Despite their impressive performance in a wide range of NLP tasks, Large Language Models (LLMs) have been reported to encode worrying-levels of gender biases. Prior work has proposed debiasing methods that require human labelled examples, data augmentation and fine-tuning of LLMs, which are computationally costly. Moreover, one might not even have access to the model parameters for performing debiasing such as in the case of closed LLMs such as GPT-4. To address this challenge, we propose bias suppression that prevents biased generations of LLMs by simply providing textual preambles constructed from manually designed templates and real-world statistics, without accessing to model parameters. We show that, using CrowsPairs dataset, our textual preambles covering counterfactual statements can suppress gender biases in English LLMs such as LLaMA2. Moreover, we find that gender-neutral descriptions of gender-biased objects can also suppress their gender biases. Moreover, we show that bias suppression has acceptable adverse effect on downstream task performance with HellaSwag and COPA.

Acquiring factual knowledge for language models (LMs) in low-resource languages poses a serious challenge, thus resorting to cross-lingual transfer in multilingual LMs (ML-LMs). In this study, we ask how ML-LMs acquire and represent factual knowledge. Using the multilingual factual knowledge probing dataset, mLAMA, we first conducted a neuron investigation of ML-LMs (specifically, multilingual BERT). We then traced the roots of facts back to the knowledge source (Wikipedia) to identify the ways in which ML-LMs acquire specific facts. We finally identified three patterns of acquiring and representing facts in ML-LMs: language-independent, cross-lingual shared and transferred, and devised methods for differentiating them. Our findings highlight the challenge of maintaining consistent factual knowledge across languages, underscoring the need for better fact representation learning in ML-LMs.

Entity disambiguation (ED) is typically solved by learning to classify a given mention into one of the entities in the model’s entity vocabulary by referring to their embeddings. However, this approach cannot address mentions of entities that are not covered by the entity vocabulary. Aiming to enhance the applicability of ED models, we propose a method of extending a state-of-the-art ED model by dynamically computing embeddings of out-of-vocabulary entities. Specifically, our method computes embeddings from entity descriptions and mention contexts. Experiments with standard benchmark datasets show that the extended model performs comparable to or better than existing models whose entity embeddings are trained for all candidate entities as well as embedding-free models. We release our source code and model checkpoints at https://github.com/studio-ousia/steel.

Probing classifiers have been extensively used to inspect whether a model component captures specific linguistic phenomena. This top-down approach is, however, costly when we have no probable hypothesis on the association between the target model component and phenomena. In this study, aiming to provide a flexible, exploratory analysis of a neural model at various levels ranging from individual neurons to the model as a whole, we present a bottom-up approach to inspect the target neural model by using neuron representations obtained from a massive corpus of text. We first feed massive amount of text to the target model and collect sentences that strongly activate each neuron. We then abstract the collected sentences to obtain neuron representations that help us interpret the corresponding neurons; we augment the sentences with linguistic annotations (e.g., part-of-speech tags) and various metadata (e.g., topic and sentiment), and apply pattern mining and clustering techniques to the augmented sentences. We demonstrate the utility of our method by inspecting the pre-trained BERT. Our exploratory analysis reveals that i) specific phrases and domains of text are captured by individual neurons in BERT, ii) a group of neurons simultaneously capture the same linguistic phenomena, and iii) deeper-level layers capture more specific linguistic phenomena.

There exist biases in individual’s language use; the same word (e.g., cool) is used for expressing different meanings (e.g., temperature range) or different words (e.g., cloudy, hazy) are used for describing the same meaning. In this study, we propose a method of modeling such personal biases in word meanings (hereafter, semantic variations) with personalized word embeddings obtained by solving a task on subjective text while regarding words used by different individuals as different words. To prevent personalized word embeddings from being contaminated by other irrelevant biases, we solve a task of identifying a review-target (objective output) from a given review. To stabilize the training of this extreme multi-class classification, we perform a multi-task learning with metadata identification. Experimental results with reviews retrieved from RateBeer confirmed that the obtained personalized word embeddings improved the accuracy of sentiment analysis as well as the target task. Analysis of the obtained personalized word embeddings revealed trends in semantic variations related to frequent and adjective words.