Agent-Driven Corpus Linguistics: A Framework for Autonomous Linguistic Discovery

Jia Yu; Weiwei Yu; Pengfei Xiao; Fukun Xing

cs.CLApr 8, 2026

Agent-Driven Corpus Linguistics: A Framework for Autonomous Linguistic Discovery

Authors: Jia Yu, Weiwei Yu, Pengfei Xiao, Fukun Xing

Organizations: Zhejiang International Studies University, Hangzhou, China · Tianjin University, Tianjin, China

Abstract

Corpus linguistics has traditionally relied on human researchers to formulate hypotheses, construct queries, and interpret results - a process demanding specialized technical skills and considerable time. We propose Agent-Driven Corpus Linguistics, an approach in which a large language model (LLM), connected to a corpus query engine via a structured tool-use interface, takes over the investigative cycle: generating hypotheses, querying the corpus, interpreting results, and refining analysis across multiple rounds. The human researcher sets direction and evaluates final output. Unlike unconstrained LLM generation, every finding is anchored in verifiable corpus evidence. We treat this not as a replacement for the corpus-based/corpus-driven distinction but as a complementary dimension: it concerns who conducts the inquiry, not the epistemological relationship between theory and data. We demonstrate the framework by linking an LLM agent to a CQP-indexed Gutenberg corpus (5 million tokens) via the Model Context Protocol (MCP). Given only "investigate English intensifiers," the agent identified a diachronic relay chain (so+ADJ > very > really), three pathways of semantic change (delexicalization, polarity fixation, metaphorical constraint), and register-sensitive distributions. A controlled baseline experiment shows that corpus grounding contributes quantification and falsifiability that the model cannot produce from training data alone. To test external validity, the agent replicated two published studies on the CLMET corpus (40 million tokens) - Claridge (2025) and De Smet (2013) - with close quantitative agreement. Agent-driven corpus research can thus produce empirically grounded findings at machine speed, lowering the technical barrier for a broader range of researchers.

Figures & tables

System	Corpus access	Hyp. gen.	Query	Interpret	Iterate
UDagent	UD treebanks (Python)	—	✓	—	—
AntConc+ChatAI	User-uploaded	—	✓	partial	—
CorpusChat	User-uploaded	—	✓	partial	—
AI Scientist	ML benchmarks (code)	✓	✓	✓	✓
Ours	CQP corpora (tool-use)	✓	✓	✓	✓

Table 1: Comparison of AI-assisted corpus research systems along the autonomy spectrum.

Dimension	Human-Driven	Agent-Driven
Cognitive agent	Human researcher	AI (LLM)
Epistemological mode	Corpus-based, corpus-driven, or hybrid	Same (determined by task design)
Hypothesis source	Researcher’s domain knowledge	Agent’s parametric knowledge + data
Query construction	Manual	Autonomous
Result interpretation	Human analysis	AI analysis, human review
Iteration speed	Human-paced	Machine-paced

Table 2: Two largely independent dimensions of corpus linguistics methodology. The epistemological axis (corpus-based vs. corpus-driven) concerns the direction of inquiry; the agent axis concerns who performs the inquiry. Agent-driven research can operate in either epistemological mode.

Panel of Figure 1 — Figure 1: System architecture of the agent-driven corpus linguistics framework. The orange loop at the centre is the agent’s research cycle: hypothesize, query, interpret, revise. The corpus engine (gray, dashed) feeds grounded evidence into this loop through a tool-use interface; it is an implementation detail, swappable for any MCP-compatible backend. The human researcher sets the direction and has the final word.

Panel of Figure 2 — Figure 2: Five-stage iterative workflow. Blue stages involve human participation; orange stages are autonomous agent operations. The dashed arrow indicates the iteration loop.

	anc.	med.	e.mod.	18C	e.19C	l.19C	e.20C	l.20C
very	472	576	631	1,804	1,287	1,281	959	964
so +ADJ	459	966	1,017	712	823	705	629	418
quite	179	76	159	225	222	265	211	431
really	77	54	40	151	42	199	190	185
truly	0	54	99	170	84	106	64	50
utterly	13	101	20	18	74	31	29	48

Table 3: Normalized frequency (per million words) of selected intensifiers across historical periods.

Panel of Figure 3 — Figure 3: Semantic change trajectories of English intensifiers, arranged by collocational freedom (left = maximal; right = constrained). The horizontal axis represents collocational diversity, not a single underlying continuum: complete delexicalization, polarity fixation (subjectification), and metaphorical constraint are qualitatively distinct processes ( Hopper and Traugott, 2003 ; Traugott, 1982 ; Sweetser, 1990 ) that produce different collocational profiles.

	Drama	Poetry	Journalism	Cooking
very	657	635	1,619	2,160
really	352	17	162	119
so +ADJ	800	919	742	340
quite	219	146	289	85
exceedingly	—	8	38	—

Table 4: Normalized frequency (pmw) of intensifiers in selected genres.

Capability	Without corpus (recall)	With corpus (agent-driven)
Hypothesis generation	6 well-formed hypotheses, consistent with literature	5 hypotheses, including 1 data-prompted (H5)
Frequency hierarchy	Correct rank ordering, no numbers	Precise pmw for 22 intensifiers
Diachronic trends	Correct directional predictions	8-period quantified data; very peak at 1,804 pmw in 18C
Register sensitivity	Predicted informal $>$ formal	Quantified: really 352 vs. 17 pmw (20 $\times$ )
Delexicalization	Described general multi-stage pathway	Three-pathway comparative framework with quantified polarity profiles
Collocate lists	Impressionistic examples only	807 types for very ; 67 for utterly ; full classification

Table 5: Comparison of the same LLM (Claude Opus) with and without corpus access.

	1710–1780	1780–1850	1850–1920
Agent result	105.6	90.7	53.3
Claridge (2025)	$\sim$ 120	$\sim$ 99	$\sim$ 52.5

Table 6: Replication of Claridge (2025) : normalized frequency (pmw) of reader (singular) in CLMET.

	1710–1780	1780–1850	1850–1920
remember +VBG	4	61	162
remember +to-inf	71	74	48
Gerund share	5.3%	45.2%	77.1%

Table 7: Replication of De Smet (2013) : remember +gerund vs. remember +to-infinitive in CLMET.

Finding	Description	Run 1	Run 2	Run 3
Frequency hierarchy	very $>$ so +ADJ $>$ most +ADJ	✓	✓	✓
Diachronic relay	so +ADJ $\to$ very $\to$ really	✓	✓	✓
Register sensitivity	really concentrated in drama	✓	✓	✓
Delexicalization pathways	Multiple pathways identified	3	3	2 $\dagger$
Semantic prosody of utterly	Negative/privative dominance	80%	78%	82%

Table 8: Reproducibility of core findings across three independent runs of the intensifier case study.

Corpus Linguistics

See all 15 papers

GrepSeek: Training Search Agents for Direct Corpus Interaction
May 28, 2026Alireza Salemi, Chang Zeng, Atharva Nijasure +4Tool-Augmented Language Model Agents Agentic Search
University of Massachusetts Amherst · Princeton University · Carnegie Mellon University
LLM Agents as Computational Typologists
Sep 7, 2026Changbing Yang, Christopher Hammerly, Freda Shi +1LLM Agents Linguistic Typology
University of British Columbia · University of Waterloo · Vector Institute
A large-scale pipeline for LLM-assisted corpus annotation: variation and change in the English consider construction
Oct 14, 2025Cameron Morin, Matti Marttinen LarssonLLM-Assisted Annotation
Université Paris-Cité · University of Gothenburg

May 28, 2026J/K move · Enter open · S save

GrepSeek: Training Search Agents for Direct Corpus Interaction

Alireza Salemi, Chang Zeng, Atharva Nijasure, Jui-Hui Chung+3

University of Massachusetts Amherst · Princeton University · Carnegie Mellon University

Tool-Augmented Language Model Agents Agentic Search LLM Agent Training Knowledge-Intensive QA RL for LLM Agents

Open paper AskPDF arXiv

Large Language Model (LLM) search agents have shown strong promise on knowledge-intensive tasks through iterative reasoning and retrieval. Most existing systems rely on retrievers that return ranked documents from a pre-built index. We explore a complementary paradigm in which the agent treats the corpus as the search environment and finds evidence through executable shell commands. We introduce GrepSeek, an optimized direct corpus interaction (DCI) agent that learns to find, filter, and compose evidence over large text corpora. To stabilize reinforcement learning (RL) over large corpora, we train in two stages: first, we initialize the policy using verified, causally grounded search trajectories generated by an answer-aware Tutor and an answer-blind Planner; then, we refine the policy using Group Relative Policy Optimization (GRPO). To make DCI practical at scale, we introduce two semantics-preserving execution optimizations: Pruned Adaptive Command Execution, which reduces shell-based search latency by up to

77\times

on a 14GB corpus with 21 million documents using a compact auxiliary structure, and Sharded-Parallel Corpus Search, which achieves up to

7.6\times

speedup without additional preprocessing; both preserve equivalence with sequential execution. Across eight open-domain QA benchmarks, GrepSeek achieves the strongest overall performance, with a statistically significant relative improvement of

5.7\%

over the best baseline. Our analysis shows how DCI-optimized agents conduct flexible and effective compositional search through direct corpus interaction.

Abstract

Corpus Linguistics

Explore similar work

GrepSeek: Training Search Agents for Direct Corpus Interaction

LLM Agents as Computational Typologists

A large-scale pipeline for LLM-assisted corpus annotation: variation and change in the English consider construction