Fahim Dalvi

Basel Mousi, Nadir Durrani, Fatema Ahmad et al. · utoronto

Arabic, with its rich diversity of dialects, remains significantly underrepresented in Large Language Models, particularly in dialectal variations. We address this gap by introducing seven synthetic datasets in dialects alongside Modern Standard Arabic (MSA), created using Machine Translation (MT) combined with human post-editing. We present AraDiCE, a benchmark for Arabic Dialect and Cultural Evaluation. We evaluate LLMs on dialect comprehension and generation, focusing specifically on low-resource Arabic dialects. Additionally, we introduce the first-ever fine-grained benchmark designed to evaluate cultural awareness across the Gulf, Egypt, and Levant regions, providing a novel dimension to LLM evaluation. Our findings demonstrate that while Arabic-specific models like Jais and AceGPT outperform multilingual models on dialectal tasks, significant challenges persist in dialect identification, generation, and translation. This work contributes $\approx$45K post-edited samples, a cultural benchmark, and highlights the importance of tailored training to improve LLM performance in capturing the nuances of diverse Arabic dialects and cultural contexts. We have released the dialectal translation models and benchmarks developed in this study (https://huggingface.co/datasets/QCRI/AraDiCE).

Fahim Dalvi, Maram Hasanain, Sabri Boughorbel et al.

The recent development and success of Large Language Models (LLMs) necessitate an evaluation of their performance across diverse NLP tasks in different languages. Although several frameworks have been developed and made publicly available, their customization capabilities for specific tasks and datasets are often complex for different users. In this study, we introduce the LLMeBench framework, which can be seamlessly customized to evaluate LLMs for any NLP task, regardless of language. The framework features generic dataset loaders, several model providers, and pre-implements most standard evaluation metrics. It supports in-context learning with zero- and few-shot settings. A specific dataset and task can be evaluated for a given LLM in less than 20 lines of code while allowing full flexibility to extend the framework for custom datasets, models, or tasks. The framework has been tested on 31 unique NLP tasks using 53 publicly available datasets within 90 experimental setups, involving approximately 296K data points. We open-sourced LLMeBench for the community (https://github.com/qcri/LLMeBench/) and a video demonstrating the framework is available online. (https://youtu.be/9cC2m_abk3A)

Yimin Fan, Fahim Dalvi, Nadir Durrani et al.

Neuron Interpretation has gained traction in the field of interpretability, and have provided fine-grained insights into what a model learns and how language knowledge is distributed amongst its different components. However, the lack of evaluation benchmark and metrics have led to siloed progress within these various methods, with very little work comparing them and highlighting their strengths and weaknesses. The reason for this discrepancy is the difficulty of creating ground truth datasets, for example, many neurons within a given model may learn the same phenomena, and hence there may not be one correct answer. Moreover, a learned phenomenon may spread across several neurons that work together -- surfacing these to create a gold standard challenging. In this work, we propose an evaluation framework that measures the compatibility of a neuron analysis method with other methods. We hypothesize that the more compatible a method is with the majority of the methods, the more confident one can be about its performance. We systematically evaluate our proposed framework and present a comparative analysis of a large set of neuron interpretation methods. We make the evaluation framework available to the community. It enables the evaluation of any new method using 20 concepts and across three pre-trained models.The code is released at https://github.com/fdalvi/neuron-comparative-analysis

Fahim Dalvi, Abdul Rafae Khan, Firoj Alam et al.

A large number of studies that analyze deep neural network models and their ability to encode various linguistic and non-linguistic concepts provide an interpretation of the inner mechanics of these models. The scope of the analyses is limited to pre-defined concepts that reinforce the traditional linguistic knowledge and do not reflect on how novel concepts are learned by the model. We address this limitation by discovering and analyzing latent concepts learned in neural network models in an unsupervised fashion and provide interpretations from the model's perspective. In this work, we study: i) what latent concepts exist in the pre-trained BERT model, ii) how the discovered latent concepts align or diverge from classical linguistic hierarchy and iii) how the latent concepts evolve across layers. Our findings show: i) a model learns novel concepts (e.g. animal categories and demographic groups), which do not strictly adhere to any pre-defined categorization (e.g. POS, semantic tags), ii) several latent concepts are based on multiple properties which may include semantics, syntax, and morphology, iii) the lower layers in the model dominate in learning shallow lexical concepts while the higher layers learn semantic relations and iv) the discovered latent concepts highlight potential biases learned in the model. We also release a novel BERT ConceptNet dataset (BCN) consisting of 174 concept labels and 1M annotated instances.

Hassan Sajjad, Nadir Durrani, Fahim Dalvi et al.

We propose a novel framework ConceptX, to analyze how latent concepts are encoded in representations learned within pre-trained language models. It uses clustering to discover the encoded concepts and explains them by aligning with a large set of human-defined concepts. Our analysis on seven transformer language models reveal interesting insights: i) the latent space within the learned representations overlap with different linguistic concepts to a varying degree, ii) the lower layers in the model are dominated by lexical concepts (e.g., affixation), whereas the core-linguistic concepts (e.g., morphological or syntactic relations) are better represented in the middle and higher layers, iii) some encoded concepts are multi-faceted and cannot be adequately explained using the existing human-defined concepts.

Ahmed Abdelali, Nadir Durrani, Cenk Demiroglu et al.

NatiQ is end-to-end text-to-speech system for Arabic. Our speech synthesizer uses an encoder-decoder architecture with attention. We used both tacotron-based models (tacotron-1 and tacotron-2) and the faster transformer model for generating mel-spectrograms from characters. We concatenated Tacotron1 with the WaveRNN vocoder, Tacotron2 with the WaveGlow vocoder and ESPnet transformer with the parallel wavegan vocoder to synthesize waveforms from the spectrograms. We used in-house speech data for two voices: 1) neutral male "Hamza"- narrating general content and news, and 2) expressive female "Amina"- narrating children story books to train our models. Our best systems achieve an average Mean Opinion Score (MOS) of 4.21 and 4.40 for Amina and Hamza respectively. The objective evaluation of the systems using word and character error rate (WER and CER) as well as the response time measured by real-time factor favored the end-to-end architecture ESPnet. NatiQ demo is available on-line at https://tts.qcri.org

Nadir Durrani, Hassan Sajjad, Fahim Dalvi et al.

We study the evolution of latent space in fine-tuned NLP models. Different from the commonly used probing-framework, we opt for an unsupervised method to analyze representations. More specifically, we discover latent concepts in the representational space using hierarchical clustering. We then use an alignment function to gauge the similarity between the latent space of a pre-trained model and its fine-tuned version. We use traditional linguistic concepts to facilitate our understanding and also study how the model space transforms towards task-specific information. We perform a thorough analysis, comparing pre-trained and fine-tuned models across three models and three downstream tasks. The notable findings of our work are: i) the latent space of the higher layers evolve towards task-specific concepts, ii) whereas the lower layers retain generic concepts acquired in the pre-trained model, iii) we discovered that some concepts in the higher layers acquire polarity towards the output class, and iv) that these concepts can be used for generating adversarial triggers.

Nadir Durrani, Fahim Dalvi, Hassan Sajjad

While a lot of work has been done in understanding representations learned within deep NLP models and what knowledge they capture, little attention has been paid towards individual neurons. We present a technique called as Linguistic Correlation Analysis to extract salient neurons in the model, with respect to any extrinsic property - with the goal of understanding how such a knowledge is preserved within neurons. We carry out a fine-grained analysis to answer the following questions: (i) can we identify subsets of neurons in the network that capture specific linguistic properties? (ii) how localized or distributed neurons are across the network? iii) how redundantly is the information preserved? iv) how fine-tuning pre-trained models towards downstream NLP tasks, impacts the learned linguistic knowledge? iv) how do architectures vary in learning different linguistic properties? Our data-driven, quantitative analysis illuminates interesting findings: (i) we found small subsets of neurons that can predict different linguistic tasks, ii) with neurons capturing basic lexical information (such as suffixation) localized in lower most layers, iii) while those learning complex concepts (such as syntactic role) predominantly in middle and higher layers, iii) that salient linguistic neurons are relocated from higher to lower layers during transfer learning, as the network preserve the higher layers for task specific information, iv) we found interesting differences across pre-trained models, with respect to how linguistic information is preserved within, and v) we found that concept exhibit similar neuron distribution across different languages in the multilingual transformer models. Our code is publicly available as part of the NeuroX toolkit.

Preslav Nakov, Firoj Alam, Yifan Zhang et al.

Fighting the ongoing COVID-19 infodemic has been declared as one of the most important focus areas by the World Health Organization since the onset of the COVID-19 pandemic. While the information that is consumed and disseminated consists of promoting fake cures, rumors, and conspiracy theories to spreading xenophobia and panic, at the same time there is information (e.g., containing advice, promoting cure) that can help different stakeholders such as policy-makers. Social media platforms enable the infodemic and there has been an effort to curate the content on such platforms, analyze and debunk them. While a majority of the research efforts consider one or two aspects (e.g., detecting factuality) of such information, in this study we focus on a multifaceted approach, including an API,\url{https://app.swaggerhub.com/apis/yifan2019/Tanbih/0.8.0/} and a demo system,\url{https://covid19.tanbih.org}, which we made freely and publicly available. We believe that this will facilitate researchers and different stakeholders. A screencast of the API services and demo is available.\url{https://youtu.be/zhbcSvxEKMk}

Majd Hawasly, Fahim Dalvi, Nadir Durrani

Despite the revolution caused by deep NLP models, they remain black boxes, necessitating research to understand their decision-making processes. A recent work by Dalvi et al. (2022) carried out representation analysis through the lens of clustering latent spaces within pre-trained models (PLMs), but that approach is limited to small scale due to the high cost of running Agglomerative hierarchical clustering. This paper studies clustering algorithms in order to scale the discovery of encoded concepts in PLM representations to larger datasets and models. We propose metrics for assessing the quality of discovered latent concepts and use them to compare the studied clustering algorithms. We found that K-Means-based concept discovery significantly enhances efficiency while maintaining the quality of the obtained concepts. Furthermore, we demonstrate the practicality of this newfound efficiency by scaling latent concept discovery to LLMs and phrasal concepts.

Firoj Alam, Fahim Dalvi, Nadir Durrani et al.

The opacity of deep neural networks remains a challenge in deploying solutions where explanation is as important as precision. We present ConceptX, a human-in-the-loop framework for interpreting and annotating latent representational space in pre-trained Language Models (pLMs). We use an unsupervised method to discover concepts learned in these models and enable a graphical interface for humans to generate explanations for the concepts. To facilitate the process, we provide auto-annotations of the concepts (based on traditional linguistic ontologies). Such annotations enable development of a linguistic resource that directly represents latent concepts learned within deep NLP models. These include not just traditional linguistic concepts, but also task-specific or sensitive concepts (words grouped based on gender or religious connotation) that helps the annotators to mark bias in the model. The framework consists of two parts (i) concept discovery and (ii) annotation platform.

FANAR TEAM, Ummar Abbas, Mohammad Shahmeer Ahmad et al.

We present Fanar 2.0, the second generation of Qatar's Arabic-centric Generative AI platform. Sovereignty is a first-class design principle: every component, from data pipelines to deployment infrastructure, was designed and operated entirely at QCRI, Hamad Bin Khalifa University. Fanar 2.0 is a story of resource-constrained excellence: the effort ran on 256 NVIDIA H100 GPUs, with Arabic having only ~0.5% of web data despite 400 million native speakers. Fanar 2.0 adopts a disciplined strategy of data quality over quantity, targeted continual pre-training, and model merging to achieve substantial gains within these constraints. At the core is Fanar-27B, continually pre-trained from a Gemma-3-27B backbone on a curated corpus of 120 billion high-quality tokens across three data recipes. Despite using 8x fewer pre-training tokens than Fanar 1.0, it delivers substantial benchmark improvements: Arabic knowledge (+9.1 pts), language (+7.3 pts), dialects (+3.5 pts), and English capability (+7.6 pts). Beyond the core LLM, Fanar 2.0 introduces a rich stack of new capabilities. FanarGuard is a state-of-the-art 4B bilingual moderation filter for Arabic safety and cultural alignment. The speech family Aura gains a long-form ASR model for hours-long audio. Oryx vision family adds Arabic-aware image and video understanding alongside culturally grounded image generation. An agentic tool-calling framework enables multi-step workflows. Fanar-Sadiq utilizes a multi-agent architecture for Islamic content. Fanar-Diwan provides classical Arabic poetry generation. FanarShaheen delivers LLM-powered bilingual translation. A redesigned multi-layer orchestrator coordinates all components through intent-aware routing and defense-in-depth safety validation. Taken together, Fanar 2.0 demonstrates that sovereign, resource-constrained AI development can produce systems competitive with those built at far greater scale.

Fahim Dalvi, Nadir Durrani, Hassan Sajjad et al.

The proliferation of deep neural networks in various domains has seen an increased need for the interpretability of these models, especially in scenarios where fairness and trust are as important as model performance. A lot of independent work is being carried out to: i) analyze what linguistic and non-linguistic knowledge is learned within these models, and ii) highlight the salient parts of the input. We present NxPlain, a web application that provides an explanation of a model's prediction using latent concepts. NxPlain discovers latent concepts learned in a deep NLP model, provides an interpretation of the knowledge learned in the model, and explains its predictions based on the used concepts. The application allows users to browse through the latent concepts in an intuitive order, letting them efficiently scan through the most salient concepts with a global corpus level view and a local sentence-level view. Our tool is useful for debugging, unraveling model bias, and for highlighting spurious correlations in a model. A hosted demo is available here: https://nxplain.qcri.org.

Ahmed Abdelali, Nadir Durrani, Fahim Dalvi et al.

Arabic is a Semitic language which is widely spoken with many dialects. Given the success of pre-trained language models, many transformer models trained on Arabic and its dialects have surfaced. While there have been an extrinsic evaluation of these models with respect to downstream NLP tasks, no work has been carried out to analyze and compare their internal representations. We probe how linguistic information is encoded in the transformer models, trained on different Arabic dialects. We perform a layer and neuron analysis on the models using morphological tagging tasks for different dialects of Arabic and a dialectal identification task. Our analysis enlightens interesting findings such as: i) word morphology is learned at the lower and middle layers, ii) while syntactic dependencies are predominantly captured at the higher layers, iii) despite a large overlap in their vocabulary, the MSA-based models fail to capture the nuances of Arabic dialects, iv) we found that neurons in embedding layers are polysemous in nature, while the neurons in middle layers are exclusive to specific properties

Basel Mousi, Fahim Dalvi, Shammur Chowdhury et al.

Vision-language models (VLMs) can achieve high accuracy while still accepting culturally plausible but visually incorrect interpretations. Existing hallucination benchmarks rarely test this failure mode, particularly outside Western contexts and English. We introduce M2CQA, a culturally grounded multimodal benchmark built from images spanning 17 MENA countries, paired with contrastive true and counterfactual statements in English, Arabic, and its dialects. To isolate hallucination beyond raw accuracy, we propose the CounterFactual Hallucination Rate (CFHR), which measures counterfactual acceptance conditioned on correctly answering the true statement. Evaluating state-of-the-art VLMs under multiple prompting strategies, we find that CFHR rises sharply in Arabic, especially in dialects, even when true-statement accuracy remains high. Moreover, reasoning-first prompting consistently increases counterfactual hallucination, while answering before justifying improves robustness. We will make the experimental resources and dataset publicly available for the community.

Basel Mousi, Nadir Durrani, Fahim Dalvi et al.

Despite their remarkable ability to capture linguistic nuances across diverse languages, questions persist regarding the degree of alignment between languages in multilingual embeddings. Drawing inspiration from research on high-dimensional representations in neural language models, we employ clustering to uncover latent concepts within multilingual models. Our analysis focuses on quantifying the \textit{alignment} and \textit{overlap} of these concepts across various languages within the latent space. To this end, we introduce two metrics \CA{} and \CO{} aimed at quantifying these aspects, enabling a deeper exploration of multilingual embeddings. Our study encompasses three multilingual models (\texttt{mT5}, \texttt{mBERT}, and \texttt{XLM-R}) and three downstream tasks (Machine Translation, Named Entity Recognition, and Sentiment Analysis). Key findings from our analysis include: i) deeper layers in the network demonstrate increased cross-lingual \textit{alignment} due to the presence of language-agnostic concepts, ii) fine-tuning of the models enhances \textit{alignment} within the latent space, and iii) such task-specific calibration helps in explaining the emergence of zero-shot capabilities in the models.\footnote{The code is available at \url{https://github.com/baselmousi/multilingual-latent-concepts}}

Firoj Alam, Ali Ezzat Shahroor, Md. Arid Hasan et al. · utoronto

Large-scale multimodal models achieve strong results on tasks like Visual Question Answering (VQA), but they often fail when queries require culturally grounded, everyday knowledge, particularly in low-resource and underrepresented languages. To bridge this gap, we introduce Everyday Multimodal and Multilingual QA (EverydayMMQA), a framework for creating large-scale, culturally-grounded datasets for spoken and visual question answering (SVQA). Using this framework, we developed OASIS, a multimodal dataset integrating speech, images, and text. With over ~0.92M images and 14.8M QA pairs, OASIS contains 3.7M spoken questions, enabling four unique input combinations: speech-only, text-only, speech+image, and text+image. Focused on English and Arabic varieties, 18 countries, the dataset content is curated to reflect diverse, real-world situations. OASIS tests models on tasks beyond object recognition that involve pragmatic, commonsense, and culturally aware reasoning. We benchmarked four closed-source models, three open-source models, and one fine-tuned model. EverydayMMQA and OASIS together provide a benchmark and training dataset for building multimodal LLMs for a comprehensive set of everyday tasks within cultural contexts. The framework and dataset will be made publicly available to the community.

Fahim Dalvi, Hassan Sajjad, Nadir Durrani

Neuron analysis provides insights into how knowledge is structured in representations and discovers the role of neurons in the network. In addition to developing an understanding of our models, neuron analysis enables various applications such as debiasing, domain adaptation and architectural search. We present NeuroX, a comprehensive open-source toolkit to conduct neuron analysis of natural language processing models. It implements various interpretation methods under a unified API, and provides a framework for data processing and evaluation, thus making it easier for researchers and practitioners to perform neuron analysis. The Python toolkit is available at https://www.github.com/fdalvi/NeuroX. Demo Video available at https://youtu.be/mLhs2YMx4u8.

Fanar Team, Ummar Abbas, Mohammad Shahmeer Ahmad et al.

We present Fanar, a platform for Arabic-centric multimodal generative AI systems, that supports language, speech and image generation tasks. At the heart of Fanar are Fanar Star and Fanar Prime, two highly capable Arabic Large Language Models (LLMs) that are best in the class on well established benchmarks for similar sized models. Fanar Star is a 7B (billion) parameter model that was trained from scratch on nearly 1 trillion clean and deduplicated Arabic, English and Code tokens. Fanar Prime is a 9B parameter model continually trained on the Gemma-2 9B base model on the same 1 trillion token set. Both models are concurrently deployed and designed to address different types of prompts transparently routed through a custom-built orchestrator. The Fanar platform provides many other capabilities including a customized Islamic Retrieval Augmented Generation (RAG) system for handling religious prompts, a Recency RAG for summarizing information about current or recent events that have occurred after the pre-training data cut-off date. The platform provides additional cognitive capabilities including in-house bilingual speech recognition that supports multiple Arabic dialects, voice and image generation that is fine-tuned to better reflect regional characteristics. Finally, Fanar provides an attribution service that can be used to verify the authenticity of fact based generated content. The design, development, and implementation of Fanar was entirely undertaken at Hamad Bin Khalifa University's Qatar Computing Research Institute (QCRI) and was sponsored by Qatar's Ministry of Communications and Information Technology to enable sovereign AI technology development.

Xuemin Yu, Fahim Dalvi, Nadir Durrani et al.

Interpreting and understanding the predictions made by deep learning models poses a formidable challenge due to their inherently opaque nature. Many previous efforts aimed at explaining these predictions rely on input features, specifically, the words within NLP models. However, such explanations are often less informative due to the discrete nature of these words and their lack of contextual verbosity. To address this limitation, we introduce the Latent Concept Attribution method (LACOAT), which generates explanations for predictions based on latent concepts. Our foundational intuition is that a word can exhibit multiple facets, contingent upon the context in which it is used. Therefore, given a word in context, the latent space derived from our training process reflects a specific facet of that word. LACOAT functions by mapping the representations of salient input words into the training latent space, allowing it to provide latent context-based explanations of the prediction.

Basel Mousi, Nadir Durrani, Fahim Dalvi

While Knowledge Editing (KE) has been widely explored in English, its behavior in morphologically rich languages like Arabic remains underexamined. In this work, we present the first study of Arabic KE. We evaluate four methods (ROME, MEMIT, ICE, and LTE) on Arabic translations of the ZsRE and Counterfact benchmarks, analyzing both multilingual and cross-lingual settings. Our experiments on Llama-2-7B-chat show that parameter-based methods struggle with cross-lingual generalization, while instruction-tuned methods perform more robustly. We extend Learning-To-Edit (LTE) to a multilingual setting and show that joint Arabic-English training improves both editability and transfer. We release Arabic KE benchmarks and multilingual training for LTE data to support future research.

Asım Ersoy, Basel Mousi, Shammur Chowdhury et al.

The emergence of large language models (LLMs) has demonstrated that systems trained solely on text can acquire extensive world knowledge, develop reasoning capabilities, and internalize abstract semantic concepts--showcasing properties that can be associated with general intelligence. This raises an intriguing question: Do such concepts emerge in models trained on other modalities, such as speech? Furthermore, when models are trained jointly on multiple modalities: Do they develop a richer, more structured semantic understanding? To explore this, we analyze the conceptual structures learned by speech and textual models both individually and jointly. We employ Latent Concept Analysis, an unsupervised method for uncovering and interpreting latent representations in neural networks, to examine how semantic abstractions form across modalities. For reproducibility we made scripts and other resources available to the community.

Nadir Durrani, Basel Mousi, Fahim Dalvi

While Knowledge Editing has been extensively studied in monolingual settings, it remains underexplored in multilingual contexts. This survey systematizes recent research on Multilingual Knowledge Editing (MKE), a growing subdomain of model editing focused on ensuring factual edits generalize reliably across languages. We present a comprehensive taxonomy of MKE methods, covering parameter-based, memory-based, fine-tuning, and hypernetwork approaches. We survey available benchmarks,summarize key findings on method effectiveness and transfer patterns, identify challenges in cross-lingual propagation, and highlight open problems related to language anisotropy, evaluation coverage, and edit scalability. Our analysis consolidates a rapidly evolving area and lays the groundwork for future progress in editable language-aware LLMs.

Sabri Boughorbel, Fahim Dalvi, Nadir Durrani et al.

As fine-tuning becomes the dominant paradigm for improving large language models (LLMs), understanding what changes during this process is increasingly important. Traditional benchmarking often fails to explain why one model outperforms another. In this work, we use model diffing, a mechanistic interpretability approach, to analyze the specific capability differences between Gemma-2-9b-it and a SimPO-enhanced variant. Using crosscoders, we identify and categorize latent representations that differentiate the two models. We find that SimPO acquired latent concepts predominantly enhance safety mechanisms (+32.8%), multilingual capabilities (+43.8%), and instruction-following (+151.7%), while its additional training also reduces emphasis on model self-reference (-44.1%) and hallucination management (-68.5%). Our analysis shows that model diffing can yield fine-grained insights beyond leaderboard metrics, attributing performance gaps to concrete mechanistic capabilities. This approach offers a transparent and targeted framework for comparing LLMs.

Ahmed Abdelali, Hamdy Mubarak, Shammur Absar Chowdhury et al.

Recent advancements in Large Language Models (LLMs) have significantly influenced the landscape of language and speech research. Despite this progress, these models lack specific benchmarking against state-of-the-art (SOTA) models tailored to particular languages and tasks. LAraBench addresses this gap for Arabic Natural Language Processing (NLP) and Speech Processing tasks, including sequence tagging and content classification across different domains. We utilized models such as GPT-3.5-turbo, GPT-4, BLOOMZ, Jais-13b-chat, Whisper, and USM, employing zero and few-shot learning techniques to tackle 33 distinct tasks across 61 publicly available datasets. This involved 98 experimental setups, encompassing ~296K data points, ~46 hours of speech, and 30 sentences for Text-to-Speech (TTS). This effort resulted in 330+ sets of experiments. Our analysis focused on measuring the performance gap between SOTA models and LLMs. The overarching trend observed was that SOTA models generally outperformed LLMs in zero-shot learning, with a few exceptions. Notably, larger computational models with few-shot learning techniques managed to reduce these performance gaps. Our findings provide valuable insights into the applicability of LLMs for Arabic NLP and speech processing tasks.

Basel Mousi, Nadir Durrani, Fahim Dalvi

Work done to uncover the knowledge encoded within pre-trained language models rely on annotated corpora or human-in-the-loop methods. However, these approaches are limited in terms of scalability and the scope of interpretation. We propose using a large language model, ChatGPT, as an annotator to enable fine-grained interpretation analysis of pre-trained language models. We discover latent concepts within pre-trained language models by applying agglomerative hierarchical clustering over contextualized representations and then annotate these concepts using ChatGPT. Our findings demonstrate that ChatGPT produces accurate and semantically richer annotations compared to human-annotated concepts. Additionally, we showcase how GPT-based annotations empower interpretation analysis methodologies of which we demonstrate two: probing frameworks and neuron interpretation. To facilitate further exploration and experimentation in the field, we make available a substantial ConceptNet dataset (TCN) comprising 39,000 annotated concepts.

Ahmed Abdelali, Nadir Durrani, Fahim Dalvi et al.

Arabic is a Semitic language which is widely spoken with many dialects. Given the success of pre-trained language models, many transformer models trained on Arabic and its dialects have surfaced. While these models have been compared with respect to downstream NLP tasks, no evaluation has been carried out to directly compare the internal representations. We probe how linguistic information is encoded in Arabic pretrained models, trained on different varieties of Arabic language. We perform a layer and neuron analysis on the models using three intrinsic tasks: two morphological tagging tasks based on MSA (modern standard Arabic) and dialectal POS-tagging and a dialectal identification task. Our analysis enlightens interesting findings such as: i) word morphology is learned at the lower and middle layers ii) dialectal identification necessitate more knowledge and hence preserved even in the final layers, iii) despite a large overlap in their vocabulary, the MSA-based models fail to capture the nuances of Arabic dialects, iv) we found that neurons in embedding layers are polysemous in nature, while the neurons in middle layers are exclusive to specific properties.

Hassan Sajjad, Nadir Durrani, Fahim Dalvi

The proliferation of deep neural networks in various domains has seen an increased need for interpretability of these models. Preliminary work done along this line and papers that surveyed such, are focused on high-level representation analysis. However, a recent branch of work has concentrated on interpretability at a more granular level of analyzing neurons within these models. In this paper, we survey the work done on neuron analysis including: i) methods to discover and understand neurons in a network, ii) evaluation methods, iii) major findings including cross architectural comparisons that neuron analysis has unraveled, iv) applications of neuron probing such as: controlling the model, domain adaptation etc., and v) a discussion on open issues and future research directions.

Nadir Durrani, Hassan Sajjad, Fahim Dalvi

Transfer learning from pre-trained neural language models towards downstream tasks has been a predominant theme in NLP recently. Several researchers have shown that deep NLP models learn non-trivial amount of linguistic knowledge, captured at different layers of the model. We investigate how fine-tuning towards downstream NLP tasks impacts the learned linguistic knowledge. We carry out a study across popular pre-trained models BERT, RoBERTa and XLNet using layer and neuron-level diagnostic classifiers. We found that for some GLUE tasks, the network relies on the core linguistic information and preserve it deeper in the network, while for others it forgets. Linguistic information is distributed in the pre-trained language models but becomes localized to the lower layers post fine-tuning, reserving higher layers for the task specific knowledge. The pattern varies across architectures, with BERT retaining linguistic information relatively deeper in the network compared to RoBERTa and XLNet, where it is predominantly delegated to the lower layers.

Hassan Sajjad, Narine Kokhlikyan, Fahim Dalvi et al.

This paper is a write-up for the tutorial on "Fine-grained Interpretation and Causation Analysis in Deep NLP Models" that we are presenting at NAACL 2021. We present and discuss the research work on interpreting fine-grained components of a model from two perspectives, i) fine-grained interpretation, ii) causation analysis. The former introduces methods to analyze individual neurons and a group of neurons with respect to a language property or a task. The latter studies the role of neurons and input features in explaining decisions made by the model. We also discuss application of neuron analysis such as network manipulation and domain adaptation. Moreover, we present two toolkits namely NeuroX and Captum, that support functionalities discussed in this tutorial.

Hassan Sajjad, Firoj Alam, Fahim Dalvi et al.

Post-processing of static embedding has beenshown to improve their performance on both lexical and sequence-level tasks. However, post-processing for contextualized embeddings is an under-studied problem. In this work, we question the usefulness of post-processing for contextualized embeddings obtained from different layers of pre-trained language models. More specifically, we standardize individual neuron activations using z-score, min-max normalization, and by removing top principle components using the all-but-the-top method. Additionally, we apply unit length normalization to word representations. On a diverse set of pre-trained models, we show that post-processing unwraps vital information present in the representations for both lexical tasks (such as word similarity and analogy)and sequence classification tasks. Our findings raise interesting points in relation to theresearch studies that use contextualized representations, and suggest z-score normalization as an essential step to consider when using them in an application.

Nadir Durrani, Hassan Sajjad, Fahim Dalvi et al.

While a lot of analysis has been carried to demonstrate linguistic knowledge captured by the representations learned within deep NLP models, very little attention has been paid towards individual neurons.We carry outa neuron-level analysis using core linguistic tasks of predicting morphology, syntax and semantics, on pre-trained language models, with questions like: i) do individual neurons in pre-trained models capture linguistic information? ii) which parts of the network learn more about certain linguistic phenomena? iii) how distributed or focused is the information? and iv) how do various architectures differ in learning these properties? We found small subsets of neurons to predict linguistic tasks, with lower level tasks (such as morphology) localized in fewer neurons, compared to higher level task of predicting syntax. Our study also reveals interesting cross architectural comparisons. For example, we found neurons in XLNet to be more localized and disjoint when predicting properties compared to BERT and others, where they are more distributed and coupled.

Firoj Alam, Fahim Dalvi, Shaden Shaar et al.

With the outbreak of the COVID-19 pandemic, people turned to social media to read and to share timely information including statistics, warnings, advice, and inspirational stories. Unfortunately, alongside all this useful information, there was also a new blending of medical and political misinformation and disinformation, which gave rise to the first global infodemic. While fighting this infodemic is typically thought of in terms of factuality, the problem is much broader as malicious content includes not only fake news, rumors, and conspiracy theories, but also promotion of fake cures, panic, racism, xenophobia, and mistrust in the authorities, among others. This is a complex problem that needs a holistic approach combining the perspectives of journalists, fact-checkers, policymakers, government entities, social media platforms, and society as a whole. Taking them into account we define an annotation schema and detailed annotation instructions, which reflect these perspectives. We performed initial annotations using this schema, and our initial experiments demonstrated sizable improvements over the baselines. Now, we issue a call to arms to the research community and beyond to join the fight by supporting our crowdsourcing annotation efforts.

John M. Wu, Yonatan Belinkov, Hassan Sajjad et al.

This paper investigates contextual word representation models from the lens of similarity analysis. Given a collection of trained models, we measure the similarity of their internal representations and attention. Critically, these models come from vastly different architectures. We use existing and novel similarity measures that aim to gauge the level of localization of information in the deep models, and facilitate the investigation of which design factors affect model similarity, without requiring any external linguistic annotation. The analysis reveals that models within the same family are more similar to one another, as may be expected. Surprisingly, different architectures have rather similar representations, but different individual neurons. We also observed differences in information localization in lower and higher layers and found that higher layers are more affected by fine-tuning on downstream tasks.

Firoj Alam, Shaden Shaar, Fahim Dalvi et al.

With the emergence of the COVID-19 pandemic, the political and the medical aspects of disinformation merged as the problem got elevated to a whole new level to become the first global infodemic. Fighting this infodemic has been declared one of the most important focus areas of the World Health Organization, with dangers ranging from promoting fake cures, rumors, and conspiracy theories to spreading xenophobia and panic. Addressing the issue requires solving a number of challenging problems such as identifying messages containing claims, determining their check-worthiness and factuality, and their potential to do harm as well as the nature of that harm, to mention just a few. To address this gap, we release a large dataset of 16K manually annotated tweets for fine-grained disinformation analysis that (i) focuses on COVID-19, (ii) combines the perspectives and the interests of journalists, fact-checkers, social media platforms, policy makers, and society, and (iii) covers Arabic, Bulgarian, Dutch, and English. Finally, we show strong evaluation results using pretrained Transformers, thus confirming the practical utility of the dataset in monolingual vs. multilingual, and single task vs. multitask settings.

Fahim Dalvi, Hassan Sajjad, Nadir Durrani et al.

Transformer-based deep NLP models are trained using hundreds of millions of parameters, limiting their applicability in computationally constrained environments. In this paper, we study the cause of these limitations by defining a notion of Redundancy, which we categorize into two classes: General Redundancy and Task-specific Redundancy. We dissect two popular pretrained models, BERT and XLNet, studying how much redundancy they exhibit at a representation-level and at a more fine-grained neuron-level. Our analysis reveals interesting insights, such as: i) 85% of the neurons across the network are redundant and ii) at least 92% of them can be removed when optimizing towards a downstream task. Based on our analysis, we present an efficient feature-based transfer learning procedure, which maintains 97% performance while using at-most 10% of the original neurons.

Hassan Sajjad, Fahim Dalvi, Nadir Durrani et al.

Transformer-based NLP models are trained using hundreds of millions or even billions of parameters, limiting their applicability in computationally constrained environments. While the number of parameters generally correlates with performance, it is not clear whether the entire network is required for a downstream task. Motivated by the recent work on pruning and distilling pre-trained models, we explore strategies to drop layers in pre-trained models, and observe the effect of pruning on downstream GLUE tasks. We were able to prune BERT, RoBERTa and XLNet models up to 40%, while maintaining up to 98% of their original performance. Additionally we show that our pruned models are on par with those built using knowledge distillation, both in terms of size and performance. Our experiments yield interesting observations such as, (i) the lower layers are most critical to maintain downstream task performance, (ii) some tasks such as paraphrase detection and sentence similarity are more robust to the dropping of layers, and (iii) models trained using a different objective function exhibit different learning patterns and w.r.t the layer dropping.

Yonatan Belinkov, Nadir Durrani, Fahim Dalvi et al.

Despite the recent success of deep neural networks in natural language processing (NLP), their interpretability remains a challenge. We analyze the representations learned by neural machine translation models at various levels of granularity and evaluate their quality through relevant extrinsic properties. In particular, we seek answers to the following questions: (i) How accurately is word-structure captured within the learned representations, an important aspect in translating morphologically-rich languages? (ii) Do the representations capture long-range dependencies, and effectively handle syntactically divergent languages? (iii) Do the representations capture lexical semantics? We conduct a thorough investigation along several parameters: (i) Which layers in the architecture capture each of these linguistic phenomena; (ii) How does the choice of translation unit (word, character, or subword unit) impact the linguistic properties captured by the underlying representations? (iii) Do the encoder and decoder learn differently and independently? (iv) Do the representations learned by multilingual NMT models capture the same amount of linguistic information as their bilingual counterparts? Our data-driven, quantitative evaluation illuminates important aspects in NMT models and their ability to capture various linguistic phenomena. We show that deep NMT models learn a non-trivial amount of linguistic information. Notable findings include: i) Word morphology and part-of-speech information are captured at the lower layers of the model; (ii) In contrast, lexical semantics or non-local syntactic and semantic dependencies are better represented at the higher layers; (iii) Representations learned using characters are more informed about wordmorphology compared to those learned using subword units; and (iv) Representations learned by multilingual models are richer compared to bilingual models.

Fahim Dalvi, Avery Nortonsmith, D. Anthony Bau et al.

We present a toolkit to facilitate the interpretation and understanding of neural network models. The toolkit provides several methods to identify salient neurons with respect to the model itself or an external task. A user can visualize selected neurons, ablate them to measure their effect on the model accuracy, and manipulate them to control the behavior of the model at the test time. Such an analysis has a potential to serve as a springboard in various research directions, such as understanding the model, better architectural choices, model distillation and controlling data biases.

Fahim Dalvi, Nadir Durrani, Hassan Sajjad et al.

Despite the remarkable evolution of deep neural networks in natural language processing (NLP), their interpretability remains a challenge. Previous work largely focused on what these models learn at the representation level. We break this analysis down further and study individual dimensions (neurons) in the vector representation learned by end-to-end neural models in NLP tasks. We propose two methods: Linguistic Correlation Analysis, based on a supervised method to extract the most relevant neurons with respect to an extrinsic task, and Cross-model Correlation Analysis, an unsupervised method to extract salient neurons w.r.t. the model itself. We evaluate the effectiveness of our techniques by ablating the identified neurons and reevaluating the network's performance for two tasks: neural machine translation (NMT) and neural language modeling (NLM). We further present a comprehensive analysis of neurons with the aim to address the following questions: i) how localized or distributed are different linguistic properties in the models? ii) are certain neurons exclusive to some properties and not others? iii) is the information more or less distributed in NMT vs. NLM? and iv) how important are the neurons identified through the linguistic correlation method to the overall task? Our code is publicly available as part of the NeuroX toolkit (Dalvi et al. 2019).

Anthony Bau, Yonatan Belinkov, Hassan Sajjad et al.

Neural machine translation (NMT) models learn representations containing substantial linguistic information. However, it is not clear if such information is fully distributed or if some of it can be attributed to individual neurons. We develop unsupervised methods for discovering important neurons in NMT models. Our methods rely on the intuition that different models learn similar properties, and do not require any costly external supervision. We show experimentally that translation quality depends on the discovered neurons, and find that many of them capture common linguistic phenomena. Finally, we show how to control NMT translations in predictable ways, by modifying activations of individual neurons.

Fahim Dalvi, Nadir Durrani, Hassan Sajjad et al.

We address the problem of simultaneous translation by modifying the Neural MT decoder to operate with dynamically built encoder and attention. We propose a tunable agent which decides the best segmentation strategy for a user-defined BLEU loss and Average Proportion (AP) constraint. Our agent outperforms previously proposed Wait-if-diff and Wait-if-worse agents (Cho and Esipova, 2016) on BLEU with a lower latency. Secondly we proposed data-driven changes to Neural MT training to better match the incremental decoding framework.

Nadir Durrani, Fahim Dalvi

Bilingual sequence models improve phrase-based translation and reordering by overcoming phrasal independence assumption and handling long range reordering. However, due to data sparsity, these models often fall back to very small context sizes. This problem has been previously addressed by learning sequences over generalized representations such as POS tags or word clusters. In this paper, we explore an alternative based on neural network models. More concretely we train neuralized versions of lexicalized reordering and the operation sequence models using feed-forward neural network. Our results show improvements of up to 0.6 and 0.5 BLEU points on top of the baseline German->English and English->German systems. We also observed improvements compared to the systems that used POS tags and word clusters to train these models. Because we modify the bilingual corpus to integrate reordering operations, this allows us to also train a sequence-to-sequence neural MT model having explicit reordering triggers. Our motivation was to directly enable reordering information in the encoder-decoder framework, which otherwise relies solely on the attention model to handle long range reordering. We tried both coarser and fine-grained reordering operations. However, these experiments did not yield any improvements over the baseline Neural MT systems.

Yonatan Belinkov, Lluís Màrquez, Hassan Sajjad et al.

While neural machine translation (NMT) models provide improved translation quality in an elegant, end-to-end framework, it is less clear what they learn about language. Recent work has started evaluating the quality of vector representations learned by NMT models on morphological and syntactic tasks. In this paper, we investigate the representations learned at different layers of NMT encoders. We train NMT systems on parallel data and use the trained models to extract features for training a classifier on two tasks: part-of-speech and semantic tagging. We then measure the performance of the classifier as a proxy to the quality of the original NMT model for the given task. Our quantitative analysis yields interesting insights regarding representation learning in NMT models. For instance, we find that higher layers are better at learning semantics while lower layers tend to be better for part-of-speech tagging. We also observe little effect of the target language on source-side representations, especially with higher quality NMT models.

Hassan Sajjad, Fahim Dalvi, Nadir Durrani et al.

Word segmentation plays a pivotal role in improving any Arabic NLP application. Therefore, a lot of research has been spent in improving its accuracy. Off-the-shelf tools, however, are: i) complicated to use and ii) domain/dialect dependent. We explore three language-independent alternatives to morphological segmentation using: i) data-driven sub-word units, ii) characters as a unit of learning, and iii) word embeddings learned using a character CNN (Convolution Neural Network). On the tasks of Machine Translation and POS tagging, we found these methods to achieve close to, and occasionally surpass state-of-the-art performance. In our analysis, we show that a neural machine translation system is sensitive to the ratio of source and target tokens, and a ratio close to 1 or greater, gives optimal performance.

Hassan Sajjad, Nadir Durrani, Fahim Dalvi et al.

In this paper, we explore alternative ways to train a neural machine translation system in a multi-domain scenario. We investigate data concatenation (with fine tuning), model stacking (multi-level fine tuning), data selection and multi-model ensemble. Our findings show that the best translation quality can be achieved by building an initial system on a concatenation of available out-of-domain data and then fine-tuning it on in-domain data. Model stacking works best when training begins with the furthest out-of-domain data and the model is incrementally fine-tuned with the next furthest domain and so on. Data selection did not give the best results, but can be considered as a decent compromise between training time and translation quality. A weighted ensemble of different individual models performed better than data selection. It is beneficial in a scenario when there is no time for fine-tuning an already trained model.

Yonatan Belinkov, Nadir Durrani, Fahim Dalvi et al.

Neural machine translation (MT) models obtain state-of-the-art performance while maintaining a simple, end-to-end architecture. However, little is known about what these models learn about source and target languages during the training process. In this work, we analyze the representations learned by neural MT models at various levels of granularity and empirically evaluate the quality of the representations for learning morphology through extrinsic part-of-speech and morphological tagging tasks. We conduct a thorough investigation along several parameters: word-based vs. character-based representations, depth of the encoding layer, the identity of the target language, and encoder vs. decoder representations. Our data-driven, quantitative evaluation sheds light on important aspects in the neural MT system and its ability to capture word structure.

Nadir Durrani, Fahim Dalvi, Hassan Sajjad et al.

This paper describes QCRI's machine translation systems for the IWSLT 2016 evaluation campaign. We participated in the Arabic->English and English->Arabic tracks. We built both Phrase-based and Neural machine translation models, in an effort to probe whether the newly emerged NMT framework surpasses the traditional phrase-based systems in Arabic-English language pairs. We trained a very strong phrase-based system including, a big language model, the Operation Sequence Model, Neural Network Joint Model and Class-based models along with different domain adaptation techniques such as MML filtering, mixture modeling and using fine tuning over NNJM model. However, a Neural MT system, trained by stacking data from different genres through fine-tuning, and applying ensemble over 8 models, beat our very strong phrase-based system by a significant 2 BLEU points margin in Arabic->English direction. We did not obtain similar gains in the other direction but were still able to outperform the phrase-based system. We also applied system combination on phrase-based and NMT outputs.

Hardie Cate, Fahim Dalvi, Zeshan Hussain

We use CNNs to build a system that both classifies images of faces based on a variety of different facial attributes and generates new faces given a set of desired facial characteristics. After introducing the problem and providing context in the first section, we discuss recent work related to image generation in Section 2. In Section 3, we describe the methods used to fine-tune our CNN and generate new images using a novel approach inspired by a Gaussian mixture model. In Section 4, we discuss our working dataset and describe our preprocessing steps and handling of facial attributes. Finally, in Sections 5, 6 and 7, we explain our experiments and results and conclude in the following section. Our classification system has 82\% test accuracy. Furthermore, our generation pipeline successfully creates well-formed faces.

Hardie Cate, Fahim Dalvi, Zeshan Hussain

Devices like the Myo armband available in the market today enable us to collect data about the position of a user's hands and fingers over time. We can use these technologies for sign language translation since each sign is roughly a combination of gestures across time. In this work, we utilize a dataset collected by a group at the University of South Wales, which contains parameters, such as hand position, hand rotation, and finger bend, for 95 unique signs. For each input stream representing a sign, we predict which sign class this stream falls into. We begin by implementing baseline SVM and logistic regression models, which perform reasonably well on high quality data. Lower quality data requires a more sophisticated approach, so we explore different methods in temporal classification, including long short term memory architectures and sequential pattern mining methods.