Understanding how linguistic knowledge is encoded in language models is crucial for improving their generalisation capabilities. In this paper, we investigate the processing of morphosyntactic phenomena, by leveraging a recently proposed method for probing language models via Shapley Head Values (SHVs). Using the English language BLiMP dataset, we test our approach on two widely used models, BERT and RoBERTa, and compare how linguistic constructions such as anaphor agreement and filler-gap dependencies are handled. Through quantitative pruning and qualitative clustering analysis, we demonstrate that attention heads responsible for processing related linguistic phenomena cluster together. Our results show that SHV-based attributions reveal distinct patterns across both models, providing insights into how language models organize and process linguistic information. These findings support the hypothesis that language models learn subnetworks corresponding to linguistic theory, with potential implications for cross-linguistic model analysis and interpretability in Natural Language Processing (NLP).
Large Language Models (LLMs) often encounter conflicts between their learned, internal (parametric knowledge, PK) and external knowledge provided during inference (contextual knowledge, CK). Understanding how LLMs models prioritize one knowledge source over the other remains a challenge. In this paper, we propose a novel probing framework to explore the mechanisms governing the selection between PK and CK in LLMs. Using controlled prompts designed to contradict the model's PK, we demonstrate that specific model activations are indicative of the knowledge source employed. We evaluate this framework on various LLMs of different sizes and demonstrate that mid-layer activations, particularly those related to relations in the input, are crucial in predicting knowledge source selection, paving the way for more reliable models capable of handling knowledge conflicts effectively.
Large Language Models (LLMs) have shown their impressive capabilities, while also raising concerns about the data contamination problems due to privacy issues and leakage of benchmark datasets in the pre-training phase. Therefore, it is vital to detect the contamination by checking whether an LLM has been pre-trained on the target texts. Recent studies focus on the generated texts and compute perplexities, which are superficial features and not reliable. In this study, we propose to utilize the probing technique for pre-training data detection by examining the model's internal activations. Our method is simple and effective and leads to more trustworthy pre-training data detection. Additionally, we propose ArxivMIA, a new challenging benchmark comprising arxiv abstracts from Computer Science and Mathematics categories. Our experiments demonstrate that our method outperforms all baselines, and achieves state-of-the-art performance on both WikiMIA and ArxivMIA, with additional experiments confirming its efficacy (Our code and dataset are available at https://github.com/zhliu0106/probing-lm-data).
Large Language Models (LLMs) demonstrate an impressive capacity to recall a vast range of factual knowledge. However, understanding their underlying reasoning and internal mechanisms in exploiting this knowledge remains a key research area. This work unveils the factual information an LLM represents internally for sentence-level claim verification. We propose an end-to-end framework to decode factual knowledge embedded in token representations from a vector space to a set of ground predicates, showing its layer-wise evolution using a dynamic knowledge graph. Our framework employs activation patching, a vector-level technique that alters a token representation during inference, to extract encoded knowledge. Accordingly, we neither rely on training nor external models. Using factual and common-sense claims from two claim verification datasets, we showcase interpretability analyses at local and global levels. The local analysis highlights entity centrality in LLM reasoning, from claim-related information and multi-hop reasoning to representation errors causing erroneous evaluation. On the other hand, the global reveals trends in the underlying evolution, such as word-based knowledge evolving into claim-related facts. By interpreting semantics from LLM latent representations and enabling graph-related analyses, this work enhances the understanding of the factual knowledge resolution process.
The rise of Large Language Models (LLMs) has affected various disciplines that got beyond mere text generation. Going beyond their textual nature, this project proposal aims to investigate the interaction between LLMs and non-verbal communication, specifically focusing on gestures. The proposal sets out a plan to examine the proficiency of LLMs in deciphering both explicit and implicit non-verbal cues within textual prompts and their ability to associate these gestures with various contextual factors. The research proposes to test established psycholinguistic study designs to construct a comprehensive dataset that pairs textual prompts with detailed gesture descriptions, encompassing diverse regional variations, and semantic labels. To assess LLMs' comprehension of gestures, experiments are planned, evaluating their ability to simulate human behaviour in order to replicate psycholinguistic experiments. These experiments consider cultural dimensions and measure the agreement between LLM-identified gestures and the dataset, shedding light on the models' contextual interpretation of non-verbal cues (e.g. gestures).
While the impact of social biases in language models has been recognized, prior methods for bias evaluation have been limited to binary association tests on small datasets, limiting our understanding of bias complexities. This paper proposes a novel framework for probing language models for social biases by assessing disparate treatment, which involves treating individuals differently according to their affiliation with a sensitive demographic group. We curate SoFa, a large-scale benchmark designed to address the limitations of existing fairness collections. SoFa expands the analysis beyond the binary comparison of stereotypical versus anti-stereotypical identities to include a diverse range of identities and stereotypes. Comparing our methodology with existing benchmarks, we reveal that biases within language models are more nuanced than acknowledged, indicating a broader scope of encoded biases than previously recognized. Benchmarking LMs on SoFa, we expose how identities expressing different religions lead to the most pronounced disparate treatments across all models. Finally, our findings indicate that real-life adversities faced by various groups such as women and people with disabilities are mirrored in the behavior of these models.
The probing classifiers framework has been employed for interpreting deep neural network models for a variety of natural language processing (NLP) applications. Studies, however, have largely focused on sentencelevel NLP tasks. This work is the first to apply the probing paradigm to representations learned for document-level information extraction (IE). We designed eight embedding probes to analyze surface, semantic, and event-understanding capabilities relevant to document-level event extraction. We apply them to the representations acquired by learning models from three different LLM-based document-level IE approaches on a standard dataset. We found that trained encoders from these models yield embeddings that can modestly improve argument detections and labeling but only slightly enhance event-level tasks, albeit trade-offs in information helpful for coherence and event-type prediction. We further found that encoder models struggle with document length and cross-sentence discourse.
Language models, such as GPT-3.5 and ChatGPT, demonstrate remarkable abilities to follow diverse human instructions and perform a wide range of tasks. However, when probing language models using a range of basic table-understanding tasks, we observe that today's language models are still sub-optimal in many table-related tasks, likely because they are pre-trained predominantly on \emph{one-dimensional} natural-language texts, whereas relational tables are \emph{two-dimensional} objects. In this work, we propose a new "\emph{table-tuning}" paradigm, where we continue to train/fine-tune language models like GPT-3.5 and ChatGPT, using diverse table-tasks synthesized from real tables as training data, with the goal of enhancing language models' ability to understand tables and perform table tasks. We show that our resulting Table-GPT models demonstrate (1) better \emph{table-understanding} capabilities, by consistently outperforming the vanilla GPT-3.5 and ChatGPT, on a wide-range of table tasks, including holdout unseen tasks, and (2) strong \emph{generalizability}, in its ability to respond to diverse human instructions to perform new table-tasks, in a manner similar to GPT-3.5 and ChatGPT.
Large Language Models (LLMs) have emerged as dominant foundational models in modern NLP. However, the understanding of their prediction processes and internal mechanisms, such as feed-forward networks (FFN) and multi-head self-attention (MHSA), remains largely unexplored. In this work, we probe LLMs from a human behavioral perspective, correlating values from LLMs with eye-tracking measures, which are widely recognized as meaningful indicators of human reading patterns. Our findings reveal that LLMs exhibit a similar prediction pattern with humans but distinct from that of Shallow Language Models (SLMs). Moreover, with the escalation of LLM layers from the middle layers, the correlation coefficients also increase in FFN and MHSA, indicating that the logits within FFN increasingly encapsulate word semantics suitable for predicting tokens from the vocabulary.
Large language models (LLMs) have been proven capable of memorizing their training data, which can be extracted through specifically designed prompts. As the scale of datasets continues to grow, privacy risks arising from memorization have attracted increasing attention. Quantifying language model memorization helps evaluate potential privacy risks. However, prior works on quantifying memorization require access to the precise original data or incur substantial computational overhead, making it difficult for applications in real-world language models. To this end, we propose a fine-grained, entity-level definition to quantify memorization with conditions and metrics closer to real-world scenarios. In addition, we also present an approach for efficiently extracting sensitive entities from autoregressive language models. We conduct extensive experiments based on the proposed, probing language models' ability to reconstruct sensitive entities under different settings. We find that language models have strong memorization at the entity level and are able to reproduce the training data even with partial leakages. The results demonstrate that LLMs not only memorize their training data but also understand associations between entities. These findings necessitate that trainers of LLMs exercise greater prudence regarding model memorization, adopting memorization mitigation techniques to preclude privacy violations.
Public opinion reflects and shapes societal behavior, but the traditional survey-based tools to measure it are limited. We introduce a novel approach to probe media diet models -- language models adapted to online news, TV broadcast, or radio show content -- that can emulate the opinions of subpopulations that have consumed a set of media. To validate this method, we use as ground truth the opinions expressed in U.S. nationally representative surveys on COVID-19 and consumer confidence. Our studies indicate that this approach is (1) predictive of human judgements found in survey response distributions and robust to phrasing and channels of media exposure, (2) more accurate at modeling people who follow media more closely, and (3) aligned with literature on which types of opinions are affected by media consumption. Probing language models provides a powerful new method for investigating media effects, has practical applications in supplementing polls and forecasting public opinion, and suggests a need for further study of the surprising fidelity with which neural language models can predict human responses.
Linguistic analysis of language models is one of the ways to explain and describe their reasoning, weaknesses, and limitations. In the probing part of the model interpretability research, studies concern individual languages as well as individual linguistic structures. The question arises: are the detected regularities linguistically coherent, or on the contrary, do they dissonate at the typological scale? Moreover, the majority of studies address the inherent set of languages and linguistic structures, leaving the actual typological diversity knowledge out of scope. In this paper, we present and apply the GUI-assisted framework allowing us to easily probe a massive number of languages for all the morphosyntactic features present in the Universal Dependencies data. We show that reflecting the anglo-centric trend in NLP over the past years, most of the regularities revealed in the mBERT model are typical for the western-European languages. Our framework can be integrated with the existing probing toolboxes, model cards, and leaderboards, allowing practitioners to use and share their standard probing methods to interpret multilingual models. Thus we propose a toolkit to systematize the multilingual flaws in multilingual models, providing a reproducible experimental setup for 104 languages and 80 morphosyntactic features. https://github.com/AIRI-Institute/Probing_framework
We introduce a new in-context learning paradigm to measure Large Language Models' (LLMs) ability to learn novel words during inference. In particular, we rewrite Winograd-style co-reference resolution problems by replacing the key concept word with a synthetic but plausible word that the model must understand to complete the task. Solving this task requires the model to make use of the dictionary definition of the new word given in the prompt. This benchmark addresses word acquisition, one important aspect of the diachronic degradation known to afflict LLMs. As LLMs are frozen in time at the moment they are trained, they are normally unable to reflect the way language changes over time. We show that the accuracy of LLMs compared to the original Winograd tasks decreases radically in our benchmark, thus identifying a limitation of current models and providing a benchmark to measure future improvements in LLMs ability to do in-context learning.
Language models (LMs) trained on raw texts have no direct access to the physical world. Gordon and Van Durme (2013) point out that LMs can thus suffer from reporting bias: texts rarely report on common facts, instead focusing on the unusual aspects of a situation. If LMs are only trained on text corpora and naively memorise local co-occurrence statistics, they thus naturally would learn a biased view of the physical world. While prior studies have repeatedly verified that LMs of smaller scales (e.g., RoBERTa, GPT-2) amplify reporting bias, it remains unknown whether such trends continue when models are scaled up. We investigate reporting bias from the perspective of colour in larger language models (LLMs) such as PaLM and GPT-3. Specifically, we query LLMs for the typical colour of objects, which is one simple type of perceptually grounded physical common sense. Surprisingly, we find that LLMs significantly outperform smaller LMs in determining an object's typical colour and more closely track human judgments, instead of overfitting to surface patterns stored in texts. This suggests that very large models of language alone are able to overcome certain types of reporting bias that are characterized by local co-occurrences.
In this paper, we set out to quantify the syntactic capacity of BERT in the evaluation regime of non-context free patterns, as occurring in Dutch. We devise a test suite based on a mildly context-sensitive formalism, from which we derive grammars that capture the linguistic phenomena of control verb nesting and verb raising. The grammars, paired with a small lexicon, provide us with a large collection of naturalistic utterances, annotated with verb-subject pairings, that serve as the evaluation test bed for an attention-based span selection probe. Our results, backed by extensive analysis, suggest that the models investigated fail in the implicit acquisition of the dependencies examined.
We present three Natural Language Inference (NLI) challenge sets that can evaluate NLI models on their understanding of temporal expressions. More specifically, we probe these models for three temporal properties: (a) the order between points in time, (b) the duration between two points in time, (c) the relation between the magnitude of times specified in different units. We find that although large language models fine-tuned on MNLI have some basic perception of the order between points in time, at large, these models do not have a thorough understanding of the relation between temporal expressions.
Recent research has demonstrated that large pre-trained language models reflect societal biases expressed in natural language. The present paper introduces a simple method for probing language models to conduct a multilingual study of gender bias towards politicians. We quantify the usage of adjectives and verbs generated by language models surrounding the names of politicians as a function of their gender. To this end, we curate a dataset of 250k politicians worldwide, including their names and gender. Our study is conducted in seven languages across six different language modeling architectures. The results demonstrate that pre-trained language models' stance towards politicians varies strongly across analyzed languages. We find that while some words such as dead, and designated are associated with both male and female politicians, a few specific words such as beautiful and divorced are predominantly associated with female politicians. Finally, and contrary to previous findings, our study suggests that larger language models do not tend to be significantly more gender-biased than smaller ones.