Agent-driven Corpus Linguistics: A New Framework for AI to Independently Explore Linguistic Patterns

This article introduces the agent-driven corpus linguistics framework, which combines large language models (LLMs) with corpus query engines to enable AI to independently complete hypothesis generation, corpus querying, and result interpretation. In the study of English intensifiers, this framework discovered important patterns such as diachronic transfer chains and semantic evolution paths, providing a new paradigm for linguistic research.

Traditional corpus linguistics relies on human researchers to complete the entire process of hypothesis formulation, query construction, and result interpretation, with three major issues: 1. High technical threshold (requiring mastery of query languages like CQL and statistical tools); 2. Low research efficiency (time-consuming manual parameter adjustment and result analysis); 3. Poor reproducibility (large differences in analysis paths and judgment criteria among researchers).

To address traditional bottlenecks, researchers proposed the agent-driven corpus linguistics framework. This framework connects LLMs to corpus query engines via structured tool interfaces, with AI taking over the research cycle (hypothesis generation, corpus querying, result interpretation, iterative analysis). Humans only need to set directions and evaluate outputs, and all findings are anchored in verifiable corpus evidence. The framework does not replace existing paradigms but serves as a complementary dimension, focusing on "who conducts the research" rather than the epistemological relationship between theory and data.

The research team connected LLM agents to the CQP-indexed Gutenberg Corpus (5 million words) via the Model Context Protocol (MCP). MCP provides a standardized tool interface, enabling LLMs to convert natural language intentions into precise CQL query statements, execute queries, and parse results. CQP is a powerful corpus query processor that supports complex linguistic queries.

After giving the agent the instruction to "investigate English intensifiers", the AI independently discovered: 1. Diachronic transfer chains (intergenerational replacement from so+adjective → very → really); 2. Three paths of semantic evolution (delexicalization, polarity fixation, metaphorical constraint); 3. Register sensitivity (significant differences in distribution frequency of different intensifiers across registers like spoken and written language).

Control experiments show that LLMs without corpus access can only provide qualitative descriptions and cannot give quantitative data, diachronic trends, or statistical tests; while the framework combines LLM reasoning with corpus evidence to achieve "1+1>2". In external validity tests, the agent reproduced the studies of Claridge (2025) and De Smet (2013) on the CLMET corpus, with quantitative results highly consistent with the original, proving the framework's reliability.

Significance of the framework: Reduces technical thresholds (no need to master query languages), improves efficiency (research cycle shortened from weeks to hours), enhances reproducibility (standardized processes reduce human variation), and expands research boundaries (large-scale systematic surveys become possible). Limitations: The depth of AI analysis may be affected by training data biases, over-interpretation of corpora needs to be prevented, and the role of humans needs further definition. Prospects: With the improvement of LLM capabilities and standardization of tool interfaces, it is expected to promote AI-enabled research paradigms in more disciplines.