Stereotypical bias encoded in language models (LMs) poses a threat to safe language technology, yet our understanding of how bias manifests in the parameters of LMs remains incomplete. We introduce local contrastive editing that enables the localization and editing of a subset of weights in a target model in relation to a reference model. We deploy this approach to identify and modify subsets of weights that are associated with gender stereotypes in LMs. Through a series of experiments we demonstrate that local contrastive editing can precisely localize and control a small subset (< 0.5%) of weights that encode gender bias. Our work (i) advances our understanding of how stereotypical biases can manifest in the parameter space of LMs and (ii) opens up new avenues for developing parameter-efficient strategies for controlling model properties in a contrastive manner.

Recent work has proposed explicitly inducing language-wise modularity in multilingual LMs via sparse fine-tuning (SFT) on per-language subnetworks as a means of better guiding cross-lingual sharing. In this paper, we investigate (1) the degree to which language-wise modularity *naturally* arises within models with no special modularity interventions, and (2) how cross-lingual sharing and interference differ between such models and those with explicit SFT-guided subnetwork modularity. In order to do so, we use XLM-R as our multilingual LM. Moreover, to quantify language specialization and cross-lingual interaction, we use a Training Data Attribution method that estimates the degree to which a model’s predictions are influenced by in-language or cross-language training examples. Our results show that language-specialized subnetworks do naturally arise, and that SFT, rather than always increasing modularity, can decrease language specialization of subnetworks in favor of more cross-lingual sharing.

Metaphors in natural language are a reflection of fundamental cognitive processes such as analogical reasoning and categorisation, and are deeply rooted in everyday communication. Metaphor understanding is therefore an essential task for large language models (LLMs). We release the Metaphor Understanding Challenge Dataset (MUNCH), designed to evaluate the metaphor understanding capabilities of LLMs. The dataset provides over 10k paraphrases for sentences containing metaphor use, as well as 1.5k instances containing inapt paraphrases. The inapt paraphrases were carefully selected to serve as control to determine whether the model indeed performs full metaphor interpretation or rather resorts to lexical similarity. All apt and inapt paraphrases were manually annotated. The metaphorical sentences cover natural metaphor uses across 4 genres (academic, news, fiction, and conversation), and they exhibit different levels of novelty. Experiments with LLaMA and GPT-3.5 demonstrate that MUNCH presents a challenging task for LLMs. The dataset is freely accessible at https://github.com/xiaoyuisrain/metaphor-understanding-challenge.

Texts written in different languages reflect different culturally-dependent beliefs of their writers. Thus, we expect multilingual LMs (MLMs), that are jointly trained on a concatenation of text in multiple languages, to encode different cultural values for each language. Yet, as the ‘multilinguality’ of these LMs is driven by cross-lingual sharing, we also have reason to belief that cultural values bleed over from one language into another. This limits the use of MLMs in practice, as apart from being proficient in generating text in multiple languages, creating language technology that can serve a community also requires the output of LMs to be sensitive to their biases (Naous et al. 2023). Yet, little is known about how cultural values emerge and evolve in MLMs (Hershcovich et al. 2022). We are the first to study how languages can exert influence on the cultural values encoded for different test languages, by studying how such values are revised during fine-tuning. Focusing on the fine-tuning stage allows us to study the interplay between value shifts when exposed to new linguistic experience from different data sources and languages. Lastly, we use a training data attribution method to find patterns in the fine-tuning examples, and the languages that they come from, that tend to instigate value shifts.

Large Language Models (LLMs) exhibit impressive performance on a range of NLP tasks, due to the general-purpose linguistic knowledge acquired during pretraining. Existing model interpretability research (Tenney et al., 2019) suggests that a linguistic hierarchy emerges in the LLM layers, with lower layers better suited to solving syntactic tasks and higher layers employed for semantic processing. Yet, little is known about how encodings of different linguistic phenomena interact within the models and to what extent processing of linguistically-related categories relies on the same, shared model representations. In this paper, we propose a framework for testing the joint encoding of linguistic categories in LLMs. Focusing on syntax, we find evidence of joint encoding both at the same (related part-of-speech (POS) classes) and different (POS classes and related syntactic dependency relations) levels of linguistic hierarchy. Our cross-lingual experiments show that the same patterns hold across languages in multilingual LLMs.

Multilingual language models (MLMs) are jointly trained on data from many different languages such that representation of individual languages can benefit from other languages’ data. Impressive performance in zero-shot cross-lingual transfer shows that these models are able to exploit this property. Yet, it remains unclear to what extent, and under which conditions, languages rely on each other’s data. To answer this question, we use TracIn (Pruthi et al., 2020), a training data attribution (TDA) method, to retrieve training samples from multilingual data that are most influential for test predictions in a given language. This allows us to analyse cross-lingual sharing mechanisms of MLMs from a new perspective. While previous work studied cross-lingual sharing at the model parameter level, we present the first approach to study it at the data level. We find that MLMs rely on data from multiple languages during fine-tuning and this reliance increases as fine-tuning progresses. We further find that training samples from other languages can both reinforce and complement the knowledge acquired from data of the test language itself.

Large multilingual language models typically share their parameters across all languages, which enables cross-lingual task transfer, but learning can also be hindered when training updates from different languages are in conflict. In this article, we propose novel methods for using language-specific subnetworks, which control cross-lingual parameter sharing, to reduce conflicts and increase positive transfer during fine-tuning. We introduce dynamic subnetworks, which are jointly updated with the model, and we combine our methods with meta-learning, an established, but complementary, technique for improving cross-lingual transfer. Finally, we provide extensive analyses of how each of our methods affects the models.

Multilingual sentence encoders have seen much success in cross-lingual model transfer for downstream NLP tasks. The success of this transfer is, however, dependent on the model’s ability to encode the patterns of cross-lingual similarity and variation. Yet, we know relatively little about the properties of individual languages or the general patterns of linguistic variation that the models encode. In this article, we investigate these questions by leveraging knowledge from the field of linguistic typology, which studies and documents structural and semantic variation across languages. We propose methods for separating language-specific subspaces within state-of-the-art multilingual sentence encoders (LASER, M-BERT, XLM, and XLM-R) with respect to a range of typological properties pertaining to lexical, morphological, and syntactic structure. Moreover, we investigate how typological information about languages is distributed across all layers of the models. Our results show interesting differences in encoding linguistic variation associated with different pretraining strategies. In addition, we propose a simple method to study how shared typological properties of languages are encoded in two state-of-the-art multilingual models—M-BERT and XLM-R. The results provide insight into their information-sharing mechanisms and suggest that these linguistic properties are encoded jointly across typologically similar languages in these models.

In this paper, we investigate what types of stereotypical information are captured by pretrained language models. We present the first dataset comprising stereotypical attributes of a range of social groups and propose a method to elicit stereotypes encoded by pretrained language models in an unsupervised fashion. Moreover, we link the emergent stereotypes to their manifestation as basic emotions as a means to study their emotional effects in a more generalized manner. To demonstrate how our methods can be used to analyze emotion and stereotype shifts due to linguistic experience, we use fine-tuning on news sources as a case study. Our experiments expose how attitudes towards different social groups vary across models and how quickly emotions and stereotypes can shift at the fine-tuning stage.

Multilingual representations have mostly been evaluated based on their performance on specific tasks. In this article, we look beyond engineering goals and analyze the relations between languages in computational representations. We introduce a methodology for comparing languages based on their organization of semantic concepts. We propose to conduct an adapted version of representational similarity analysis of a selected set of concepts in computational multilingual representations. Using this analysis method, we can reconstruct a phylogenetic tree that closely resembles those assumed by linguistic experts. These results indicate that multilingual distributional representations that are only trained on monolingual text and bilingual dictionaries preserve relations between languages without the need for any etymological information. In addition, we propose a measure to identify semantic drift between language families. We perform experiments on word-based and sentence-based multilingual models and provide both quantitative results and qualitative examples. Analyses of semantic drift in multilingual representations can serve two purposes: They can indicate unwanted characteristics of the computational models and they provide a quantitative means to study linguistic phenomena across languages.

In this paper, we define and apply representational stability analysis (ReStA), an intuitive way of analyzing neural language models. ReStA is a variant of the popular representational similarity analysis (RSA) in cognitive neuroscience. While RSA can be used to compare representations in models, model components, and human brains, ReStA compares instances of the same model, while systematically varying single model parameter. Using ReStA, we study four recent and successful neural language models, and evaluate how sensitive their internal representations are to the amount of prior context. Using RSA, we perform a systematic study of how similar the representational spaces in the first and second (or higher) layers of these models are to each other and to patterns of activation in the human brain. Our results reveal surprisingly strong differences between language models, and give insights into where the deep linguistic processing, that integrates information over multiple sentences, is happening in these models. The combination of ReStA and RSA on models and brains allows us to start addressing the important question of what kind of linguistic processes we can hope to observe in fMRI brain imaging data. In particular, our results suggest that the data on story reading from Wehbe et al./ (2014) contains a signal of shallow linguistic processing, but show no evidence on the more interesting deep linguistic processing.