Mohammad Taher Pilehvar

Ali Modarressi, Mohsen Fayyaz, Yadollah Yaghoobzadeh et al.

There has been a growing interest in interpreting the underlying dynamics of Transformers. While self-attention patterns were initially deemed as the primary option, recent studies have shown that integrating other components can yield more accurate explanations. This paper introduces a novel token attribution analysis method that incorporates all the components in the encoder block and aggregates this throughout layers. Through extensive quantitative and qualitative experiments, we demonstrate that our method can produce faithful and meaningful global token attributions. Our experiments reveal that incorporating almost every encoder component results in increasingly more accurate analysis in both local (single layer) and global (the whole model) settings. Our global attribution analysis significantly outperforms previous methods on various tasks regarding correlation with gradient-based saliency scores. Our code is freely available at https://github.com/mohsenfayyaz/GlobEnc.

Ali Modarressi, Mohsen Fayyaz, Ehsan Aghazadeh et al.

An emerging solution for explaining Transformer-based models is to use vector-based analysis on how the representations are formed. However, providing a faithful vector-based explanation for a multi-layer model could be challenging in three aspects: (1) Incorporating all components into the analysis, (2) Aggregating the layer dynamics to determine the information flow and mixture throughout the entire model, and (3) Identifying the connection between the vector-based analysis and the model's predictions. In this paper, we present DecompX to tackle these challenges. DecompX is based on the construction of decomposed token representations and their successive propagation throughout the model without mixing them in between layers. Additionally, our proposal provides multiple advantages over existing solutions for its inclusion of all encoder components (especially nonlinear feed-forward networks) and the classification head. The former allows acquiring precise vectors while the latter transforms the decomposition into meaningful prediction-based values, eliminating the need for norm- or summation-based vector aggregation. According to the standard faithfulness evaluations, DecompX consistently outperforms existing gradient-based and vector-based approaches on various datasets. Our code is available at https://github.com/mohsenfayyaz/DecompX.

Ali Modarressi, Hosein Mohebbi, Mohammad Taher Pilehvar

Pre-trained language models have shown stellar performance in various downstream tasks. But, this usually comes at the cost of high latency and computation, hindering their usage in resource-limited settings. In this work, we propose a novel approach for reducing the computational cost of BERT with minimal loss in downstream performance. Our method dynamically eliminates less contributing tokens through layers, resulting in shorter lengths and consequently lower computational cost. To determine the importance of each token representation, we train a Contribution Predictor for each layer using a gradient-based saliency method. Our experiments on several diverse classification tasks show speedups up to 22x during inference time without much sacrifice in performance. We also validate the quality of the selected tokens in our method using human annotations in the ERASER benchmark. In comparison to other widely used strategies for selecting important tokens, such as saliency and attention, our proposed method has a significantly lower false positive rate in generating rationales. Our code is freely available at https://github.com/amodaresi/AdapLeR .

Houman Mehrafarin, Sara Rajaee, Mohammad Taher Pilehvar

Several studies have investigated the reasons behind the effectiveness of fine-tuning, usually through the lens of probing. However, these studies often neglect the role of the size of the dataset on which the model is fine-tuned. In this paper, we highlight the importance of this factor and its undeniable role in probing performance. We show that the extent of encoded linguistic knowledge depends on the number of fine-tuning samples. The analysis also reveals that larger training data mainly affects higher layers, and that the extent of this change is a factor of the number of iterations updating the model during fine-tuning rather than the diversity of the training samples. Finally, we show through a set of experiments that fine-tuning data size affects the recoverability of the changes made to the model's linguistic knowledge.

Mohsen Fayyaz, Ehsan Aghazadeh, Ali Modarressi et al.

Current pre-trained language models rely on large datasets for achieving state-of-the-art performance. However, past research has shown that not all examples in a dataset are equally important during training. In fact, it is sometimes possible to prune a considerable fraction of the training set while maintaining the test performance. Established on standard vision benchmarks, two gradient-based scoring metrics for finding important examples are GraNd and its estimated version, EL2N. In this work, we employ these two metrics for the first time in NLP. We demonstrate that these metrics need to be computed after at least one epoch of fine-tuning and they are not reliable in early steps. Furthermore, we show that by pruning a small portion of the examples with the highest GraNd/EL2N scores, we can not only preserve the test accuracy, but also surpass it. This paper details adjustments and implementation choices which enable GraNd and EL2N to be applied to NLP.

Mahdi Zakizadeh, Kaveh Eskandari Miandoab, Mohammad Taher Pilehvar

Numerous debiasing techniques have been proposed to mitigate the gender bias that is prevalent in pretrained language models. These are often evaluated on datasets that check the extent to which the model is gender-neutral in its predictions. Importantly, this evaluation protocol overlooks the possible adverse impact of bias mitigation on useful gender knowledge. To fill this gap, we propose DiFair, a manually curated dataset based on masked language modeling objectives. DiFair allows us to introduce a unified metric, gender invariance score, that not only quantifies a model's biased behavior, but also checks if useful gender knowledge is preserved. We use DiFair as a benchmark for a number of widely-used pretained language models and debiasing techniques. Experimental results corroborate previous findings on the existing gender biases, while also demonstrating that although debiasing techniques ameliorate the issue of gender bias, this improvement usually comes at the price of lowering useful gender knowledge of the model.

Sara Rajaee, Yadollah Yaghoobzadeh, Mohammad Taher Pilehvar

It has been shown that NLI models are usually biased with respect to the word-overlap between premise and hypothesis; they take this feature as a primary cue for predicting the entailment label. In this paper, we focus on an overlooked aspect of the overlap bias in NLI models: the reverse word-overlap bias. Our experimental results demonstrate that current NLI models are highly biased towards the non-entailment label on instances with low overlap, and the existing debiasing methods, which are reportedly successful on existing challenge datasets, are generally ineffective in addressing this category of bias. We investigate the reasons for the emergence of the overlap bias and the role of minority examples in its mitigation. For the former, we find that the word-overlap bias does not stem from pre-training, and for the latter, we observe that in contrast to the accepted assumption, eliminating minority examples does not affect the generalizability of debiasing methods with respect to the overlap bias.

Ali Modarressi, Hossein Amirkhani, Mohammad Taher Pilehvar

Several proposals have been put forward in recent years for improving out-of-distribution (OOD) performance through mitigating dataset biases. A popular workaround is to train a robust model by re-weighting training examples based on a secondary biased model. Here, the underlying assumption is that the biased model resorts to shortcut features. Hence, those training examples that are correctly predicted by the biased model are flagged as being biased and are down-weighted during the training of the main model. However, assessing the importance of an instance merely based on the predictions of the biased model may be too naive. It is possible that the prediction of the main model can be derived from another decision-making process that is distinct from the behavior of the biased model. To circumvent this, we introduce a fine-tuning strategy that incorporates the similarity between the main and biased model attribution scores in a Product of Experts (PoE) loss function to further improve OOD performance. With experiments conducted on natural language inference and fact verification benchmarks, we show that our method improves OOD results while maintaining in-distribution (ID) performance.

Mohammad Akbar-Tajari, Sara Rajaee, Mohammad Taher Pilehvar

Parameter-efficient fine-tuning approaches have recently garnered a lot of attention. Having considerably lower number of trainable weights, these methods can bring about scalability and computational effectiveness. In this paper, we look for optimal sub-networks and investigate the capability of different transformer modules in transferring knowledge from a pre-trained model to a downstream task. Our empirical results suggest that every transformer module in BERT can act as a winning ticket: fine-tuning each specific module while keeping the rest of the network frozen can lead to comparable performance to the full fine-tuning. Among different modules, LayerNorms exhibit the best capacity for knowledge transfer with limited trainable weights, to the extent that, with only 0.003% of all parameters in the layer-wise analysis, they show acceptable performance on various target tasks. On the reasons behind their effectiveness, we argue that their notable performance could be attributed to their high-magnitude weights compared to that of the other modules in the pre-trained BERT.

Jingxuan Chen, Mohammad Taher Pilehvar, Jose Camacho-Collados

Users often rely on Large Language Models (LLMs) for processing multiple documents or performing analysis over a number of instances. For example, analysing the overall sentiment of a number of movie reviews requires an LLM to process the sentiment of each review individually in order to provide a final aggregated answer. While LLM performance on such individual tasks is generally high, there has been little research on how LLMs perform when dealing with multi-instance inputs. In this paper, we perform a comprehensive evaluation of the multi-instance processing (MIP) ability of LLMs for tasks in which they excel individually. The results show that all LLMs follow a pattern of slight performance degradation for small numbers of instances (approximately 20-100), followed by a performance collapse on larger instance counts. Crucially, our analysis shows that while context length is associated with this degradation, the number of instances has a stronger effect on the final results. This finding suggests that when optimising LLM performance for MIP, attention should be paid to both context length and, in particular, instance count.

Kiamehr Rezaee, Jose Camacho-Collados, Mohammad Taher Pilehvar

Large Language Models (LLMs) have demonstrated impressive capabilities in solving complex tasks, including those requiring a certain level of reasoning. In this paper, we focus on state tracking, a problem where models need to keep track of the state governing a number of entities. To isolate the state tracking component from other factors, we propose a benchmark based on three well-defined state tracking tasks and analyse the performance of LLMs in different scenarios. The results indicate that the recent generation of LLMs (specifically, GPT-4 and Llama3) are capable of tracking state, especially when integrated with mechanisms such as Chain of Thought. However, models from the former generation, while understanding the task and being able to solve it at the initial stages, often fail at this task after a certain number of steps.

Erfan Moosavi Monazzah, Vahid Rahimzadeh, Yadollah Yaghoobzadeh et al.

Large language models predominantly reflect Western cultures, largely due to the dominance of English-centric training data. This imbalance presents a significant challenge, as LLMs are increasingly used across diverse contexts without adequate evaluation of their cultural competence in non-English languages, including Persian. To address this gap, we introduce PerCul, a carefully constructed dataset designed to assess the sensitivity of LLMs toward Persian culture. PerCul features story-based, multiple-choice questions that capture culturally nuanced scenarios. Unlike existing benchmarks, PerCul is curated with input from native Persian annotators to ensure authenticity and to prevent the use of translation as a shortcut. We evaluate several state-of-the-art multilingual and Persian-specific LLMs, establishing a foundation for future research in cross-cultural NLP evaluation. Our experiments demonstrate a 11.3% gap between best closed source model and layperson baseline while the gap increases to 21.3% by using the best open-weight model. You can access the dataset from here: https://huggingface.co/datasets/teias-ai/percul

Majid Zarharan, Pascal Wullschleger, Babak Behkam Kia et al.

This paper presents a comprehensive analysis of explainable fact-checking through a series of experiments, focusing on the ability of large language models to verify public health claims and provide explanations or justifications for their veracity assessments. We examine the effectiveness of zero/few-shot prompting and parameter-efficient fine-tuning across various open and closed-source models, examining their performance in both isolated and joint tasks of veracity prediction and explanation generation. Importantly, we employ a dual evaluation approach comprising previously established automatic metrics and a novel set of criteria through human evaluation. Our automatic evaluation indicates that, within the zero-shot scenario, GPT-4 emerges as the standout performer, but in few-shot and parameter-efficient fine-tuning contexts, open-source models demonstrate their capacity to not only bridge the performance gap but, in some instances, surpass GPT-4. Human evaluation reveals yet more nuance as well as indicating potential problems with the gold explanations.

Mohammad Akbar-Tajari, Mohammad Taher Pilehvar, Mohammad Mahmoody

The challenge of ensuring Large Language Models (LLMs) align with societal standards is of increasing interest, as these models are still prone to adversarial jailbreaks that bypass their safety mechanisms. Identifying these vulnerabilities is crucial for enhancing the robustness of LLMs against such exploits. We propose Graph of ATtacks (GoAT), a method for generating adversarial prompts to test the robustness of LLM alignment using the Graph of Thoughts framework [Besta et al., 2024]. GoAT excels at generating highly effective jailbreak prompts with fewer queries to the victim model than state-of-the-art attacks, achieving up to five times better jailbreak success rate against robust models like Llama. Notably, GoAT creates high-quality, human-readable prompts without requiring access to the targeted model's parameters, making it a black-box attack. Unlike approaches constrained by tree-based reasoning, GoAT's reasoning is based on a more intricate graph structure. By making simultaneous attack paths aware of each other's progress, this dynamic framework allows a deeper integration and refinement of reasoning paths, significantly enhancing the collaborative exploration of adversarial vulnerabilities in LLMs. At a technical level, GoAT starts with a graph structure and iteratively refines it by combining and improving thoughts, enabling synergy between different thought paths. The code for our implementation can be found at: https://github.com/GoAT-pydev/Graph_of_Attacks.

Faridoun Mehri, Mahdieh Soleymani Baghshah, Mohammad Taher Pilehvar

Why do gradient-based explanations struggle with Transformers, and how can we improve them? We identify gradient flow imbalances in Transformers that violate FullGrad-completeness, a critical property for attribution faithfulness that CNNs naturally possess. To address this issue, we introduce LibraGrad -- a theoretically grounded post-hoc approach that corrects gradient imbalances through pruning and scaling of backward paths, without changing the forward pass or adding computational overhead. We evaluate LibraGrad using three metric families: Faithfulness, which quantifies prediction changes under perturbations of the most and least relevant features; Completeness Error, which measures attribution conservation relative to model outputs; and Segmentation AP, which assesses alignment with human perception. Extensive experiments across 8 architectures, 4 model sizes, and 4 datasets show that LibraGrad universally enhances gradient-based methods, outperforming existing white-box methods -- including Transformer-specific approaches -- across all metrics. We demonstrate superior qualitative results through two complementary evaluations: precise text-prompted region highlighting on CLIP models and accurate class discrimination between co-occurring animals on ImageNet-finetuned models -- two settings on which existing methods often struggle. LibraGrad is effective even on the attention-free MLP-Mixer architecture, indicating potential for extension to other modern architectures. Our code is freely available at https://github.com/NightMachinery/LibraGrad.

Junho Myung, Nayeon Lee, Yi Zhou et al.

Large language models (LLMs) often lack culture-specific knowledge of daily life, especially across diverse regions and non-English languages. Existing benchmarks for evaluating LLMs' cultural sensitivities are limited to a single language or collected from online sources such as Wikipedia, which do not reflect the mundane everyday lifestyles of diverse regions. That is, information about the food people eat for their birthday celebrations, spices they typically use, musical instruments youngsters play, or the sports they practice in school is common cultural knowledge but uncommon in easily collected online sources, especially for underrepresented cultures. To address this issue, we introduce BLEnD, a hand-crafted benchmark designed to evaluate LLMs' everyday knowledge across diverse cultures and languages. BLEnD comprises 52.6k question-answer pairs from 16 countries/regions, in 13 different languages, including low-resource ones such as Amharic, Assamese, Azerbaijani, Hausa, and Sundanese. We construct the benchmark to include two formats of questions: short-answer and multiple-choice. We show that LLMs perform better for cultures that are highly represented online, with a maximum 57.34% difference in GPT-4, the best-performing model, in the short-answer format. For cultures represented by mid-to-high-resource languages, LLMs perform better in their local languages, but for cultures represented by low-resource languages, LLMs perform better in English than the local languages. We make our dataset publicly available at: https://github.com/nlee0212/BLEnD.

Matteo Marcuzzo, Alessandro Zangari, Andrea Albarelli et al.

As LLMs excel on standard reading comprehension benchmarks, attention is shifting toward evaluating their capacity for complex abstract reasoning and inference. Literature-based benchmarks, with their rich narrative and moral depth, provide a compelling framework for evaluating such deeper comprehension skills. Here, we present MORABLES, a human-verified benchmark built from fables and short stories drawn from historical literature. The main task is structured as multiple-choice questions targeting moral inference, with carefully crafted distractors that challenge models to go beyond shallow, extractive question answering. To further stress-test model robustness, we introduce adversarial variants designed to surface LLM vulnerabilities and shortcuts due to issues such as data contamination. Our findings show that, while larger models outperform smaller ones, they remain susceptible to adversarial manipulation and often rely on superficial patterns rather than true moral reasoning. This brittleness results in significant self-contradiction, with the best models refuting their own answers in roughly 20% of cases depending on the framing of the moral choice. Interestingly, reasoning-enhanced models fail to bridge this gap, suggesting that scale - not reasoning ability - is the primary driver of performance.

Sina Abbasi, Mohammad Reza Modarres, Mohammad Taher Pilehvar

With new large language models (LLMs) emerging frequently, it is important to consider the potential value of model-agnostic approaches that can provide interpretability across a variety of architectures. While recent advances in LLM interpretability show promise, many rely on complex, model-specific methods with high computational costs. To address these limitations, we propose NormXLogit, a novel technique for assessing the significance of individual input tokens. This method operates based on the input and output representations associated with each token. First, we demonstrate that during the pre-training of LLMs, the norms of word embeddings effectively capture token importance. Second, we reveal a significant relationship between a token's importance and the extent to which its representation can resemble the model's final prediction. Extensive analyses reveal that our approach outperforms existing gradient-based methods in terms of faithfulness and offers competitive performance in layer-wise explanations compared to leading architecture-specific techniques.

Vahid Rahimzadeh, Erfan Moosavi Monazzah, Mohammad Taher Pilehvar et al.

Persona-driven LLMs have emerged as powerful tools in computational social science, yet existing approaches fall at opposite extremes, either relying on costly human-curated data or producing synthetic personas that lack consistency and realism. We introduce SYNTHIA, a dataset of 30,000 backstories derived from 10,000 real social media users from BlueSky open platform across three time windows, bridging this spectrum by grounding synthetic generation in authentic user activity. Our evaluation demonstrates that SYNTHIA achieves competitive performance with state-of-the-art methods in demographic diversity and social survey alignment while significantly outperforming them in narrative consistency. Uniquely, SYNTHIA incorporates temporal dimensionality and provides rich social interaction metadata from the underlying network, enabling new research directions in computational social science and persona-driven language modeling.

Majid Zarharan, Maryam Hashemi, Malika Behroozrazegh et al.

We introduce FarExStance, a new dataset for explainable stance detection in Farsi. Each instance in this dataset contains a claim, the stance of an article or social media post towards that claim, and an extractive explanation which provides evidence for the stance label. We compare the performance of a fine-tuned multilingual RoBERTa model to several large language models in zero-shot, few-shot, and parameter-efficient fine-tuned settings on our new dataset. On stance detection, the most accurate models are the fine-tuned RoBERTa model, the LLM Aya-23-8B which has been fine-tuned using parameter-efficient fine-tuning, and few-shot Claude-3.5-Sonnet. Regarding the quality of the explanations, our automatic evaluation metrics indicate that few-shot GPT-4o generates the most coherent explanations, while our human evaluation reveals that the best Overall Explanation Score (OES) belongs to few-shot Claude-3.5-Sonnet. The fine-tuned Aya-32-8B model produced explanations most closely aligned with the reference explanations.

Alessandro Zangari, Matteo Marcuzzo, Andrea Albarelli et al.

Puns are a form of humorous wordplay that exploits polysemy and phonetic similarity. While LLMs have shown promise in detecting puns, we show in this paper that their understanding often remains shallow, lacking the nuanced grasp typical of human interpretation. By systematically analyzing and reformulating existing pun benchmarks, we demonstrate how subtle changes in puns are sufficient to mislead LLMs. Our contributions include comprehensive and nuanced pun detection benchmarks, human evaluation across recent LLMs, and an analysis of the robustness challenges these models face in processing puns.

Mahdi Zakizadeh, Mohammad Taher Pilehvar

Gender bias in pretrained language models (PLMs) poses significant social and ethical challenges. Despite growing awareness, there is a lack of comprehensive investigation into how different models internally represent and propagate such biases. This study adopts an information-theoretic approach to analyze how gender biases are encoded within various encoder-based architectures. We focus on three key aspects: identifying how models encode gender information and biases, examining the impact of bias mitigation techniques and fine-tuning on the encoded biases and their effectiveness, and exploring how model design differences influence the encoding of biases. Through rigorous and systematic investigation, our findings reveal a consistent pattern of gender encoding across diverse models. Surprisingly, debiasing techniques often exhibit limited efficacy, sometimes inadvertently increasing the encoded bias in internal representations while reducing bias in model output distributions. This highlights a disconnect between mitigating bias in output distributions and addressing its internal representations. This work provides valuable guidance for advancing bias mitigation strategies and fostering the development of more equitable language models.

Mahdi Zakizadeh, Mohammad Taher Pilehvar

Accurately measuring gender stereotypical bias in language models is a complex task with many hidden aspects. Current benchmarks have underestimated this multifaceted challenge and failed to capture the full extent of the problem. This paper examines the inconsistencies between intrinsic stereotype benchmarks. We propose that currently available benchmarks each capture only partial facets of gender stereotypes, and when considered in isolation, they provide just a fragmented view of the broader landscape of bias in language models. Using StereoSet and CrowS-Pairs as case studies, we investigated how data distribution affects benchmark results. By applying a framework from social psychology to balance the data of these benchmarks across various components of gender stereotypes, we demonstrated that even simple balancing techniques can significantly improve the correlation between different measurement approaches. Our findings underscore the complexity of gender stereotyping in language models and point to new directions for developing more refined techniques to detect and reduce bias.

Mohammad Reza Modarres, Sina Abbasi, Mohammad Taher Pilehvar

Advances in dataset analysis techniques have enabled more sophisticated approaches to analyzing and characterizing training data instances, often categorizing data based on attributes such as ``difficulty''. In this work, we introduce RepMatch, a novel method that characterizes data through the lens of similarity. RepMatch quantifies the similarity between subsets of training instances by comparing the knowledge encoded in models trained on them, overcoming the limitations of existing analysis methods that focus solely on individual instances and are restricted to within-dataset analysis. Our framework allows for a broader evaluation, enabling similarity comparisons across arbitrary subsets of instances, supporting both dataset-to-dataset and instance-to-dataset analyses. We validate the effectiveness of RepMatch across multiple NLP tasks, datasets, and models. Through extensive experimentation, we demonstrate that RepMatch can effectively compare datasets, identify more representative subsets of a dataset (that lead to better performance than randomly selected subsets of equivalent size), and uncover heuristics underlying the construction of some challenge datasets.

Sara Rajaee, Mohammad Taher Pilehvar

Several studies have explored various advantages of multilingual pre-trained models (such as multilingual BERT) in capturing shared linguistic knowledge. However, less attention has been paid to their limitations. In this paper, we investigate the multilingual BERT for two known issues of the monolingual models: anisotropic embedding space and outlier dimensions. We show that, unlike its monolingual counterpart, the multilingual BERT model exhibits no outlier dimension in its representations while it has a highly anisotropic space. There are a few dimensions in the monolingual BERT with high contributions to the anisotropic distribution. However, we observe no such dimensions in the multilingual BERT. Furthermore, our experimental results demonstrate that increasing the isotropy of multilingual space can significantly improve its representation power and performance, similarly to what had been observed for monolingual CWRs on semantic similarity tasks. Our analysis indicates that, despite having different degenerated directions, the embedding spaces in various languages tend to be partially similar with respect to their structures.

Mohsen Fayyaz, Ehsan Aghazadeh, Ali Modarressi et al.

Most of the recent works on probing representations have focused on BERT, with the presumption that the findings might be similar to the other models. In this work, we extend the probing studies to two other models in the family, namely ELECTRA and XLNet, showing that variations in the pre-training objectives or architectural choices can result in different behaviors in encoding linguistic information in the representations. Most notably, we observe that ELECTRA tends to encode linguistic knowledge in the deeper layers, whereas XLNet instead concentrates that in the earlier layers. Also, the former model undergoes a slight change during fine-tuning, whereas the latter experiences significant adjustments. Moreover, we show that drawing conclusions based on the weight mixing evaluation strategy -- which is widely used in the context of layer-wise probing -- can be misleading given the norm disparity of the representations across different layers. Instead, we adopt an alternative information-theoretic probing with minimum description length, which has recently been proven to provide more reliable and informative results.

Sara Rajaee, Mohammad Taher Pilehvar

It is widely accepted that fine-tuning pre-trained language models usually brings about performance improvements in downstream tasks. However, there are limited studies on the reasons behind this effectiveness, particularly from the viewpoint of structural changes in the embedding space. Trying to fill this gap, in this paper, we analyze the extent to which the isotropy of the embedding space changes after fine-tuning. We demonstrate that, even though isotropy is a desirable geometrical property, fine-tuning does not necessarily result in isotropy enhancements. Moreover, local structures in pre-trained contextual word representations (CWRs), such as those encoding token types or frequency, undergo a massive change during fine-tuning. Our experiments show dramatic growth in the number of elongated directions in the embedding space, which, in contrast to pre-trained CWRs, carry the essential linguistic knowledge in the fine-tuned embedding space, making existing isotropy enhancement methods ineffective.

Hossein Amirkhani, Mohammad Taher Pilehvar

Existing techniques for mitigating dataset bias often leverage a biased model to identify biased instances. The role of these biased instances is then reduced during the training of the main model to enhance its robustness to out-of-distribution data. A common core assumption of these techniques is that the main model handles biased instances similarly to the biased model, in that it will resort to biases whenever available. In this paper, we show that this assumption does not hold in general. We carry out a critical investigation on two well-known datasets in the domain, MNLI and FEVER, along with two biased instance detection methods, partial-input and limited-capacity models. Our experiments show that in around a third to a half of instances, the biased model is unable to predict the main model's behavior, highlighted by the significantly different parts of the input on which they base their decisions. Based on a manual validation, we also show that this estimate is highly in line with human interpretation. Our findings suggest that down-weighting of instances detected by bias detection methods, which is a widely-practiced procedure, is an unnecessary waste of training data. We release our code to facilitate reproducibility and future research.

Sara Rajaee, Mohammad Taher Pilehvar

The representation degeneration problem in Contextual Word Representations (CWRs) hurts the expressiveness of the embedding space by forming an anisotropic cone where even unrelated words have excessively positive correlations. Existing techniques for tackling this issue require a learning process to re-train models with additional objectives and mostly employ a global assessment to study isotropy. Our quantitative analysis over isotropy shows that a local assessment could be more accurate due to the clustered structure of CWRs. Based on this observation, we propose a local cluster-based method to address the degeneration issue in contextual embedding spaces. We show that in clusters including punctuations and stop words, local dominant directions encode structural information, removing which can improve CWRs performance on semantic tasks. Moreover, we find that tense information in verb representations dominates sense semantics. We show that removing dominant directions of verb representations can transform the space to better suit semantic applications. Our experiments demonstrate that the proposed cluster-based method can mitigate the degeneration problem on multiple tasks.

Hosein Mohebbi, Ali Modarressi, Mohammad Taher Pilehvar

Several studies have been carried out on revealing linguistic features captured by BERT. This is usually achieved by training a diagnostic classifier on the representations obtained from different layers of BERT. The subsequent classification accuracy is then interpreted as the ability of the model in encoding the corresponding linguistic property. Despite providing insights, these studies have left out the potential role of token representations. In this paper, we provide a more in-depth analysis on the representation space of BERT in search for distinct and meaningful subspaces that can explain the reasons behind these probing results. Based on a set of probing tasks and with the help of attribution methods we show that BERT tends to encode meaningful knowledge in specific token representations (which are often ignored in standard classification setups), allowing the model to detect syntactic and semantic abnormalities, and to distinctively separate grammatical number and tense subspaces.

Alessandro Raganato, Tommaso Pasini, Jose Camacho-Collados et al.

The ability to correctly model distinct meanings of a word is crucial for the effectiveness of semantic representation techniques. However, most existing evaluation benchmarks for assessing this criterion are tied to sense inventories (usually WordNet), restricting their usage to a small subset of knowledge-based representation techniques. The Word-in-Context dataset (WiC) addresses the dependence on sense inventories by reformulating the standard disambiguation task as a binary classification problem; but, it is limited to the English language. We put forward a large multilingual benchmark, XL-WiC, featuring gold standards in 12 new languages from varied language families and with different degrees of resource availability, opening room for evaluation scenarios such as zero-shot cross-lingual transfer. We perform a series of experiments to determine the reliability of the datasets and to set performance baselines for several recent contextualized multilingual models. Experimental results show that even when no tagged instances are available for a target language, models trained solely on the English data can attain competitive performance in the task of distinguishing different meanings of a word, even for distant languages. XL-WiC is available at https://pilehvar.github.io/xlwic/.

Daniel Loureiro, Kiamehr Rezaee, Mohammad Taher Pilehvar et al.

Transformer-based language models have taken many fields in NLP by storm. BERT and its derivatives dominate most of the existing evaluation benchmarks, including those for Word Sense Disambiguation (WSD), thanks to their ability in capturing context-sensitive semantic nuances. However, there is still little knowledge about their capabilities and potential limitations in encoding and recovering word senses. In this article, we provide an in-depth quantitative and qualitative analysis of the celebrated BERT model with respect to lexical ambiguity. One of the main conclusions of our analysis is that BERT can accurately capture high-level sense distinctions, even when a limited number of examples is available for each word sense. Our analysis also reveals that in some cases language models come close to solving coarse-grained noun disambiguation under ideal conditions in terms of availability of training data and computing resources. However, this scenario rarely occurs in real-world settings and, hence, many practical challenges remain even in the coarse-grained setting. We also perform an in-depth comparison of the two main language model based WSD strategies, i.e., fine-tuning and feature extraction, finding that the latter approach is more robust with respect to sense bias and it can better exploit limited available training data. In fact, the simple feature extraction strategy of averaging contextualized embeddings proves robust even using only three training sentences per word sense, with minimal improvements obtained by increasing the size of this training data.

Costanza Conforti, Jakob Berndt, Mohammad Taher Pilehvar et al.

We present a new challenging stance detection dataset, called Will-They-Won't-They (WT-WT), which contains 51,284 tweets in English, making it by far the largest available dataset of the type. All the annotations are carried out by experts; therefore, the dataset constitutes a high-quality and reliable benchmark for future research in stance detection. Our experiments with a wide range of recent state-of-the-art stance detection systems show that the dataset poses a strong challenge to existing models in this domain.

Anna Breit, Artem Revenko, Kiamehr Rezaee et al.

We present WiC-TSV, a new multi-domain evaluation benchmark for Word Sense Disambiguation. More specifically, we introduce a framework for Target Sense Verification of Words in Context which grounds its uniqueness in the formulation as a binary classification task thus being independent of external sense inventories, and the coverage of various domains. This makes the dataset highly flexible for the evaluation of a diverse set of models and systems in and across domains. WiC-TSV provides three different evaluation settings, depending on the input signals provided to the model. We set baseline performance on the dataset using state-of-the-art language models. Experimental results show that even though these models can perform decently on the task, there remains a gap between machine and human performance, especially in out-of-domain settings. WiC-TSV data is available at https://competitions.codalab.org/competitions/23683

Victor Prokhorov, Ehsan Shareghi, Yingzhen Li et al.

Variational Autoencoders (VAEs) are known to suffer from learning uninformative latent representation of the input due to issues such as approximated posterior collapse, or entanglement of the latent space. We impose an explicit constraint on the Kullback-Leibler (KL) divergence term inside the VAE objective function. While the explicit constraint naturally avoids posterior collapse, we use it to further understand the significance of the KL term in controlling the information transmitted through the VAE channel. Within this framework, we explore different properties of the estimated posterior distribution, and highlight the trade-off between the amount of information encoded in a latent code during training, and the generative capacity of the model.

Victor Prokhorov, Mohammad Taher Pilehvar, Nigel Collier

We present a novel method for mapping unrestricted text to knowledge graph entities by framing the task as a sequence-to-sequence problem. Specifically, given the encoded state of an input text, our decoder directly predicts paths in the knowledge graph, starting from the root and ending at the target node following hypernym-hyponym relationships. In this way, and in contrast to other text-to-entity mapping systems, our model outputs hierarchically structured predictions that are fully interpretable in the context of the underlying ontology, in an end-to-end manner. We present a proof-of-concept experiment with encouraging results, comparable to those of state-of-the-art systems.

Victor Prokhorov, Mohammad Taher Pilehvar, Dimitri Kartsaklis et al.

Word embedding techniques heavily rely on the abundance of training data for individual words. Given the Zipfian distribution of words in natural language texts, a large number of words do not usually appear frequently or at all in the training data. In this paper we put forward a technique that exploits the knowledge encoded in lexical resources, such as WordNet, to induce embeddings for unseen words. Our approach adapts graph embedding and cross-lingual vector space transformation techniques in order to merge lexical knowledge encoded in ontologies with that derived from corpus statistics. We show that the approach can provide consistent performance improvements across multiple evaluation benchmarks: in-vitro, on multiple rare word similarity datasets, and in-vivo, in two downstream text classification tasks.

Milan Gritta, Mohammad Taher Pilehvar, Nigel Collier

Empirical methods in geoparsing have thus far lacked a standard evaluation framework describing the task, metrics and data used to compare state-of-the-art systems. Evaluation is further made inconsistent, even unrepresentative of real-world usage by the lack of distinction between the different types of toponyms, which necessitates new guidelines, a consolidation of metrics and a detailed toponym taxonomy with implications for Named Entity Recognition (NER) and beyond. To address these deficiencies, our manuscript introduces a new framework in three parts. Part 1) Task Definition: clarified via corpus linguistic analysis proposing a fine-grained Pragmatic Taxonomy of Toponyms. Part 2) Metrics: discussed and reviewed for a rigorous evaluation including recommendations for NER/Geoparsing practitioners. Part 3) Evaluation Data: shared via a new dataset called GeoWebNews to provide test/train examples and enable immediate use of our contributions. In addition to fine-grained Geotagging and Toponym Resolution (Geocoding), this dataset is also suitable for prototyping and evaluating machine learning NLP models.

Mohammad Taher Pilehvar, Dimitri Kartsaklis, Victor Prokhorov et al.

Rare word representation has recently enjoyed a surge of interest, owing to the crucial role that effective handling of infrequent words can play in accurate semantic understanding. However, there is a paucity of reliable benchmarks for evaluation and comparison of these techniques. We show in this paper that the only existing benchmark (the Stanford Rare Word dataset) suffers from low-confidence annotations and limited vocabulary; hence, it does not constitute a solid comparison framework. In order to fill this evaluation gap, we propose CAmbridge Rare word Dataset (Card-660), an expert-annotated word similarity dataset which provides a highly reliable, yet challenging, benchmark for rare word representation techniques. Through a set of experiments we show that even the best mainstream word embeddings, with millions of words in their vocabularies, are unable to achieve performances higher than 0.43 (Pearson correlation) on the dataset, compared to a human-level upperbound of 0.90. We release the dataset and the annotation materials at https://pilehvar.github.io/card-660/.

Mohammad Taher Pilehvar, Jose Camacho-Collados

By design, word embeddings are unable to model the dynamic nature of words' semantics, i.e., the property of words to correspond to potentially different meanings. To address this limitation, dozens of specialized meaning representation techniques such as sense or contextualized embeddings have been proposed. However, despite the popularity of research on this topic, very few evaluation benchmarks exist that specifically focus on the dynamic semantics of words. In this paper we show that existing models have surpassed the performance ceiling of the standard evaluation dataset for the purpose, i.e., Stanford Contextual Word Similarity, and highlight its shortcomings. To address the lack of a suitable benchmark, we put forward a large-scale Word in Context dataset, called WiC, based on annotations curated by experts, for generic evaluation of context-sensitive representations. WiC is released in https://pilehvar.github.io/wic/.

Dimitri Kartsaklis, Mohammad Taher Pilehvar, Nigel Collier

This paper addresses the problem of mapping natural language text to knowledge base entities. The mapping process is approached as a composition of a phrase or a sentence into a point in a multi-dimensional entity space obtained from a knowledge graph. The compositional model is an LSTM equipped with a dynamic disambiguation mechanism on the input word embeddings (a Multi-Sense LSTM), addressing polysemy issues. Further, the knowledge base space is prepared by collecting random walks from a graph enhanced with textual features, which act as a set of semantic bridges between text and knowledge base entities. The ideas of this work are demonstrated on large-scale text-to-entity mapping and entity classification tasks, with state of the art results.

Jose Camacho-Collados, Luis Espinosa-Anke, Mohammad Taher Pilehvar

Incorporating linguistic, world and common sense knowledge into AI/NLP systems is currently an important research area, with several open problems and challenges. At the same time, processing and storing this knowledge in lexical resources is not a straightforward task. This tutorial proposes to address these complementary goals from two methodological perspectives: the use of NLP methods to help the process of constructing and enriching lexical resources and the use of lexical resources for improving NLP applications. Two main types of audience can benefit from this tutorial: those working on language resources who are interested in becoming acquainted with automatic NLP techniques, with the end goal of speeding and/or easing up the process of resource curation; and on the other hand, researchers in NLP who would like to benefit from the knowledge of lexical resources to improve their systems and models. The slides of the tutorial are available at https://bitbucket.org/luisespinosa/lr-nlp/

Jose Camacho-Collados, Mohammad Taher Pilehvar

Over the past years, distributed semantic representations have proved to be effective and flexible keepers of prior knowledge to be integrated into downstream applications. This survey focuses on the representation of meaning. We start from the theoretical background behind word vector space models and highlight one of their major limitations: the meaning conflation deficiency, which arises from representing a word with all its possible meanings as a single vector. Then, we explain how this deficiency can be addressed through a transition from the word level to the more fine-grained level of word senses (in its broader acceptation) as a method for modelling unambiguous lexical meaning. We present a comprehensive overview of the wide range of techniques in the two main branches of sense representation, i.e., unsupervised and knowledge-based. Finally, this survey covers the main evaluation procedures and applications for this type of representation, and provides an analysis of four of its important aspects: interpretability, sense granularity, adaptability to different domains and compositionality.

Mohammad Taher Pilehvar, Jose Camacho-Collados, Roberto Navigli et al.

Lexical ambiguity can impede NLP systems from accurate understanding of semantics. Despite its potential benefits, the integration of sense-level information into NLP systems has remained understudied. By incorporating a novel disambiguation algorithm into a state-of-the-art classification model, we create a pipeline to integrate sense-level information into downstream NLP applications. We show that a simple disambiguation of the input text can lead to consistent performance improvement on multiple topic categorization and polarity detection datasets, particularly when the fine granularity of the underlying sense inventory is reduced and the document is sufficiently large. Our results also point to the need for sense representation research to focus more on in vivo evaluations which target the performance in downstream NLP applications rather than artificial benchmarks.

Victor Prokhorov, Mohammad Taher Pilehvar, Dimitri Kartsaklis et al.

We propose a methodology that adapts graph embedding techniques (DeepWalk (Perozzi et al., 2014) and node2vec (Grover and Leskovec, 2016)) as well as cross-lingual vector space mapping approaches (Least Squares and Canonical Correlation Analysis) in order to merge the corpus and ontological sources of lexical knowledge. We also perform comparative analysis of the used algorithms in order to identify the best combination for the proposed system. We then apply this to the task of enhancing the coverage of an existing word embedding's vocabulary with rare and unseen words. We show that our technique can provide considerable extra coverage (over 99%), leading to consistent performance gain (around 10% absolute gain is achieved with w2v-gn-500K cf.§3.3) on the Rare Word Similarity dataset.

Jose Camacho-Collados, Mohammad Taher Pilehvar

Text preprocessing is often the first step in the pipeline of a Natural Language Processing (NLP) system, with potential impact in its final performance. Despite its importance, text preprocessing has not received much attention in the deep learning literature. In this paper we investigate the impact of simple text preprocessing decisions (particularly tokenizing, lemmatizing, lowercasing and multiword grouping) on the performance of a standard neural text classifier. We perform an extensive evaluation on standard benchmarks from text categorization and sentiment analysis. While our experiments show that a simple tokenization of input text is generally adequate, they also highlight significant degrees of variability across preprocessing techniques. This reveals the importance of paying attention to this usually-overlooked step in the pipeline, particularly when comparing different models. Finally, our evaluation provides insights into the best preprocessing practices for training word embeddings.

Mohammad Taher Pilehvar, Nigel Collier

One major deficiency of most semantic representation techniques is that they usually model a word type as a single point in the semantic space, hence conflating all the meanings that the word can have. Addressing this issue by learning distinct representations for individual meanings of words has been the subject of several research studies in the past few years. However, the generated sense representations are either not linked to any sense inventory or are unreliable for infrequent word senses. We propose a technique that tackles these problems by de-conflating the representations of words based on the deep knowledge it derives from a semantic network. Our approach provides multiple advantages in comparison to the past work, including its high coverage and the ability to generate accurate representations even for infrequent word senses. We carry out evaluations on six datasets across two semantic similarity tasks and report state-of-the-art results on most of them.

José Camacho-Collados, Ignacio Iacobacci, Roberto Navigli et al.

Representing the semantics of linguistic items in a machine-interpretable form has been a major goal of Natural Language Processing since its earliest days. Among the range of different linguistic items, words have attracted the most research attention. However, word representations have an important limitation: they conflate different meanings of a word into a single vector. Representations of word senses have the potential to overcome this inherent limitation. Indeed, the representation of individual word senses and concepts has recently gained in popularity with several experimental results showing that a considerable performance improvement can be achieved across different NLP applications upon moving from word level to the deeper sense and concept levels. Another interesting point regarding the representation of concepts and word senses is that these models can be seamlessly applied to other linguistic items, such as words, phrases and sentences.