A Reproducible Multi-Architecture Baseline for Token-Level Chinese Metaphor Identification under the MIPVU Framework

# A Reproducible Multi-Architecture Baseline for Token-Level Chinese Metaphor Identification under the MIPVU Framework
Source: [https://arxiv.org/html/2605.07170](https://arxiv.org/html/2605.07170)
\[ BoldFont = lmroman10\-bold\.otf, ItalicFont = lmroman10\-italic\.otf, BoldItalicFont = lmroman10\-bolditalic\.otf, \] \[ BoldFont = lmsans10\-bold\.otf, ItalicFont = lmsans10\-oblique\.otf, \] \[ ItalicFont = lmmono10\-italic\.otf, AutoFakeBold = 2\.0, \]\\setCJKmainfont\[Extension=\.otf, BoldFont=FandolSong\-Bold\]FandolSong\-Regular

###### Abstract

Metaphor is pervasive in everyday language, yet token\-level computational identification of metaphor\-related words in Chinese under the MIPVU framework remains under\-explored relative to English\. This paper presents a reproducible multi\-architecture baseline for token\-level metaphor identification on the PSU Chinese Metaphor Corpus \(PSU CMC\), the only widely available MIPVU\-annotated Chinese corpus\. We systematically compare three model families: \(i\) encoder fine\-tuning with Chinese RoBERTa\-wwm\-ext\-large; \(ii\) MelBERT adapted to Chinese using a newly constructed basic\-meaning resource derived from the Modern Chinese Dictionary, 7th edition \(MCD7\), comprising 74,823 entries with 71\.51% PSU CMC vocabulary coverage; and \(iii\) Qwen3\.5\-9B fine\-tuned with QLoRA as an instruction\-tuned generative baseline\. Across five fixed seeds, MelBERT MIP\-only achieves the strongest performance at 0\.7281±\\pm0\.0050 test positive F1, marginally above MelBERT Full \(0\.7270±\\pm0\.0069\) and clearly above plain RoBERTa \(0\.7142±\\pm0\.0121\)\. The Qwen QLoRA generative configuration trails encoder baselines by approximately 11 F1 points \(0\.6157±\\pm0\.0113\)\. Three findings merit attention: \(1\) the SPV channel of MelBERT does not contribute reliable positive signal in Chinese, consistent with the dominance of conventional metaphor; \(2\) the Qwen–encoder gap is concentrated in recall, reflecting the discrete\-commitment limitation of generative output; \(3\) several Qwen task formulations fail due to format design rather than model capacity\. We release all split manifests, per\-seed outputs, the MCD7 basic\-meaning embedding pipeline, and training scripts to serve as a common reference for future Chinese metaphor identification research\.

## 1Introduction

Metaphor is pervasive in everyday language\(Steenet al\.,[2010](https://arxiv.org/html/2605.07170#bib.bib3)\), and its computational identification is a long\-standing problem with applications in sentiment analysis, machine translation, discourse understanding, and language education\. The Metaphor Identification Procedure \(MIP\)\(Pragglejaz Group,[2007](https://arxiv.org/html/2605.07170#bib.bib11)\)and its operationalization MIPVU\(Steenet al\.,[2010](https://arxiv.org/html/2605.07170#bib.bib3)\)provide the dominant linguistic protocol for token\-level metaphor annotation: each lexical unit is judged metaphor\-related if its contextual meaning differs from, but can be understood by comparison with, a more concrete or basic meaning\.

For English, MIPVU has driven a substantial research program\. Successive systems—from biLSTM\-CRF taggers to Transformer\-based sequence labelers—have been evaluated on the VU Amsterdam Metaphor Corpus and TOEFL essays\(Maoet al\.,[2019](https://arxiv.org/html/2605.07170#bib.bib12); Suet al\.,[2020](https://arxiv.org/html/2605.07170#bib.bib13); Gonget al\.,[2020](https://arxiv.org/html/2605.07170#bib.bib14)\), and the most influential recent architecture, MelBERT\(Choiet al\.,[2021](https://arxiv.org/html/2605.07170#bib.bib5)\), explicitly grounds its design in MIP and Selectional Preference Violation theory\. For Chinese, by contrast, the picture is markedly thinner\.Lu and Wang \([2017](https://arxiv.org/html/2605.07170#bib.bib1)\)introduced the PSU Chinese Metaphor Corpus \(PSU CMC\), the only widely available token\-level MIPVU\-annotated Chinese corpus, and yet PSU CMC has remained an under\-evaluated benchmark: modern encoder fine\-tuning, MelBERT\-style lexical fusion, and instruction\-tuned LLMs have not been systematically compared on it under a shared seed protocol with publicly released artifacts\.

This paper addresses that gap\. We treat the question pragmatically:*what does a careful, reproducible baseline for token\-level metaphor identification on PSU CMC look like in 2026?*Concretely, we compare three model families that span the methodological space: \(i\) standard encoder fine\-tuning with Chinese RoBERTa\-wwm\-ext\-large\(Cuiet al\.,[2021](https://arxiv.org/html/2605.07170#bib.bib6)\); \(ii\) MelBERT adapted to Chinese, requiring a Chinese basic\-meaning resource that does not yet exist in open form; and \(iii\) Qwen3\.5\-9B fine\-tuned with QLoRA\(Dettmerset al\.,[2023](https://arxiv.org/html/2605.07170#bib.bib8); Huet al\.,[2022](https://arxiv.org/html/2605.07170#bib.bib7)\)as an instruction\-tuned generative baseline\.

The MelBERT adaptation requires resolving a resource gap\. The original MelBERT relies on WordNet first\-sense glosses to supply each token’s basic meaning\. No analogous open Chinese resource exists\. We therefore construct a basic\-meaning resource derived from the Modern Chinese Dictionary, 7th edition \(MCD7\), comprising 74,823 entries with parseable basic meanings encoded as 1024\-dimensional vectors, with71\.51%71\.51\\%coverage of PSU CMC vocabulary\. This resource is itself a contribution of the paper; the released artifacts \(embeddings, mappings, pipeline\) are licensed for research use, while the original copyrighted gloss text is not redistributed\.

Across five fixed seeds, the strongest configuration we evaluate is MelBERT MIP\-only \(i\.e\., MelBERT with the SPV channel removed\), at0\.7281±0\.00500\.7281\\pm 0\.0050test positive F1\. This is marginally above the full three\-channel MelBERT \(0\.7270±0\.00690\.7270\\pm 0\.0069\) and clearly above plain RoBERTa fine\-tuning \(0\.7142±0\.01210\.7142\\pm 0\.0121\)\. Qwen3\.5\-9B with QLoRA trails the encoder\-based baselines by roughly 11 absolute F1 points \(0\.6157±0\.01130\.6157\\pm 0\.0113\), with the gap concentrated in recall and amplified on the fiction register\.

Beyond the headline numbers, three findings merit attention\. First, the SPV channel of MelBERT does not contribute reliable positive signal in our Chinese setting: the MIP\-only ablation has both higher mean F1 and notably lower seed variance than the Full configuration\. Second, the Qwen lag is asymmetric: precision is comparable to encoder baselines while recall is substantially lower, a pattern consistent with the discrete commitment imposed by generative output and the limited capacity of low\-rank adaptation for fine\-grained token\-level signals\. Third, several Qwen task formulations fail in qualitatively different ways—BIO span tagging collapses on predominantly single\-token Chinese metaphors, and structured generation collapses under supervision\-token truncation—failures that reflect format design rather than model capacity\.

The contributions of this paper are:

1. 1\.A reproducible multi\-architecture baseline for token\-level metaphor identification on PSU CMC, with five\-seed runs across four configurations and full release of split, code, and per\-seed outputs\.
2. 2\.A Chinese MIP basic\-meaning resource derived from MCD7, enabling MelBERT\-style lexical fusion in Chinese for the first time\.
3. 3\.Empirical findings on Chinese metaphor identification, including the unexpectedly competitive MelBERT MIP\-only ablation, the precision\-favoring asymmetry of QLoRA\-adapted generative LLMs, and the format\-design sensitivity of generative task formulations\.

The remainder of the paper is organized as follows\. Section 2 situates this work in the metaphor identification literature\. Section 3 describes the dataset, the MCD7 basic\-meaning resource, the model architectures, and the experimental protocol\. Section 4 presents the results and analyzes the key findings\. Section 5 discusses limitations and future directions, and Section 6 concludes\.

## 2Related Work

We situate our work along three axes: the MIPVU annotation framework and its Chinese adaptation \(§2\.1\), computational methods for metaphor identification \(§2\.2\), and lexical resources for basic\-meaning representation \(§2\.3\)\.

### 2\.1MIPVU and Token\-level Metaphor Resources

The Metaphor Identification Procedure \(MIP\) was introduced by the Pragglejaz Group\(Pragglejaz Group,[2007](https://arxiv.org/html/2605.07170#bib.bib11)\)as a reproducible inter\-annotator protocol for tagging metaphor\-related words in running text\.Steenet al\.\([2010](https://arxiv.org/html/2605.07170#bib.bib3)\)extended it into MIPVU, which adds explicit treatment of indirect metaphor, direct metaphor, and borderline cases, and was applied to construct the VU Amsterdam Metaphor Corpus, the canonical English MIPVU\-annotated benchmark\. For Chinese,Lu and Wang \([2017](https://arxiv.org/html/2605.07170#bib.bib1)\)adapted MIPVU to Mandarin and constructed the PSU Chinese Metaphor Corpus \(PSU CMC\) by sampling documents from the Lancaster Corpus of Mandarin Chinese\(McEnery and Xiao,[2004](https://arxiv.org/html/2605.07170#bib.bib2)\); PSU CMC is the corpus we use in this paper\.

Other Chinese metaphor corpora have been released, but they target different tasks and annotation schemes\. CMC\(Liet al\.,[2023](https://arxiv.org/html/2605.07170#bib.bib9)\)provides sentence\-level metaphor labels with a heavy positive\-class skew \(∼\\sim91% positive\), making it a metaphor\-rich classification benchmark rather than a representative running\-text identification benchmark\. CMDAG\(Shaoet al\.,[2024](https://arxiv.org/html/2605.07170#bib.bib10)\)annotates metaphor with grounds \(rationales\) for metaphor generation\. Neither aligns directly with MIPVU’s token\-level identification task on naturally distributed text, so we do not evaluate cross\-corpus transfer in this paper\.

Wang et al\. \(2019\) adapted the MIPVU protocol specifically for Chinese in a chapter of the multi\-language MIPVU volume\(Wanget al\.,[2019](https://arxiv.org/html/2605.07170#bib.bib18); Naceyet al\.,[2019](https://arxiv.org/html/2605.07170#bib.bib17)\), documenting challenges including word segmentation ambiguity, grammaticalized prepositions, and compound\-internal metaphor\. Their adapted protocol defines Chinese\-specific metaphor flag \(MFlag\) words— pre\-source markers \(像, 好像, 如, 如同, 犹如, 好比\) and post\-source markers \(一样, 似的, 般\)—which signal direct metaphor but are themselves not annotated as metaphor\-related in PSU CMC\.

### 2\.2Metaphor Identification Methods

Computational approaches to token\-level metaphor identification have evolved through three overlapping waves\. Sequence\-tagging neural models, beginning with biLSTM\-CRF and ELMo\-based architectures\(Maoet al\.,[2019](https://arxiv.org/html/2605.07170#bib.bib12)\), treat metaphor identification as a standard tagging task and were the dominant paradigm prior to the widespread use of pretrained Transformer encoders\. The introduction of BERT\(Devlinet al\.,[2019](https://arxiv.org/html/2605.07170#bib.bib15)\)and RoBERTa\(Liuet al\.,[2019](https://arxiv.org/html/2605.07170#bib.bib16)\)reshaped the field: DeepMet\(Suet al\.,[2020](https://arxiv.org/html/2605.07170#bib.bib13)\)won the 2020 VUA Metaphor Detection Shared Task using a reading\-comprehension formulation over RoBERTa with linguistic features; IlliniMet\(Gonget al\.,[2020](https://arxiv.org/html/2605.07170#bib.bib14)\)combined RoBERTa with WordNet, VerbNet, POS, and concreteness features\. MelBERT\(Choiet al\.,[2021](https://arxiv.org/html/2605.07170#bib.bib5)\)departed from this “RoBERTa\-plus\-features” paradigm by encoding two metaphor\-theoretic inductive biases \(MIP and SPV\) directly into its architecture\. MelBERT remains the strongest published BERT\-family approach we adapt; its dependence on a basic\-meaning resource is what motivates our MCD7 construction\.

The third wave—instruction\-tuned large language models—is more recent and less settled for token\-level metaphor identification\. Generative LLMs have been applied to figurative language tasks including simile recognition, metaphor generation, and metaphor explanation, but token\-level MIPVU identification poses a structural challenge: the model must commit to a discrete metaphor label per token without a calibrated probability head, in contrast to encoder classifiers\. Several authors have explored prompt\-engineering, multi\-stage prompting, and task\-form variants to address this mismatch\. Our Qwen3\.5\-9B with QLoRA experiments contribute to this line of work in the Chinese setting, where systematic comparison against MIPVU encoder baselines has not previously been reported\.

For Chinese specifically, Zhang et al\. \(2021\) proposed SaGE, a syntax\-aware GCN with ELECTRA model achieving 85\.22% macro\-F1 on the CCL2018 Chinese metaphor evaluation dataset\(Zhanget al\.,[2021](https://arxiv.org/html/2605.07170#bib.bib19)\)\. However, CCL2018 is a sentence\-level three\-class task \(verb metaphor / noun metaphor / literal\), fundamentally different from PSU CMC’s token\-level binary identification under MIPVU\. No prior work has reported encoder fine\-tuning, MelBERT\-style lexical fusion, or systematic LLM comparison on PSU CMC\.

In the LLM prompting paradigm, Huang and Liu \(2026\) reported a GPT\-4\-based, interpretable MIPVU rule\-script framework on PSU CMC in an arXiv preprint\(Huang and Liu,[2026](https://arxiv.org/html/2605.07170#bib.bib20)\)\. Fuoli et al\. \(2025\) systematically compared prompt engineering, retrieval\-augmented generation \(RAG\), and fine\-tuning for English metaphor identification, and reported strongest performance for the fine\-tuned setting\(Fuoliet al\.,[2025](https://arxiv.org/html/2605.07170#bib.bib21)\)\. Our work extends this prompt\-versus\-fine\-tune comparison to the Chinese MIPVU setting, where we additionally include encoder baselines that the LLM literature has not compared against\.

### 2\.3Lexical Resources for Basic Meaning

MelBERT\-style lexical fusion requires per\-token basic\-meaning representations\. For English, this role is filled by WordNet first\-sense glosses, which are open\-licensed and computationally accessible\. For Chinese, the lexical\-resource situation is heterogeneous: HowNet, Chinese WordNet, and BCC have been used for various Chinese NLP tasks, but each has limitations as a MelBERT substrate—HowNet’s sememe representation does not align with MIPVU’s notion of basic meaning, Chinese WordNet has sparser coverage than its English counterpart, and BCC is a corpus rather than a lexicon\. We therefore construct a basic\-meaning resource directly from MCD7, the most widely cited authoritative reference dictionary of modern Chinese, with full extraction and encoding pipelines released as part of this work\.

## 3Method

This section describes the evaluation corpus and its split protocol \(§3\.1–3\.2\), the basic\-meaning resource we construct for MelBERT adaptation \(§3\.3\), the three model configurations under comparison \(§3\.4\), and the shared training and evaluation protocol \(§3\.5\)\.

### 3\.1PSU Chinese Metaphor Corpus

The PSU Chinese Metaphor Corpus \(PSU CMC\)\(Lu and Wang,[2017](https://arxiv.org/html/2605.07170#bib.bib1)\)is a multi\-register Chinese corpus annotated with token\-level metaphor labels following the Metaphor Identification Procedure VU \(MIPVU\)\(Steenet al\.,[2010](https://arxiv.org/html/2605.07170#bib.bib3)\)\. The text is sampled from the Lancaster Corpus of Mandarin Chinese \(LCMC\)\(McEnery and Xiao,[2004](https://arxiv.org/html/2605.07170#bib.bib2)\), a one\-million\-word balanced corpus of written Mandarin, from whichLu and Wang \([2017](https://arxiv.org/html/2605.07170#bib.bib1)\)drew 75 documents covering three registers: academic prose, fiction, and news\.

Each lexical unit is assigned a binary metaphor flag based on the contextual\-vs\-basic meaning contrast central to MIPVU\. As shown in Table[1](https://arxiv.org/html/2605.07170#S3.T1), the corpus contains 1,724 sentences and 35,746 tokens, of which 3,272 tokens \(9\.16%\) are labeled as metaphor\-related words\. Metaphor density varies markedly across registers, ranging from 6\.36% in news to 13\.67% in academic\.

Table 1:PSU CMC dataset statistics by register and split\.
### 3\.2Data Split

We adopt a*file\-level*70/10/20 train/dev/test split with seed 42, ensuring that no document appears in more than one partition\. Sentence\-level splits, common in earlier metaphor identification work, can leak document\-level cues since adjacent sentences from the same source share register, topic, and stylistic conventions, leading to optimistic estimates of generalization\.

The resulting split contains 1,182 train, 198 dev, and 344 test sentences \(from 52, 8, and 15 documents respectively; see Table[1](https://arxiv.org/html/2605.07170#S3.T1)\)\. The split manifest—i\.e\., the document IDs assigned to each partition—is released as part of our reproducibility package, and all experiments throughout this paper use this split without modification\.

### 3\.3Modern Chinese Dictionary Basic\-Meaning Resource

A central component of the MelBERT architecture \(Section[3\.4\.2](https://arxiv.org/html/2605.07170#S3.SS4.SSS2)\) is the*basic meaning*of each lexical unit—the most concrete or primary sense, used as the contrastive anchor in the MIP channel\. The original MelBERT formulation for English uses WordNet first\-sense glosses\(Choiet al\.,[2021](https://arxiv.org/html/2605.07170#bib.bib5)\)\. For Chinese, no equivalent open lexical resource exists\. We therefore construct one from the*Modern Chinese Dictionary*, 7th edition\(Dictionary Editorial Office, Institute of Linguistics, Chinese Academy of Social Sciences,[2016](https://arxiv.org/html/2605.07170#bib.bib4)\), hereafter MCD7—the authoritative Chinese\-language reference dictionary maintained by the Institute of Linguistics, Chinese Academy of Social Sciences, and published by the Commercial Press\.

#### 3\.3\.1Construction Pipeline

Our extraction pipeline proceeds in five stages:

1. 1\.Format extraction: We decode the dictionary from MDX format, yielding 74,823 raw entries\.
2. 2\.Sense parsing: For each entry, we parse multiple sense definitions, separated by sense markers in the source\.
3. 3\.Basic\-meaning selection: Per MIPVU principles, we select the most concrete and primary sense \(typically the first sense definition\) as the basic meaning\. Entries with no parseable sense receive a fallback to their full headword definition\.
4. 4\.Cross\-reference resolution: 5,947 entries \(7\.95%\) contain reference indicators \(e\.g\.,jian‘see’,tong‘same as’,cankan‘cf\.’\) that redirect to other entries\. We recursively resolve these references with a maximum depth of 5 and circular\-reference detection\. Of the 5,947 referencing entries, 4,861 \(81\.74%\) resolve successfully; 1,081 fail due to missing target entries, and 5 form cycles\.
5. 5\.Embedding encoding: Each final basic\-meaning text is encoded into a 1024\-dimensional vector using Chinese RoBERTa\-wwm\-ext\-large\(Cuiet al\.,[2021](https://arxiv.org/html/2605.07170#bib.bib6)\)\[CLS\]token representation\.

The resulting resource provides a one\-to\-one mapping from dictionary headwords to 1024\-dimensional basic\-meaning vectors, ready for direct consumption by the MelBERT MIP channel described in Section 3\.4\.2\.

#### 3\.3\.2Statistics and Coverage

Table[2](https://arxiv.org/html/2605.07170#S3.T2)summarizes the resource\. The dictionary contains 74,823 entries; 99\.33% have parseable basic\-meaning text \(the remaining 0\.67% retain only the headword as fallback\)\. Of these entries, 80\.29% are single\-sense, while 19\.71% are multi\-sense \(mean 1\.31, max 24 senses per entry\)\.

Coverage of the PSU CMC vocabulary is 71\.51%—that is, 5,094 of 7,123 unique tokens in PSU CMC have a corresponding dictionary entry\. The remaining 28\.49% of PSU CMC tokens fall outside the dictionary, primarily proper nouns, register\-specific compounds, and rare expressions\. These tokens receive a zero\-vector fallback at MelBERT input, marked via an out\-of\-vocabulary mask\.

Table 2:Statistics and coverage of the MCD7 basic\-meaning resource\.
#### 3\.3\.3Use in MelBERT and Limitations

For each input token, MelBERT looks up its dictionary entry and uses the encoded basic\-meaning vector as the input to its MIP channel\. We use*only the first sense*per entry, following the original MelBERT design\. This means the 19\.71% of multi\-sense entries are underutilized; richer integration of the full sense inventory is a natural extension we discuss in Section[5](https://arxiv.org/html/2605.07170#S5)\.

#### 3\.3\.4License and Release

The Modern Chinese Dictionary is copyrighted by the Commercial Press; we therefore do not redistribute the original gloss text\. We release: \(a\) the per\-entry 1024\-dimensional embedding vectors derived from the basic\-meaning text; \(b\) entry token\-to\-index mappings; and \(c\) the full extraction and encoding pipeline as scripts\. Users wishing to reproduce or extend the resource must obtain the dictionary independently\. Released artifacts \(a\)–\(c\) are distributed under MIT license \(code\) and CC BY 4\.0 \(derived data\) for research use\.

### 3\.4Model Architectures

We compare three model families that span the current methodological space for token\-level metaphor identification\.

#### 3\.4\.1RoBERTa\-wwm\-ext\-large \(Encoder Fine\-tuning\)

Our first baseline is standard token\-classification fine\-tuning of Chinese RoBERTa\-wwm\-ext\-large\(Cuiet al\.,[2021](https://arxiv.org/html/2605.07170#bib.bib6)\), a 24\-layer Transformer encoder with whole\-word\-masking pre\-training on Chinese Wikipedia and EXT data\. We add a linear classification head on top of each token’s final hidden state, producing a binary metaphor probability per token\. Training uses standard cross\-entropy loss; no auxiliary signals or external resources are involved\.

#### 3\.4\.2MelBERT

##### Architecture\.

MelBERT\(Choiet al\.,[2021](https://arxiv.org/html/2605.07170#bib.bib5)\)extends an encoder backbone with two parallel channels grounded in metaphor identification theory: \(1\) the Metaphor Identification Procedure \(MIP\) channel contrasts each token’s contextualized representation against its basic\-meaning vector \(Section[3\.3](https://arxiv.org/html/2605.07170#S3.SS3)\); \(2\) the Selectional Preference Violation \(SPV\) channel contrasts the token’s local context with the token’s contextualized representation, flagging selectional anomalies\. Both channels feed into independent classification heads, and a third*fusion head*combines them\. The training loss is a weighted sum of three binary cross\-entropy losses \(one per head\)\.

##### Chinese adaptation\.

We use the same Chinese RoBERTa\-wwm\-ext\-large backbone as in Section[3\.4\.1](https://arxiv.org/html/2605.07170#S3.SS4.SSS1)and replace the WordNet\-based basic\-meaning embeddings of the original English MelBERT with our MCD7\-derived embeddings \(Section[3\.3](https://arxiv.org/html/2605.07170#S3.SS3)\)\. Tokens not in MCD7 receive a zero\-vector fallback marked via an out\-of\-vocabulary mask\.

##### Channel ablations\.

In addition to the standard three\-channel*Full*configuration, we evaluate two ablations:*MIP\-only*\(SPV channel and fusion head removed; loss is single\-channel binary cross\-entropy\) and*SPV\-only*\(mirror ablation\)\.

#### 3\.4\.3Qwen3\.5\-9B with QLoRA \(Instruction\-tuned Generation\)

##### Backbone\.

Our third model is Qwen3\.5\-9B, a 9\-billion\-parameter instruction\-tuned language model\. Unlike RoBERTa and MelBERT, Qwen treats metaphor identification as a generation task: given a sentence as prompt, the model generates a structured output identifying metaphor positions, which we then parse deterministically\.

##### Parameter\-efficient adaptation\.

We use QLoRA\(Dettmerset al\.,[2023](https://arxiv.org/html/2605.07170#bib.bib8)\), which combines 4\-bit NF4 quantization of the frozen backbone with low\-rank adapter training\(Huet al\.,[2022](https://arxiv.org/html/2605.07170#bib.bib7)\)\. Adapters are inserted into the four attention projection matrices \(query, key, value, output\) with rankr=16r=16and scaling factorα=32\\alpha=32\. This reduces trainable parameters to under 1% of the full model while keeping the backbone in 4\-bit precision \(peak GPU memory: 15\.3 GB on a single RTX 5090\)\.

##### Task formulations\.

The choice of how to encode token\-level metaphor identification as a generation task is non\-trivial\. We systematically compare six task formulations \(Q1–Q8 in Table[3](https://arxiv.org/html/2605.07170#S3.T3)\), spanning classification\-style and generation\-style designs:

- •Classification\-style: token\-level head over hidden states \(Q1\), and BIO\-tagged span prediction \(Q4\)\.
- •Generation\-style: free\-form JSON listing metaphor tokens \(Q2\), QA\-style natural language \(Q6\), and structured generation with token\-id constraints \(Q8 v1/v2\)\.

Table 3:Test positive F1 of six Qwen3\.5\-9B \+ QLoRA task formulations on PSU CMC test set, single seed \(=42\)\.Q2 \(Generative JSON Extraction\) emerged as the strongest formulation in single\-seed experiments \(Table[3](https://arxiv.org/html/2605.07170#S3.T3)\); we adopt it for the 5\-seed main comparison reported in Section[4\.1](https://arxiv.org/html/2605.07170#S4.SS1)\.

### 3\.5Training and Evaluation Protocol

All experiments run on a single NVIDIA RTX 5090 \(32 GB VRAM\), with Qwen3\.5\-9B loaded in 4\-bit NF4 quantization and bf16 compute\.

For each main model \(RoBERTa, MelBERT Full, MelBERT MIP\-only, and Qwen Q2\), we run training and evaluation across five fixed seeds: \{42, 123, 2024, 7, 31415\}\. The seed list is fixed in advance and released alongside our code, so all multi\-seed numbers in this paper correspond to the same seed set across architectures\. We report mean±\\pmstd with population standard deviation \(ddof==0\) throughout, matching the convention used in our internal aggregation scripts\.

Table[4](https://arxiv.org/html/2605.07170#S3.T4)summarizes the final training configurations\. RoBERTa and MelBERT both use the Chinese RoBERTa\-wwm\-ext\-large backbone with learning rate5×10−55\\\!\\times\\\!10^\{\-5\}, effective batch size 16, up to 10 epochs, and maximum sequence length 256\. Qwen Q2 uses learning rate2×10−42\\\!\\times\\\!10^\{\-4\}, effective batch size 16 \(batch 1 with gradient accumulation 16\), up to 3 epochs, and maximum sequence length 1024 to accommodate the prompt\-plus\-response generative format\. The Qwen configuration is the one verified against per\-seedtrain\_summary\.jsonfiles; an early planning document contained inaccurate values, which we do not adopt\.

Table 4:Hyperparameters for the four model configurations\.For all models, training uses early stopping with patience 3 \(for RoBERTa and MelBERT\) or patience 2 \(for Qwen\) on dev\-set monitoring\. RoBERTa and MelBERT monitor dev positive\-class F1 directly; Qwen Q2 monitors dev cross\-entropy loss, since per\-batch dev F1 was not logged during training\.

We report*positive\-class F1*\(denoted*Test pos\-F1*\) as the primary metric\. PSU CMC is heavily class\-imbalanced \(9\.16% positive at the token level\), so macro F1 is dominated by the trivial negative\-class score and inflates absolute numbers without reflecting metaphor identification quality\. We report macro F1 alongside positive F1 for completeness, but all comparisons in this paper, including the register\-level breakdowns, use positive\-class F1\.

For the Qwen generative configuration, predictions are obtained by deterministic JSON parsing of the model’s output\. Tokens listed in the parsed JSON receive a positive label; all others receive a negative label\. Failures of JSON parsing \(e\.g\., malformed brackets\) are counted as a separate*parse failure rate*; in practice this rate is below 0\.3% across all 5 Qwen seeds and does not materially affect F1\.

Per\-register F1 is computed by partitioning the test set into the three registers \(academic, news, fiction\) and computing positive\-class F1 within each subset\. Register subsets share the same gold labels and are non\-overlapping\. All metric aggregation uses ddof==0 across seeds\.

## 4Experiments

We first report the main cross\-architecture comparison \(§4\.1\), per\-register performance \(§4\.2\), Qwen task\-form comparison \(§4\.3\), and MelBERT channel ablation \(§4\.4\), followed by an analysis of the Qwen precision–recall asymmetry \(§4\.5\)\.

### 4\.1Main Comparison

Table[5](https://arxiv.org/html/2605.07170#S4.T5)reports five\-seed test performance for RoBERTa, MelBERT \(Full and MIP\-only\), and Qwen3\.5\-9B with QLoRA\. The strongest configuration is MelBERT MIP\-only at0\.7281±0\.00500\.7281\\pm 0\.0050test positive\-class F1, marginally above MelBERT Full \(0\.7270±0\.00690\.7270\\pm 0\.0069\) and clearly above plain RoBERTa fine\-tuning \(0\.7142±0\.01210\.7142\\pm 0\.0121\)\. The Qwen Q2 generative configuration trails the encoder\-based baselines by roughly 11 absolute F1 points at0\.6157±0\.01130\.6157\\pm 0\.0113\.

Several patterns are worth noting\. First, MelBERT MIP\-only is not only the highest\-scoring configuration on positive F1 but also the most stable across seeds: its standard deviation \(0\.00500\.0050\) is approximately28%28\\%smaller than MelBERT Full’s \(0\.00690\.0069\) and less than half of RoBERTa’s \(0\.01210\.0121\)\. Second, the gap between MelBERT Full and MelBERT MIP\-only \(\+0\.0011\+0\.0011in pos\-F1\) is well within one standard deviation of either configuration; we do not claim statistical superiority of MIP\-only, but treat the consistency of the pattern across seeds and metrics as evidence that the SPV channel does not contribute reliably positive signal in our setting\. Third, although Qwen lags on positive F1, its precision \(0\.6963±0\.01030\.6963\\pm 0\.0103\) is comparable to encoder baselines while its recall \(0\.5526±0\.02350\.5526\\pm 0\.0235\) is markedly lower\.

Table 5:Test set performance of four model configurations on PSU CMC, reported as mean±\\pmstandard deviation over 5 seeds \(population std, ddof==0\)\.The competitive performance of MIP\-only merits further examination\. The original MelBERT paper\(Choiet al\.,[2021](https://arxiv.org/html/2605.07170#bib.bib5)\)motivates the SPV channel as a complementary signal to MIP: where MIP detects contextual\-vs\-basic meaning contrasts, SPV detects selectional violations between a token and its local context\. On English VUA, removing the SPV channel hurts performance\. On PSU CMC, we observe the opposite trend: MIP\-only matches or slightly exceeds the Full configuration on every aggregate metric \(Table[5](https://arxiv.org/html/2605.07170#S4.T5), Table[6](https://arxiv.org/html/2605.07170#S4.T6)\), with notably lower seed variance\.

We offer two non\-exclusive hypotheses, presented as motivation for future work rather than confirmed claims:

1. 1\.Conventional metaphor dominance in Chinese\.A substantial fraction of metaphor in Chinese running text is conventional rather than novel: the metaphorical sense is itself well\-established in the lexicon\. For such metaphors, the contextual\-vs\-basic\-meaning contrast \(MIP\) is informative, while the selectional\-violation contrast \(SPV\) is weak: the metaphor\-context pairing is so frequent that it does not look*anomalous*to the SPV channel\.
2. 2\.Optimization noise from the three\-way loss\.The Full configuration trains a three\-head loss \(MIP head, SPV head, and a fusion head\) jointly\. Reducing this to a single MIP head halves the number of competing optimization objectives; the observed reduction in seed variance \(−28%\-28\\%\) is consistent with a less\-noisy optimization landscape, independent of any linguistic claim about SPV’s signal value\.

### 4\.2Per\-Register Breakdown

Figure[1](https://arxiv.org/html/2605.07170#S4.F1)visualizes positive F1 separately for the three registers in PSU CMC\. Across all four configurations, academic prose is the easiest register, followed by news, with fiction the hardest\. The ordering is consistent with metaphor density: academic prose has the highest metaphor density \(13\.67%13\.67\\%, see Table[1](https://arxiv.org/html/2605.07170#S3.T1)\) while fiction has only7\.54%7\.54\\%and features more novel, context\-dependent metaphor\.

The cross\-architecture pattern is preserved within each register: MelBERT MIP\-only and Full are essentially tied on academic \(0\.7532±0\.00480\.7532\\pm 0\.0048vs\.0\.7538±0\.00840\.7538\\pm 0\.0084\) and news \(0\.7357±0\.01430\.7357\\pm 0\.0143vs\.0\.7327±0\.00680\.7327\\pm 0\.0068\), with MIP\-only modestly ahead on fiction \(0\.6725±0\.02710\.6725\\pm 0\.0271vs\.0\.6694±0\.02370\.6694\\pm 0\.0237\)\. Notably, fiction is also where seed variance is largest for every model \(Qwen reachesσ=0\.0378\\sigma=0\.0378on fiction versus0\.00550\.0055on academic\), and where Qwen’s gap to the encoders is most severe \(roughly−0\.13\-0\.13absolute pos\-F1 versus MelBERT MIP\-only on fiction, versus−0\.09\-0\.09on academic\)\.

![Refer to caption](https://arxiv.org/html/2605.07170v1/x1.png)Figure 1:Per\-register positive F1 across the four model configurations on PSU CMC test set\. Bars show 5\-seed means; error bars represent±1​σ\\pm 1\\sigma\(population std\)\. Registers ordered by MelBERT \(MIP\-only\) F1, descending\.The fiction register’s lower performance and higher variance reflect two factors\. First, fiction has the lowest metaphor density of the three registers \(7\.54%7\.54\\%, Table[1](https://arxiv.org/html/2605.07170#S3.T1)\), so each individual fiction prediction contributes more variance to the per\-register F1\. Second, fiction metaphor is, on average, more novel and context\-dependent than the conventionalized metaphor common in academic prose\. For Qwen specifically, fiction is also the register where seed 7 collapses \(0\.47020\.4702vs\. a non\-seed\-7 mean of0\.56000\.5600\), pulling the overall seed 7 number down to the cluster outlier visible in Figure[3](https://arxiv.org/html/2605.07170#S4.F3)\.

### 4\.3Qwen Task\-Form Comparison

We previously reported \(Table[3](https://arxiv.org/html/2605.07170#S3.T3), single\-seed\) that Q2 generative JSON extraction is the strongest of six task formulations we tried for Qwen3\.5\-9B\. Figure[2](https://arxiv.org/html/2605.07170#S4.F2)restates this ranking as a ladder against the RoBERTa five\-seed baseline\.

Two findings are robust across formulations:

1. 1\.No Qwen formulation reaches encoder\-baseline performance\.Even Q2 \(0\.62750\.6275\) is approximately0\.0870\.087below RoBERTa’s mean \(0\.71420\.7142\)\. The next\-strongest formulation, Q1 token\-level classification \(0\.56800\.5680baseline;0\.58080\.5808best after ablation\), is about0\.130\.13below RoBERTa\.
2. 2\.Generation\-style formulations dominate classification\-style ones at the top of the ranking, but the two paradigms are not strictly ordered\.Q2 \(generation\) outperforms Q1 \(classification\) by0\.0600\.060, but Q1 in turn outperforms Q8 v2 \(generation, longer context\)\. The poorest two formulations are Q4 \(BIO span, classification\) and Q8 v1 \(structured generation with truncated supervision\)\.

![Refer to caption](https://arxiv.org/html/2605.07170v1/x2.png)Figure 2:Test positive F1 across six Qwen3\.5\-9B task formulations, single seed \(=42\)\. Color groups classification \(Q1, Q4\) versus generation \(Q2, Q6, Q8\) paradigms\. Dashed horizontal line: RoBERTa\-large 5\-seed mean for reference\.Two failure modes are worth highlighting for their generalizability\.

- •Q4 \(BIO Span\)achieves0\.40490\.4049overall F1 but F1=0=0on the I\-tag class\. Chinese metaphors are predominantly single\-token at the lexical\-unit granularity used by PSU CMC; multi\-token metaphor spans \(which BIO tagging is designed to handle\) are rare, and the I\-tag head learns to predict near\-zero probability everywhere\.
- •Q8 v1 \(max length 256\)collapses to F1≈0\\approx 0, while Q8 v2 \(max length 512\) recovers to0\.52990\.5299\. The cause is supervision\-token truncation: in Q8 v1, the model output is frequently cut off before the relevant tokens are generated\.

### 4\.4MelBERT Channel Ablation

Table[6](https://arxiv.org/html/2605.07170#S4.T6)reports the three\-way MelBERT ablation: Full \(0\.7270±0\.00690\.7270\\pm 0\.0069\), MIP\-only \(0\.7281±0\.00500\.7281\\pm 0\.0050\), and SPV\-only \(0\.7206±0\.00640\.7206\\pm 0\.0064\)\. The standalone SPV channel is the weakest of the three; the MIP channel alone is competitive with the fused configuration\.

Table 6:MelBERT channel ablation, 5\-seed mean±\\pmstd \(population std, ddof==0\)\.Figure[3](https://arxiv.org/html/2605.07170#S4.F3)visualizes per\-seed positive F1 for the four main configurations\. Two observations stand out: the MelBERT configurations \(Full and MIP\-only\) cluster tightly around their respective means, with no seed\-level outliers, while Qwen Q2 has one clear outlier at seed 7 \(0\.59440\.5944, dragged primarily by fiction\-register performance\)\.

![Refer to caption](https://arxiv.org/html/2605.07170v1/x3.png)Figure 3:Per\-seed test positive F1 for the four model configurations\. Each point is one seed; box plots show median, IQR, and range across the 5 seeds\. Qwen seed 7 is highlighted as a fiction\-register outlier\.
### 4\.5Why Does Qwen Lag, and Why Is Recall the Weak Side?

The Qwen Q2 generative configuration is roughly1111absolute F1 points below the encoder baselines on PSU CMC test \(Table[5](https://arxiv.org/html/2605.07170#S4.T5)\)\. Crucially, this gap is concentrated in*recall*\(0\.5526±0\.02350\.5526\\pm 0\.0235\) rather than precision \(0\.6963±0\.01030\.6963\\pm 0\.0103\): Qwen tends to be conservative in flagging metaphor, missing borderline cases that the encoder baselines catch\.

Two structural causes are likely\. First, generative output forces a discrete commitment per token \(the token either appears in the generated JSON list or it does not\), with no calibrated probability gradient as in classification heads\. Tokens whose metaphor status is genuinely ambiguous are systematically dropped rather than included with low confidence\. Second, the QLoRA low\-rank adaptation modifies only the four attention projection matrices and may underfit a fine\-grained token\-level signal compared to a dense classification head trained on top of full\-rank hidden states\. The combination—a hard discrete decision plus low\-rank adaptation—biases Qwen toward the precision\-favoring regime we observe\.

## 5Discussion

We discuss our reproducibility commitments and released artifacts \(§5\.1\), the limitations of this study \(§5\.2\), and directions for future work \(§5\.3\)\.

### 5\.1Reproducibility and Released Artifacts

We release: \(a\) the file\-level train/dev/test split manifest \(Section[3\.2](https://arxiv.org/html/2605.07170#S3.SS2)\); \(b\) per\-seed checkpoints, evaluation outputs, and aggregated metrics for all four model configurations; \(c\) the MCD7 dictionary basic\-meaning embedding pipeline and the released embeddings under the licensing terms in Section[3\.3\.4](https://arxiv.org/html/2605.07170#S3.SS3.SSS4); \(d\) all training and aggregation scripts\. Our intended use is to provide a common reference point for future work on token\-level metaphor identification on PSU CMC: we expect future systems to compare against the configurations and seed protocol reported here, rather than against informal estimates from the literature\.

### 5\.2Limitations

Several limitations are worth flagging explicitly\. First, our MelBERT implementation uses only the first sense per dictionary entry \(Section[3\.3\.3](https://arxiv.org/html/2605.07170#S3.SS3.SSS3)\)\. The19\.71%19\.71\\%of multi\-sense entries in MCD7 are underutilized; richer sense\-aware integration \(e\.g\., attention\-weighted multi\-sense aggregation\) is a natural follow\-up\. Second, our Qwen Q2 configuration was selected by single\-seed task\-form comparison \(Section[4\.3](https://arxiv.org/html/2605.07170#S4.SS3)\); a more thorough hyperparameter search across the better\-performing formulations could narrow the gap to encoder baselines, though we do not expect it to close\. Third, dev\-set early\-stopping for Qwen used cross\-entropy loss rather than dev positive F1 \(Section[3\.5](https://arxiv.org/html/2605.07170#S3.SS5)\), a logging artifact rather than a methodological choice; future runs should monitor dev positive F1 directly\. Fourth, we evaluate only on PSU CMC test; cross\-corpus generalization \(e\.g\., to CMC\(Liet al\.,[2023](https://arxiv.org/html/2605.07170#bib.bib9)\)or CMDAG\(Shaoet al\.,[2024](https://arxiv.org/html/2605.07170#bib.bib10)\)\) is left for future work, particularly because those corpora use different annotation schemes and are not directly comparable\.

### 5\.3Future Work

The largest open direction we identify is multi\-sense MelBERT: the machinery to consume multiple basic\-meaning senses per token is straightforward \(e\.g\., gated aggregation or cross\-sense attention\) and the resource is already in hand \(MCD7,19\.71%19\.71\\%multi\-sense entries\)\. A second direction is domain\-adaptive pre\-training: a balanced Chinese written corpus with metaphor\-aware pre\-training objectives could plausibly improve all encoder baselines simultaneously, providing a ceiling for the MelBERT MIP\-only finding\. A third direction is paradigm hybridization—using encoder predictions as a high\-precision filter and an LLM as a recall booster on the residual—which our precision\-recall analysis \(Section[4\.5](https://arxiv.org/html/2605.07170#S4.SS5)\) suggests is structurally sensible\.

## 6Conclusion

We presented a reproducible multi\-architecture baseline for token\-level Chinese metaphor identification under the MIPVU framework, evaluated on the PSU Chinese Metaphor Corpus, comparing encoder fine\-tuning \(RoBERTa\-wwm\-ext\-large\), specialized lexical fusion \(MelBERT, Full and channel\-ablated variants\), and instruction\-tuned generation with parameter\-efficient adaptation \(Qwen3\.5\-9B with QLoRA\)\. The strongest configuration we evaluated is MelBERT MIP\-only at0\.7281±0\.00500\.7281\\pm 0\.0050test positive F1 across five seeds, marginally above MelBERT Full and notably above plain RoBERTa fine\-tuning\. We additionally release a basic\-meaning resource derived from the Modern Chinese Dictionary \(7th edition\), supporting MIPVU\-style lexical fusion in Chinese for the first time\. Beyond the headline numbers, we report several methodological findings: the SPV channel of MelBERT contributes little reliable signal in Chinese; the Qwen generative configuration’s gap to encoders is concentrated in recall and amplified on fiction register; and several Qwen task formulations fail in qualitatively distinct ways that reflect format\-design rather than model\-capacity issues\. We hope these baselines, configurations, and released artifacts provide a useful reference point for future work on Chinese metaphor identification\.

## References

- M\. Choi, S\. Lee, E\. Choi, H\. Park, J\. Lee, D\. Lee, and J\. Lee \(2021\)MelBERT: metaphor detection via contextualized late interaction using metaphorical identification theories\.InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(NAACL\-HLT\),pp\. 1763–1773\.Cited by:[§1](https://arxiv.org/html/2605.07170#S1.p2.1),[§2\.2](https://arxiv.org/html/2605.07170#S2.SS2.p1.1),[§3\.3](https://arxiv.org/html/2605.07170#S3.SS3.p1.1),[§3\.4\.2](https://arxiv.org/html/2605.07170#S3.SS4.SSS2.Px1.p1.1),[§4\.1](https://arxiv.org/html/2605.07170#S4.SS1.p3.1)\.
- Y\. Cui, W\. Che, T\. Liu, B\. Qin, and Z\. Yang \(2021\)Pre\-training with whole word masking for Chinese BERT\.IEEE/ACM Transactions on Audio, Speech, and Language Processing29,pp\. 3504–3514\.Cited by:[§1](https://arxiv.org/html/2605.07170#S1.p3.1),[item 5](https://arxiv.org/html/2605.07170#S3.I1.i5.p1.1),[§3\.4\.1](https://arxiv.org/html/2605.07170#S3.SS4.SSS1.p1.1)\.
- T\. Dettmers, A\. Pagnoni, A\. Holtzman, and L\. Zettlemoyer \(2023\)QLoRA: efficient finetuning of quantized LLMs\.InAdvances in Neural Information Processing Systems,Cited by:[§1](https://arxiv.org/html/2605.07170#S1.p3.1),[§3\.4\.3](https://arxiv.org/html/2605.07170#S3.SS4.SSS3.Px2.p1.2)\.
- J\. Devlin, M\. Chang, K\. Lee, and K\. Toutanova \(2019\)BERT: pre\-training of deep bidirectional transformers for language understanding\.InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(NAACL\-HLT\),pp\. 4171–4186\.Cited by:[§2\.2](https://arxiv.org/html/2605.07170#S2.SS2.p1.1)\.
- Dictionary Editorial Office, Institute of Linguistics, Chinese Academy of Social Sciences \(2016\)Xiandai hanyu cidian \(modern chinese dictionary\)\.7 edition,The Commercial Press,Beijing\.External Links:ISBN 978\-7\-100\-12450\-8Cited by:[§3\.3](https://arxiv.org/html/2605.07170#S3.SS3.p1.1)\.
- M\. Fuoli, W\. Huang, J\. Littlemore, S\. Turner, and E\. Wilding \(2025\)Metaphor identification using large language models: a comparison of rag, prompt engineering, and fine\-tuning\.External Links:2509\.24866,[Link](https://arxiv.org/abs/2509.24866)Cited by:[§2\.2](https://arxiv.org/html/2605.07170#S2.SS2.p4.1)\.
- H\. Gong, K\. Gupta, A\. Jain, and S\. Bhat \(2020\)IlliniMet: illinois system for metaphor detection with contextual and linguistic information\.InProceedings of the Second Workshop on Figurative Language Processing,pp\. 146–153\.Cited by:[§1](https://arxiv.org/html/2605.07170#S1.p2.1),[§2\.2](https://arxiv.org/html/2605.07170#S2.SS2.p1.1)\.
- E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen \(2022\)LoRA: low\-rank adaptation of large language models\.InInternational Conference on Learning Representations \(ICLR\),Cited by:[§1](https://arxiv.org/html/2605.07170#S1.p3.1),[§3\.4\.3](https://arxiv.org/html/2605.07170#S3.SS4.SSS3.Px2.p1.2)\.
- W\. Huang and M\. Liu \(2026\)Interpretable chinese metaphor identification via llm\-assisted mipvu rule script generation: a comparative protocol study\.External Links:2603\.10784,[Link](https://arxiv.org/abs/2603.10784)Cited by:[§2\.2](https://arxiv.org/html/2605.07170#S2.SS2.p4.1)\.
- Y\. Li, S\. Wang, C\. Lin, F\. Guerin, and L\. Barrault \(2023\)A chinese metaphor corpus and its use for metaphor recognition\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),Cited by:[§2\.1](https://arxiv.org/html/2605.07170#S2.SS1.p2.1),[§5\.2](https://arxiv.org/html/2605.07170#S5.SS2.p1.1)\.
- Y\. Liu, M\. Ott, N\. Goyal, J\. Du, M\. Joshi, D\. Chen, O\. Levy, M\. Lewis, L\. Zettlemoyer, and V\. Stoyanov \(2019\)RoBERTa: a robustly optimized BERT pretraining approach\.arXiv preprint arXiv:1907\.11692\.Cited by:[§2\.2](https://arxiv.org/html/2605.07170#S2.SS2.p1.1)\.
- X\. Lu and B\. P\. Wang \(2017\)Towards a metaphor\-annotated corpus of Mandarin Chinese\.Language Resources and Evaluation51\(3\),pp\. 663–694\.External Links:[Document](https://dx.doi.org/10.1007/s10579-017-9392-9)Cited by:[§1](https://arxiv.org/html/2605.07170#S1.p2.1),[§2\.1](https://arxiv.org/html/2605.07170#S2.SS1.p1.1),[§3\.1](https://arxiv.org/html/2605.07170#S3.SS1.p1.1)\.
- R\. Mao, C\. Lin, and F\. Guerin \(2019\)End\-to\-end sequential metaphor identification inspired by linguistic theories\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics \(ACL\),Florence, Italy,pp\. 3888–3898\.External Links:[Document](https://dx.doi.org/10.18653/v1/P19-1378)Cited by:[§1](https://arxiv.org/html/2605.07170#S1.p2.1),[§2\.2](https://arxiv.org/html/2605.07170#S2.SS2.p1.1)\.
- A\. McEnery and Z\. Xiao \(2004\)The Lancaster corpus of Mandarin Chinese: a corpus for monolingual and contrastive language study\.InProceedings of the Fourth International Conference on Language Resources and Evaluation \(LREC’04\),Lisbon, Portugal\.Cited by:[§2\.1](https://arxiv.org/html/2605.07170#S2.SS1.p1.1),[§3\.1](https://arxiv.org/html/2605.07170#S3.SS1.p1.1)\.
- S\. Nacey, L\. Greve, A\. Dorst, and T\. Krennmayr \(Eds\.\) \(2019\)Metaphor identification in multiple languages: mipvu around the world\.John Benjamins,Amsterdam\.Cited by:[§2\.1](https://arxiv.org/html/2605.07170#S2.SS1.p3.1)\.
- Pragglejaz Group \(2007\)MIP: a method for identifying metaphorically used words in discourse\.Metaphor and Symbol22\(1\),pp\. 1–39\.Cited by:[§1](https://arxiv.org/html/2605.07170#S1.p1.1),[§2\.1](https://arxiv.org/html/2605.07170#S2.SS1.p1.1)\.
- Y\. Shao, L\. Liu, Y\. Lan, L\. Wang, and T\. Zhao \(2024\)CMDAG: a Chinese metaphor dataset with annotated grounds for metaphor generation\.InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation \(LREC\-COLING\),Cited by:[§2\.1](https://arxiv.org/html/2605.07170#S2.SS1.p2.1),[§5\.2](https://arxiv.org/html/2605.07170#S5.SS2.p1.1)\.
- G\. J\. Steen, A\. G\. Dorst, J\. B\. Herrmann, A\. A\. Kaal, T\. Krennmayr, and T\. Pasma \(2010\)A method for linguistic metaphor identification: from MIP to MIPVU\.John Benjamins,Amsterdam\.Cited by:[§1](https://arxiv.org/html/2605.07170#S1.p1.1),[§2\.1](https://arxiv.org/html/2605.07170#S2.SS1.p1.1),[§3\.1](https://arxiv.org/html/2605.07170#S3.SS1.p1.1)\.
- C\. Su, F\. Fukumoto, X\. Huang, J\. Li, R\. Wang, and Z\. Chen \(2020\)DeepMet: a reading comprehension paradigm for token\-level metaphor detection\.InProceedings of the Second Workshop on Figurative Language Processing,pp\. 30–39\.Cited by:[§1](https://arxiv.org/html/2605.07170#S1.p2.1),[§2\.2](https://arxiv.org/html/2605.07170#S2.SS2.p1.1)\.
- B\. P\. Wang, X\. Lu, C\. Hsu, E\. P\. Lin, and H\. Ai \(2019\)Metaphor identification in chinese\.InMetaphor Identification in Multiple Languages: MIPVU Around the World,S\. Nacey, L\. Greve, A\. Dorst, and T\. Krennmayr \(Eds\.\),Cited by:[§2\.1](https://arxiv.org/html/2605.07170#S2.SS1.p3.1)\.
- S\. Zhang, Y\. Liu, and Y\. Ma \(2021\)SaGE: syntax\-aware GCN with ELECTRA for chinese metaphor detection\.InProceedings of the 20th Chinese National Conference on Computational Linguistics,Huhhot, China,pp\. 667–677\.External Links:[Link](https://aclanthology.org/2021.ccl-1.60/)Cited by:[§2\.2](https://arxiv.org/html/2605.07170#S2.SS2.p3.1)\.