which we show fail silently and give contradictory results when given too little data. Beyond point estimates, confidence bands are necessary to rigorously establish the relationship between different approaches. We present the first method to construct valid confidence bands for tuning curves. The bands are exact, simultaneous, and distribution-free, thus they provide a robust basis for comparing methods. Empirical analysis shows that while bootstrap confidence bands, which serve as a baseline, fail to approximate their target confidence, ours achieve it exactly. We validate our design with ablations, analyze the effect of sample size, and provide guidance on comparing models with our method. To promote confident comparisons in future work, we release a library implementing the method at https://github.com/nalourie/opda .

Semantic textual similarity is the task of estimating the similarity between the meaning of two texts. In this paper, we fine-tune transformer architectures for semantic textual similarity on the Semantic Textual Similarity Benchmark by tuning the model partially and then end-to-end. We experiment with BERT, RoBERTa, and DeBERTaV3 cross-encoders by approaching the problem as a binary classification task or a regression task. We combine the outputs of the transformer models and use handmade features as inputs for boosting algorithms. Due to worse test set results coupled with improvements on the validation set, we experiment with different dataset splits to further investigate this occurrence. We also provide an error analysis, focused on the edges of the prediction range.

This paper describes our submission to Task 10 at SemEval 2023-Explainable Detection of Online Sexism (EDOS), divided into three subtasks. The recent rise in social media platforms has seen an increase in disproportionate levels of sexism experienced by women on social media platforms. This has made detecting and explaining online sexist content more important than ever to make social media safer and more accessible for women. Our approach consists of experimenting and finetuning BERT-based models and using a Majority Voting ensemble model that outperforms individual baseline model scores. Our system achieves a macro F1 score of 0.8392 for Task A, 0.6092 for Task B, and 0.4319 for Task C.

This paper presents a new pre-trained language model, DeBERTaV3, which improves the original DeBERTa model by replacing mask language modeling (MLM) with replaced token detection (RTD), a more sample-efficient pre-training task. Our analysis shows that vanilla embedding sharing in ELECTRA hurts training efficiency and model performance. This is because the training losses of the discriminator and the generator pull token embeddings in different directions, creating the "tug-of-war" dynamics. We thus propose a new gradient-disentangled embedding sharing method that avoids the tug-of-war dynamics, improving both training efficiency and the quality of the pre-trained model. We have pre-trained DeBERTaV3 using the same settings as DeBERTa to demonstrate its exceptional performance on a wide range of downstream natural language understanding (NLU) tasks. Taking the GLUE benchmark with eight tasks as an example, the DeBERTaV3 Large model achieves a 91.37% average score, which is 1.37% over DeBERTa and 1.91% over ELECTRA, setting a new state-of-the-art (SOTA) among the models with a similar structure. Furthermore, we have pre-trained a multi-lingual model mDeBERTa and observed a larger improvement over strong baselines compared to English models. For example, the mDeBERTa Base achieves a 79.8% zero-shot cross-lingual accuracy on XNLI and a 3.6% improvement over XLM-R Base, creating a new SOTA on this benchmark. We have made our pre-trained models and inference code publicly available at https://github.com/microsoft/DeBERTa.

Fact verification has attracted a lot of research attention recently, e.g., in journalism, marketing, and policymaking, as misinformation and disinformation online can sway one's opinion and affect one's actions. While fact-checking is a hard task in general, in many cases, false statements can be easily debunked based on analytics over tables with reliable information. Hence, table-based fact verification has recently emerged as an important and growing research area. Yet, progress has been limited due to the lack of datasets that can be used to pre-train language models (LMs) to be aware of common table operations, such as aggregating a column or comparing tuples. To bridge this gap, in this paper we introduce PASTA, a novel state-of-the-art framework for table-based fact verification via pre-training with synthesized sentence-table cloze questions. In particular, we design six types of common sentence-table cloze tasks, including Filter, Aggregation, Superlative, Comparative, Ordinal, and Unique, based on which we synthesize a large corpus consisting of 1.2 million sentence-table pairs from WikiTables. PASTA uses a recent pre-trained LM, DeBERTaV3, and further pretrains it on our corpus. Our experimental results show that PASTA achieves new state-of-the-art performance on two table-based fact verification benchmarks: TabFact and SEM-TAB-FACTS. In particular, on the complex set of TabFact, which contains multiple operations, PASTA largely outperforms the previous state of the art by 4.7 points (85.6% vs. 80.9%), and the gap between PASTA and human performance on the small TabFact test set is narrowed to just 1.5 points (90.6% vs. 92.1%).