Area Overview

The objective of this project is to examine the usefulness of synthetic data in linguistic research. Here, synthetic data refers broadly to data obtained through Large Language Models (LLMs), including texts generated by LLMs as well as their by-product internal representations (e.g., embeddings). Specifically, we investigate the extent to which synthetic data can reproduce empirical data, such as human-generated texts, established linguistic knowledge, and neural activity.

If synthetic data can sufficiently approximate empirical data, it may accelerate linguistic research by enabling the use of low-cost synthetic data in place of costly empirical data. Expected advantages include: (i) quantitative evaluation of linguistic phenomena—such as meaning and cognition—that have been difficult to assess using conventional methods; (ii) data augmentation, including missing-value completion; (iii) alternatives to experiments that are difficult to conduct due to physical or ethical constraints; and (iv) reduction of costs associated with experiments requiring human participants.

To achieve these goals, this research area is organized around four Planned Research projects: Synthetic Corpus Studies, Vector-based Glottometrics, Linguistic Simulationology, and Comparative Representational Neuroscience.

Overview of the Planned Research Projects

A01 Synthetic Corpus Studies

When conducting research on language, the linguistic data under investigation—that is, a corpus—plays a crucial role. By analyzing a corpus, we can address a wide range of questions, such as in what contexts particular expressions or words occur, how frequently they appear, in what senses they are used, and how their usage varies across writers or genres. However, constructing a corpus is an extremely labor-intensive task.

In our project, we therefore explore the use of corpora artificially generated by LLMs—so-called synthetic corpora—as substitutes for human-generated corpora. We aim to answer questions such as how similar synthetic corpora are to human-generated corpora and how to make synthetic corpora more similar to human-generated ones. Furthermore, by investigating whether LLMs possess internal conceptual structures often attributed to humans (for example, what is referred to as construal in cognitive linguistics), we seek to broaden the scope of LLM applications in linguistic research. In addition, because LLMs can be trained under a wide range of conditions, we examine how different training conditions affect the similarity between LLMs and humans, and to what extent LLM learning approximates human language acquisition and language change.

A02 Vector-based Glottometrics

The objective of this research project is to develop methods for measuring hidden characteristics of language. We refer to these characteristics as the intrinsic properties of language. This involves measuring properties not explicitly revealed on the surface of linguistic data, such as impressions and semantic shifts in words. To put it another way, we create tools that would function as microscopes or X-rays for linguistic data. Traditionally, intrinsic properties have been measured primarily based on linguists’ introspection, experiments with human subjects, and manual analysis of texts. While the effectiveness of these methods is beyond doubt, they all involve manual investigations, making it difficult to scale up the scope of measurement. Besides, some experiments and analyses cannot be conducted due to physical or ethical constraints.

The primary approach of our research utilizes the internal state of LLMs. When language data is fed into an LLM, an internal state (more precisely, the hidden state of the language model) is obtained as a byproduct. The internal state is represented as a set of numerical values (a vector) that reflects the input data. We can measure intrinsic properties by applying various calculations to this internal state. We also develop methods to measure intrinsic properties by treating LLMs as experimental subjects. In collaboration with the other projects, we aim to clarify what intrinsic properties of language can be measured and to what extent.

A03 Linguistic Simulationology

Our project aims to explore the potential of LLMs as simulators for linguistic research. Specifically, we investigate how and to what extent LLMs can be utilized to model linguistic phenomena that are difficult to observe directly (e.g., language change and language contact), as well as to serve as alternatives to experiments that are challenging to conduct due to physical constraints or high costs (e.g., cross-linguistic experiments requiring speakers of different languages). Through this investigation, we seek to clarify both the possibilities and the limitations of LLM-based simulation. We will evaluate model validity by examining the extent to which LLM-generated simulation results reproduce empirical data and established linguistic findings.

Through simulation-based data augmentation, LLMs are expected to complement linguistic data that are difficult or impossible to obtain in reality—for example, reconstructing plausible linguistic forms that are not attested in historical records but are assumed to have existed. Furthermore, virtual experimental settings that allow specific factors to be systematically introduced or removed will enable us to observe model behavior in counterfactual environments. By adopting a constructive methodology that seeks to understand phenomena through their reproduction, this project aims to provide new insights into fundamental linguistic questions concerning the dynamic nature of language, as well as linguistic universality and diversity. In collaboration with other projects, we will further examine the potential and the limitations of LLMs as simulators across a wide range of linguistic phenomena.

A04 Comparative Representational Neuroscience

Our project positions LLMs not merely as models of human language processing, but as models for understanding general-purpose semantic representations in the human brain. While LLM representations can accurately predict brain activity during language tasks, their relationship to non-linguistic neural representations—such as vision and other sensory modalities—remains unclear. We therefore examine the relationship between brain activity and LLM representations across the processing cascade from perception to linguistic expression, to clarify semantic representations that span both linguistic and non-linguistic domains.

Building on this, we establish and extend an in silico framework that integrates LLM-based brain activity generation and brain decoding. We develop models that predict and generate brain activity from arbitrary text, testing whether known findings can be reproduced and enabling data-driven exploration of hypotheses. We also compare neural and LLM representations, and develop methods to characterize their correspondence. This integrated framework allows us to evaluate hypotheses without relying on measured data, thereby extending the methodological foundations of brain information analysis and providing a neuroscientific basis for LLM-based approaches across the entire research area.