Chinese Essay Rhetoric Recognition Using LoRA, In-context Learning and Model Ensemble

Yuxuan Lai; Xiajing Wang; Chen Zheng

cs.CLMar 24, 2026

Chinese Essay Rhetoric Recognition Using LoRA, In-context Learning and Model Ensemble

Authors: Yuxuan Lai, Xiajing Wang, Chen Zheng

Abstract

Rhetoric recognition is a critical component in automated essay scoring. By identifying rhetorical elements in student writing, AI systems can better assess linguistic and higher-order thinking skills, making it an essential task in the area of AI for education. In this paper, we leverage Large Language Models (LLMs) for the Chinese rhetoric recognition task. Specifically, we explore Low-Rank Adaptation (LoRA) based fine-tuning and in-context learning to integrate rhetoric knowledge into LLMs. We formulate the outputs as JSON to obtain structural outputs and translate keys to Chinese. To further enhance the performance, we also investigate several model ensemble methods. Our method achieves the best performance on all three tracks of CCL 2025 Chinese essay rhetoric recognition evaluation task, winning the first prize.

Figures & tables

Input : Sentence-1.庄稼汉们站在地头，望着这片黄澄澄像狗尾巴的稻谷。Sentence-2.他感到十分开心。

Track-1 : Sentence-id:

\left\{1\right\}

. form-level fine-grained type: 比喻-明喻(Metaphor-Simile)

Track-2 : Sentence-id:

\left\{1\right\}

. content-level fine-grained type: 比喻-实在物比喻(Metaphor-Concrete)

Track-3 : Sentence-id:

\left\{1\right\}

. 本体(tenor)-稻谷(swathes of rice) 喻体(vehicle)-狗尾巴(tails of dogs) 喻词(comparator)-像(like)

Table 1: Example of expected outputs for Track-1, Track-2, and Track-3.

Figure 1: An illustration of our methods.

Inputs : 庄稼汉们站在地头，望着这片黄澄澄像狗尾巴的稻谷。他感到十分开心。

Original Annotation : Sentence ID: 1, rhetoric: 比喻, form: 明喻, content: 实在物, conjunction: [像], conjunctionBeginIdx:[16], conjunctionEndIdx:[16], tenor:[稻谷], tenorBeginIdx:[21], tenorEndIdx:[22], vehicle:[狗尾巴], vehicleBeginIdx:[17], vehicleEndIdx:[19]

JSON output + translation : [{“内容片段”: “庄稼汉们站在地头，望着这片黄澄澄像狗尾巴的稻谷。”, “修辞手法”: “比喻”, “形式上的细粒度修辞分类”: “明喻”, “内容上的细粒度修辞分类”: “实在物”, “喻词”: “像”, “本体”: “稻谷”, “喻体”: “狗尾巴”}]

Table 2: An Example for the JSON-format output.

{“messages”: [{“role”: “system”, “content”: SYSTEM PROMPT }, {“role”: “user”, “content”: INPUT DOCUMENT }, {“role”: “assistant”, “content”: JSON OUTPUT }]}

Table 4

[{“role”: “system”, “content”: SYSTEM PROMPT }, {“role”: “user”, “content”: INPUT DOCUMENT 1 }, {“role”: “assistant”, “content”: JSON OUTPUT 1 }, ……, {“role”: “assistant”, “content”: JSON OUTPUT k }, {“role”: “user”, “content”: INPUT DOCUMENT

{}_{\texttt{TEST}}

}]

Table 5

Name	Backend Model	Detail Settings
$M_{\text{LoRA-1}}$	$M_{\text{LoRA}}$	$k=20$ , JOIN
$M_{\text{LoRA-2}}$	$M_{\text{LoRA}}$	$k=0$ , JOIN
$M_{1}$	Qwen-max	$k=50$ , SEPA
$M_{2}$	Qwen-long	$k=50$ , SEPA
$M_{3}$	Qwen-long	$k=20$ , JOIN
$M_{4}$	Qwen-long	$k=32$ , JOIN , INSTRUCT

Table 3: Different model settings.

Teams	Track-1	Track-2	Track-3	Avg.
ours	47.18	54.03	39.94	47.05
BLCU	43.47	51.71	38.27	44.48
YNU	40.30	52.23	26.25	39.59
official baseline	43.68	51.14	37.48	44.19
Single Model-1 (smaller $k$ )	41.88 ( $M_{3}$ )	51.05 ( $M_{3}$ )	37.07 ( $M_{\text{LoRA-2}}$ )	–
Single Model-2 (larger $k$ )	45.23 ( $M_{1}$ )	52.52 ( $M_{1}$ )	39.39 ( $M_{\text{LoRA-1}}$ )	–

Table 4: Results of the official ranking list and some ablation settings. The details of our models are elaborated in Table 3 and § 4.1 .

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Track-1 : 比喻(明喻, 暗喻, 借喻), 比拟(名词, 动词, 形容词), 夸张(直接夸张, 间接夸张), 排比(成分排比, 句子排比), 反复(间隔反复, 连续反复), 设问(问答连属, 问答不连属), 反问(单句反问, 复句反问), 摹状(通感, 直感)
Track-2 : 比喻(实在物, 动作, 抽象概念)，比拟(拟人, 拟物)，夸张(扩大夸张, 缩小夸张, 超前夸张)，排比(并列, 承接, 递进)
Track-3 : 比喻-明喻（本体，喻体，喻词），比喻-暗喻（本体，喻体），比喻-借喻（喻体），比拟（比拟的特征，比拟的对象），夸张（夸张的对象，夸张的描述），排比（排比词）

Appendix

Table 5: Label sets for Track-1 and Track-2, and the component list for Track-3.

statistics of the raw training set :

total number of training instances: 50

total number of fine-grained rhetoric devices: 63 (content-level), 105 (form-level)

total number of rhetoric component: 145

average character number per instance: 98.3

average sentence number per instance: 3.7

numbers of each rhetoric device in the training set:

Appendix

Table 6: Statistics of the training data

[{“role”: “system”, “content”: SYSTEM PROMPT }, {“role”: “user”, “content”: INPUT DOCUMENT 1 }, {“role”: “assistant”, “content”: JSON OUTPUT 1 }, {“role”: “user”, “content”: INPUT DOCUMENT 2 }, {“role”: “assistant”, “content”: JSON OUTPUT 2 }, ……, {“role”: “user”, “content”: INPUT DOCUMENT k }, {“role”: “assistant”, “content”: JSON OUTPUT k }, {“role”: “user”, “content”: INPUT DOCUMENT

{}_{\texttt{TEST}}

}]

Appendix

Table 10

model name	No Output (%)	JSON Parsing Error (%)	# Wrong Coarse-grained Type	# Fine-grained Type
$M_{\text{LoRA-1}}$	0.15	1.20	428	5
$M_{\text{LoRA-2}}$	0.05	0.47	1024	4
$M_{1}$ (Joint Training Part. Rhetoric types: metaphor, personification, hyperbole, and parallelism)	0.07	1.19	358	3
$M_{1}$ (Trace-1-only Part. Rhetoric types: rhetorical question, hypophora, repetition, synesthesia)	0.01	0.07	1	0

Appendix

Table 7: Statistics of inference error rates. The details of our models are elaborated in Table 3 and § 4.1 .

model name	train cost	inference cost	price per million tokens
$M_{\text{LoRA}}$	460.36	–	150
$M_{\text{LoRA-1}}$	–	1187.66	4(input), 12(output)
$M_{\text{LoRA-2}}$	–	430.50	4(input), 12(output)
$M_{1}$ (Joint Training Part. Rhetoric types: metaphor, personification, hyperbole, and parallelism)	–	1083.42	2.4(input), 9.6(output)

Appendix

Table 8: Cost for model training and inference (CNY).

M_{\text{LoRA}}

is the Qwen 2.5 72B model after LoRA continue-training.

M_{\text{LoRA-2}}

is the inference model with

M_{\text{LoRA}}

, joint setting with all rhetoric types, without in-context learning example.

M_{\text{LoRA-1}}

is similar to

M_{\text{LoRA-2}}

, but incorporating the in-context learning strategy, with in-context example number

k=20

and similarity-based example selection and ordering strategies.

M_{1}

is an in-context learning based inference model, with Qwen-max as the backend model, in-context example number

k=50

, and the similarity-based example ordering strategy, focusing on the joint training part (see § 3.3 ).

model :

M_{\text{LoRA-1}}

content : ”[”内容片段”: ”上次还没被妈妈骂够吗？”, ”修辞手法”: ”反问”, ”形式上的细粒度修辞分类”: ”单句反问”]

error type : erroneous predicts ” (\u0022) as ” (\u201d) .

model :

M_{\text{LoRA-1}}

content : [”内容片段”: ”我的家，是一个“小小动物园”里面有四个动物，分别是：“黄牛妈妈、大棕能爸爸、猴子弟弟和小鱼的我。”, ”修辞手法”: ”比喻”, ”形式上的细粒度修辞分类”: ”暗喻”, ”内容上的细粒度修辞分类”: ”实在物”, ”本体”: ”我的家”, ”喻体”: ”一个“小小动物园””, ”内容片段”: ”我的家，是一个“小小动物园”里面有四个动物，分别是：“黄牛妈妈、大棕能爸爸、猴子弟弟和小鱼的我。”, ”修辞手法”: ”排比”, ”形式上的细粒度修辞分类”: ”成分排比”, ”内容上的细粒度修辞分类”: ”并列”, ”排比词”: ”动物”, ”内容片段”: ”说起我的妈妈，她肯定是一个不折不扣的＂老黄牛”啦。”, ”修辞手法”: ”比喻”, ”形式上的细粒度修辞分类”: ”暗喻”, ”内容上的细粒度修辞分类”: ”实在物”, ”本体”: ”我的妈妈”, ”喻体”: ”一个不折不扣的＂老黄牛””]

error type : multiple errors in generating quotation marks.

model :

M_{1}

(Joint Training Part. Rhetoric types: metaphor, personification, hyperbole, and parallelism)

Appendix

Table 9: Case study for JSON parsing error.

C-ReD: A Comprehensive Chinese Benchmark for AI-Generated Text Detection Derived from Real-World Prompts
Apr 13, 2026Chenxi Qing, Junxi Wu, Zheng Liu +5AI-Generated Text Detection Robustness of Machine-Generated Text Detection
Tsinghua University · Nankai University · Harbin Institute of Technology, Shenzhen +2
Does the Linguistic Register of Thinking Matter? Wenyan as a Compact Reasoning Register for Chinese LLMs
Oct 14, 2026Minghan Li, Guodong ZhouEfficient Language Model Reasoning Language Modeling
Soochow University, Suzhou, China
EnsemJudge: Enhancing Reliability in Chinese LLM-Generated Text Detection through Diverse Model Ensembles
Mar 30, 2026Zhuoshang Wang, Yubing Ren, Guoyu Zhao +3AI-Generated Text Detection Robustness of Machine-Generated Text Detection
Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China

Apr 13, 2026J/K move · Enter open · S save

C-ReD: A Comprehensive Chinese Benchmark for AI-Generated Text Detection Derived from Real-World Prompts

Chenxi Qing, Junxi Wu, Zheng Liu, Yixiang Qiu+4

Tsinghua University · Nankai University · Harbin Institute of Technology, Shenzhen+2

AI-Generated Text Detection Robustness of Machine-Generated Text Detection

Open paper AskPDF arXiv

Recently, large language models (LLMs) are capable of generating highly fluent textual content. While they offer significant convenience to humans, they also introduce various risks, like phishing and academic dishonesty. Numerous research efforts have been dedicated to developing algorithms for detecting AI-generated text and constructing relevant datasets. However, in the domain of Chinese corpora, challenges remain, including limited model diversity and data homogeneity. To address these issues, we propose C-ReD: a comprehensive Chinese Real-prompt AI-generated Detection benchmark. Experiments demonstrate that C-ReD not only enables reliable in-domain detection but also supports strong generalization to unseen LLMs and external Chinese datasets-addressing critical gaps in model diversity, domain coverage, and prompt realism that have limited prior Chinese detection benchmarks. We release our resources at https://github.com/HeraldofLight/C-ReD.

Abstract

Explore similar work

C-ReD: A Comprehensive Chinese Benchmark for AI-Generated Text Detection Derived from Real-World Prompts

Does the Linguistic Register of Thinking Matter? Wenyan as a Compact Reasoning Register for Chinese LLMs

EnsemJudge: Enhancing Reliability in Chinese LLM-Generated Text Detection through Diverse Model Ensembles