Paper deep dive
DALDALL: Data Augmentation for Lexical and Semantic Diverse in Legal Domain by leveraging LLM-Persona
Janghyeok Choi, Jaewon Lee, Sungzoon Cho
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 97%
Last extracted: 3/26/2026, 1:38:22 AM
Summary
DALDALL is a persona-based data augmentation framework designed for legal information retrieval (IR). It leverages LLMs to generate synthetic queries from diverse professional perspectives (e.g., attorneys, judges, prosecutors) to address data scarcity in low-resource legal domains. Experiments on CLERC and COLIEE benchmarks demonstrate that DALDALL significantly improves lexical and semantic diversity compared to vanilla prompting, leading to superior or competitive performance in dense retrieval tasks.
Entities (5)
Relation Signals (4)
DALDALL â evaluatedon â COLIEE
confidence 100% ¡ Experiments on the CLERC and COLIEE benchmarks demonstrate that persona-based augmentation achieves improvement
DALDALL â evaluatedon â CLERC
confidence 100% ¡ Experiments on the CLERC and COLIEE benchmarks demonstrate that persona-based augmentation achieves improvement
DALDALL â uses â LLM
confidence 98% ¡ our method employs domain-specific professional personas... to generate synthetic queries
DALDALL â augments â Legal IR
confidence 95% ¡ we introduce DALDALL, a persona-based data augmentation framework tailored for legal information retrieval (IR).
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Data scarcity remains a persistent challenge in low-resource domains. While existing data augmentation methods leverage the generative capabilities of large language models (LLMs) to produce large volumes of synthetic data, these approaches often prioritize quantity over quality and lack domain-specific strategies. In this work, we introduce DALDALL, a persona-based data augmentation framework tailored for legal information retrieval (IR). Our method employs domain-specific professional personas--such as attorneys, prosecutors, and judges--to generate synthetic queries that exhibit substantially greater lexical and semantic diversity than vanilla prompting approaches. Experiments on the CLERC and COLIEE benchmarks demonstrate that persona-based augmentation achieves improvement in lexical diversity as measured by Self-BLEU scores, while preserving semantic fidelity to the original queries. Furthermore, dense retrievers fine-tuned on persona-augmented data consistently achieve competitive or superior recall performance compared to those trained on original data or generic augmentations. These findings establish persona-based prompting as an effective strategy for generating high-quality training data in specialized, low-resource domains.
Tags
Links
- Source: https://arxiv.org/abs/2603.22765v1
- Canonical: https://arxiv.org/abs/2603.22765v1
Trouble viewing inline? Open PDF directly â
Full Text
54,827 characters extracted from source content.
Expand or collapse full text
DALDALL: Data Augmentation for Lexical and Semantic Diverse in Legal Domain by leveraging LLM-Persona Janghyeok Choi 1 1 * Jaewon Lee 2 1 * Sungzoon Cho 3 1 Abstract Data scarcity remains a persistent challenge in low-resource domains. While existing data aug- mentation methods leverage the generative ca- pabilities of large language models (LLMs) to produce large volumes of synthetic data, these approaches often prioritize quantity over qual- ity and lack domain-specific strategies. In this work, we introduce DALDALL, a persona-based data augmentation framework tailored for legal information retrieval (IR). Our method employs domain-specific professional personasâsuch as attorneys, prosecutors, and judgesâto generate synthetic queries that exhibit substantially greater lexical and semantic diversity than vanilla prompt- ing approaches. Experiments on the CLERC and COLIEE benchmarks demonstrate that persona- based augmentation achieves improvement in lex- ical diversity as measured by Self-BLEU scores, while preserving semantic fidelity to the original queries. Furthermore, dense retrievers fine-tuned on persona-augmented data consistently achieve competitive or superior recall performance com- pared to those trained on original data or generic augmentations. These findings establish persona- based prompting as an effective strategy for gen- erating high-quality training data in specialized, low-resource domains. 1. Introduction Information retrieval (IR) in low-resource domains has long suffered from data scarcity (Fadaee et al., 2017; Fang et al., 2022). The advent of large language models (LLMs) such as GPT (Brown et al., 2020) has provided a highly convenient means of generating augmented datasets (Anaby-Tavor et al., 2020; Dai et al., 2025), leading to numerous approaches that leverage the generative capabilities of LLMs to cre- * Equal contribution 1 Department of Industrial Engineering, Seoul National University, Seoul, South Korea. Correspondence to: Sungzoon Cho <zoon@snu.ac.kr>. Preprint. March 25, 2026. ate AI-labeled data from limited seed examples. Among these, few-shot prompting is widely used (Bonifacio et al., 2022; Dai et al., 2022; Jeronymo et al., 2023; Krastev et al., 2025), where seed data and their labels serve as in-context examples within a prompt. However, few-shot examples bias the model toward reusing vocabulary from the pro- vided examples, reducing the diversity of augmentations. This raises concerns about the quality of augmented data in low-resource domains (Rachmat et al., 2025). Furthermore, while many data augmentation methods have demonstrated the ability to generate large volumes of synthetic samples, domain-specific approachesâparticularly for low-resource domainsâremain largely underexplored (Jeong et al., 2025; Seo et al., 2024). In such settings, generating high-quality data is often more important than producing large quantities of mediocre samples, especially in specialized domains such as law and medicine. In this work, we focus on the legal domain. Legal IR is chal- lenging because of the limited availability of high-quality annotated data (Hong et al., 2020; Kim et al., 2025). To address this challenge, we employ persona-based prompting to generate high-quality augmentations. By assigning a spe- cific role or persona to the modelâfor example, simulating a clientâs behavior by adopting their perspectiveâLLMs can exhibit more accurate and structured reasoning compared to generic prompting (Luo et al., 2024). Prior work suggests that the role assigned to an LLM is a significant factor in shaping its outputs. Crucially, the legal domain features well-defined professional rolesâsuch as attorneys, pros- ecutors, and judgesâeach with distinct perspectives and linguistic patterns when formulating queries. This makes persona-based prompting a natural fit for generating diverse synthetic queries. As a low-resource domain, legal IR requires quality-focused approaches to augment datasets. We decompose quality in the IR context into two dimensions: lexical diversity and se- mantic diversity. Lexical diversity is important for sparse re- trieval: as many practical search systems still rely on sparse retrieval methods such as BM25. Dense retrievers benefit from exposure to diverse semantic patterns during training and are widely used for re-ranking documents retrieved by sparse retrievers. Based on this perspective, we propose 1 arXiv:2603.22765v1 [cs.CL] 24 Mar 2026 Submission and Formatting Instructions for ICML 2026 Figure 1. The DALDALL framework for persona-based data augmentation. Given a legal case statement, the Persona method (top) generates synthetic queries from multiple professional perspectives (attorney, prosecutor, judge, law professor), while the Vanilla method (bottom) does not incorporate persona information. Compared to Vanilla, Persona augmentation produces outputs with 20% greater lexical diversity (Table 1). a data augmentation framework that leverages the genera- tive capabilities of LLMs and introduces a persona-based augmentation strategy to induce substantial lexical diversity in legal queries (Figure 1). We analyze both the diversity characteristics and the downstream retrieval effectiveness of our approach. Our contributions are as follows. 1. We introduce a persona-based augmentation strategy using legal-domain-specific professional roles, which induces substantial lexical and semantic diversity, as measured by Self-BLEU scores and intra-cosine simi- larity. 2.We provide an empirical analysis of the conditions under which persona-based prompting yields diverse outputs and demonstrate its effectiveness through im- proved or maintained recall scores after fine-tuning multiple retrieval models. 2. Related Works 2.1. Legal Information Retrieval Legal IR focuses on finding relevant legal documentsâ such as case law, statutes, and precedentsâgiven a user query. Legal IR presents unique challenges that distinguish it from general-domain retrieval. First, legal case state- ments are often extremely long, frequently exceeding the context limits of commonly used retrieval models (Nguyen et al., 2025). This necessitates summarization or chunking strategies (Tran et al., 2020) that may lose critical informa- tion. Second, legal language is highly specialized, featuring domain-specific terminology, complex sentence structures, and references to statutes and precedents that require expert understanding (Hong et al., 2020). Third, and most critically for this work, high-quality annotated datasets for legal IR are scarce, as creating such datasets requires expensive legal expertise. In response to these challenges, researchers have proposed various approaches. Dense embedding approaches have shown promising results by encoding semantic similarity. Other works focus on query reformulation (Kim et al., 2025) or leverage transfer learning from related domains (Fang et al., 2022). 2.2.LLM-based Data Augmentation and Persona-based Prompting 2.2.1. LLM-BASED DATA AUGMENTATION The emergence of large language models (LLMs) has trans- formed data augmentation by enabling high-volume gen- eration of synthetic training data. LLMs such as the GPT family (Brown et al., 2020) demonstrate remarkable capa- bilities to generate synthetic data of reasonable quality in large volumes (Anaby-Tavor et al., 2020). In the IR domain specifically, several works leverage LLMs to generate syn- thetic queries for documents. InPars (Bonifacio et al., 2022) pioneered this approach by using few-shot prompting to generate queries, followed by InPars-v2 (Jeronymo et al., 2023) and InPars+ (Krastev et al., 2025), which refined the generation and filtering pipeline. Promptagator (Dai et al., 2022) demonstrated that effective dense retrievers can be trained from as few as eight examples. AugGPT (Dai et al., 2025) further explored ChatGPTâs potential for general text data augmentation. In specialized domains such as law and medicine, generating high-quality data is often more important than producing 2 Submission and Formatting Instructions for ICML 2026 large volumes of mediocre samples (Rachmat et al., 2025). This has led to more sophisticated augmentation pipelines that incorporate retrieval-augmented generation (RAG) (Seo et al., 2024) or domain-specific filtering mechanisms. Given the importance of prompt design in LLM-based aug- mentation, we now turn to a specific prompting strategy that has shown particular promise: persona-based prompting. 2.2.2. PERSONA-BASED PROMPTING Prompt engineering has emerged as a critical technique for eliciting desired behaviors from LLMs. Research has shown that even minor modifications to prompts can lead to substantial performance improvements (Radford et al., 2021). This sensitivity makes prompt design a crucial factor in any LLM-based pipeline, including data augmentation. One influential direction in prompt engineering is persona- based prompting, which involves assigning a specific role or identity to the LLM before it performs a task. This finding suggests that personas can guide LLMs toward more structured and domain-appropriate outputs. The persona methodology has been extended to various ap- plications. Large-scale persona datasets, such as Persona Hub with one billion personas (Ge et al., 2024), enable di- verse synthetic data generation across domains. For data augmentation specifically, recent works show that persona- driven approaches can improve both the quality and diver- sity of generated text (Jeong et al., 2025). The Debate-to- Write framework (Hu et al., 2025) uses multiple personas to generate diverse arguments, demonstrating that different perspectives yield lexically and semantically varied out- puts. PersonaMath (Luo et al., 2024) demonstrated that LLMs achieve higher accuracy on mathematical problems when using persona-based data augmentation for the train- ing dataset. These findings are particularly relevant to legal IR. The le- gal domain features well-defined professional rolesâsuch as attorneys, prosecutors, and judgesâeach with distinct perspectives, objectives, and linguistic patterns when formu- lating queries. Our work addresses this gap by leveraging legal personas to induce lexical diversity in synthetic query generation. 2.3. Datasets for Legal Information Retrieval Several benchmark datasets have been developed to advance research in legal IR. In this section, we review two promi- nent datasets that are particularly relevant to our work: COL- IEE and CLERC. COLIEE.The Competition on Legal Information Extrac- tion/Entailment (Nguyen et al., 2025) is an annually held competition that has become a cornerstone benchmark for legal IR research. Now in its eleventh edition, COLIEE com- prises four primary tasks spanning case law and statute law. Task 1 focuses on legal case retrieval, where systems must identify relevant precedent cases (ânoticed casesâ) given a query case, which we employ for our seed data. Based on the COLIEE 2024 statistics, the dataset contains a total of 7,350 case law files, with 5,616 cases in the labeled training set, of which 1,278 serve as query cases. CLERC.The Case Law Evaluation and Retrieval Corpus (Hou et al., 2025) is a recently introduced large-scale dataset designed to support both legal case retrieval and retrieval- augmented generation tasks. Developed in collaboration with legal professionals, CLERC is built upon the Caselaw Access Project (CAP), which contains over 1.84 million U.S. federal case documents with more than 20.7 million citations. Unlike COLIEE, which uses case-to-case retrieval, CLERC frames the task as citation retrievalâfinding prece- dent cases that should be cited to support a given piece of legal analysis. 3. Method & Approach 3.1. DALDALL: Persona-Based Data Augmentation We propose DALDALL (Data Augmentation for Lexical and Semantic Diversity in the Legal Domain by Leverag- ing LLM-Persona), a data augmentation methodology tai- lored for legal IR. Given a query caseQand its associ- ated positive casesD 1 , D 2 , . . ., benchmark datasets such as CLERC and COLIEE provide labeled pairs of the form (Q,D 1 , D 2 , . . .). Our approach generates synthetic queriesA = f(Q, P)us- ing a promptPwhile preserving the original positive labels. For each queryQ, we employ two types of prompts. The first is the Vanilla PromptP V , which yields augmentation A V . The second is the Persona PromptP P , which yields augmentationA P . The key distinction is thatP P incorpo- rates persona information absent fromP V . In this work, we compareA P andA V in terms of lexical and seman- tic diversity as well as downstream retrieval performance. Through this comparison, we evaluate the effectiveness of persona-based augmentation, which we hypothesize to be more beneficial for low-resource domains than few-shot ap- proaches that generate synthetic queries by reusing vocabu- lary from the examples. For the overview of our experiment, see Figure 2. We use five personas for comparison and fine-tuning, as our ablation study (Section 5) demonstrates that this sub- set yields optimal performance. Full descriptions of the personas and prompt templates are provided in Appendix A. 3 Submission and Formatting Instructions for ICML 2026 Figure 2. Overview of our experiment. Left: Augmentation generation pipeline, where the original statement and extracted essentials are combined and processed through either Persona or Vanilla augmentation. Right: Evaluation framework measuring lexical diversity (Self-BLEU, BM25 Recall), semantic diversity (Dense Recall, Intra-cosine), and augmentation quality via fine- tuning. 3.2. Prompt Design We provide the full prompt templates in the Appendix A. The augmentation process comprises two stages: essential extraction and query augmentation. Stage 1: Essential Extraction. We first extract four es- sential components from the original query: legal issue, legal test or standard, key precedents, and key statutes or rules. These components correspond to the core elements of legal analysis as formalized in the IRAC framework (Issue, Rule, Application, and Conclusion), which is the standard methodology for legal reasoning. Recent work on rhetori- cal role segmentation in legal NLP has similarly identified these elements as semantically distinct units within legal documents (Malik et al., 2022; Bhattacharya et al., 2019). By extracting these invariant components before augmen- tation, we ensure that the generated text preserves the legal meaning of the source while allowing lexical and structural variation in non-essential elements (Shorten et al., 2021). Stage 2: Augmentation.Given the original text, extracted essentials, and an augmentation prompt, the LLM generates synthetic queries. We employ two types of prompts: ⢠Vanilla prompts provide augmentation instructions without additional context (Section A.2). â˘Persona prompts include instructions for the LLM to adopt a specific professional perspective (Section A.3). Generation Strategy.Each augmentation is generated in- dependently using a single prompt per query. We observed that generating all augmentations simultaneously with a single prompt resulted in lower diversity compared to gen- erating them individually. 3.3. Evaluation Metrics We evaluate augmented data along two complementary di- mensions: lexical diversity and semantic quality. Ad- ditionally, we assess downstream utility through retrieval performance. Lexical Diversity.We measure lexical diversity using two metrics: 1.Self-BLEU (Zhu et al., 2018). This metric measures similarity among generated samples; lower scores in- dicate greater diversity, which is desirable for training robust models. 2. BM25 recall comparison. We compare BM25 recall scores between augmented and original queries to ver- ify that augmented queries are not restricted to a narrow vocabulary. Semantic Diversity. We assess whether augmented queries preserve the meaning of the originals: 1.Intra-Cosine Similarity. We measure the cosine simi- larity among augmented queries, to ensure they span a diverse region of the semantic space, rather than clus- tering around a single representation. 2.Retrieval score consistency. We confirm that retrieval scores obtained with augmented queries are compara- ble to those of the originals, demonstrating effective preservation of semantic content. Downstream Retrieval Performance. We show that a retriever fine-tuned on augmented queries achieves perfor- mance equal to or better than one trained on original data alone. 3.4. Fine-tuning 3.4.1. DATASET PREPROCESSING We fine-tune retrievers on two datasets: CLERC and COL- IEE (see Table 7 in Appendix B for the number of training samples per configuration). These datasets differ signifi- cantly in text length, requiring different preprocessing ap- proaches. COLIEE. COLIEE contains full legal case statements with an average token count exceeding 6,000, requiring passage-based segmentation. We segment both documents and queries into passages. Segmenting the 1,000 origi- nal queries yields over 5,000 query-passage pairs, which matches the scale of the 5,000 augmentations per method. CLERC. Queries in CLERC average 347 tokens, well within the context window of most dense retrievers. We use the original queries and augmentations directly without segmentation. The training set consists of 1,000 original queries and 5,000 augmentations per method (Vanilla and Persona). 4 Submission and Formatting Instructions for ICML 2026 Figure 3. Triplet composition for fine-tuning on the COLIEE dataset. We chunk a query and positive documents, and find k-best semantic pair with base model. Ifnis lower thani, then all query chunks are used for the triplet. 3.4.2. PASSAGE-BASED RETRIEVAL To handle COLIEEâs long documents, we adopt a passage- based retrieval approach.Following (Nguyen et al., 2025), we consider three passage-based scoring approaches: FirstP, MaxP, and SumP. We adopt MaxP, which selects the highest similarity score among query-passage pairs (Q, P i ), whereP i denotes thei-th passage in the document. MaxP is particularly well-suited for legal IR because key le- gal reasoning is often localized within a specific paragraph, such as the holding or reasoning section of a judgment. In addition to segmenting documents, we also segment query cases into passages. ColBERT introduces the Late Interaction mechanism, which aggregates token-level simi- larities as follows: score(d) = X i max jâd sim(q i , d j )(1) A related variant, the Global Max approach, computes: score(d) = max i,j sim(q i , d j )(2) We adopt Global Max for evaluation, as preliminary exper- iments on the COLIEE test set using BGE-base-en-v1.5 with default configurations showed that it provides stable baseline recall performance. During evaluation on the COL- IEE dataset, to compute Recall@k, we calculate similarity scores between all query chunks and all document passages, selecting the globally maximum similarity pair for each query-document comparison. 3.4.3. TRIPLET CONSTRUCTION COLIEE. Each queryQis segmented intonchunks: Q 1 , Q 2 , . . . , Q n . For each query, there arempositive docu- mentsD 1 , . . . , D m andiaugmentationsA 1 , . . . , A i , where DatasetApproachToken Len. Avg.Self-BLEUIntra-Cos CLERCOriginal347.7â Persona124.830.3680.855 Vanilla86.140.4350.819 COLIEEOriginal6,379.8â Persona118.90.3180.778 Vanilla79.20.4140.799 Table 1. Lexical and semantic diversity metrics for persona-based and vanilla augmentations on CLERC and COLIEE. Each method generates 5,000 augmented queries per dataset. iâ3, 5, 7, 10in our experiments. Since each augmenta- tion is under 500 tokens, no further segmentation is required for the augmented queries. Prior to fine-tuning, we construct training triplets with two objectives: (1) ensuring that the number of triplets derived from original query chunks does not exceed the number of augmentation-based triplets, and (2) reducing training noise by excluding low-relevance chunk pairs. Specifically, we select the top-iquery-chunkâdocument-chunk pairs ranked by baseline dense retriever recall score. Ifn < i, all avail- able chunks are included without repetition, rather than artificially padding to reachi. The goal of this setup is not to demonstrate that the augmentation-based training labels outperform the original ones, but rather to verify that the augmentations do not distort the modelâs learned representationâthereby confirming that the persona-based augmentation approach contributes a meaningful, comple- mentary training signal. Generating the triplets for fine- tuning is illustrated in Figure 3. CLERC.Since CLERC queries require no segmentation, triplet construction is straightforward: each original query or augmentation is paired directly with its positive and negative documents. 3.4.4. TRAINING CONFIGURATIONS We evaluate five dense retrievers: DPR, LegalBERT, BGE- base-en-v1.5, multilingual-e5-base, and BGE-M3, along with BM25 as a sparse baseline. For each dense retriever, we consider six configurations: 1. Baseline: No fine-tuning 2. Original: Original Queries 3. Vanilla only: Vanilla augmentations 4. Vanilla mix: Original + vanilla augmentations 5. Persona only: Persona augmentations 6. Persona mix: Original + persona augmentations For âmixâ configurations, we randomly sample from the combined pool of original and augmented queries to main- tain the same training set size. 5 Submission and Formatting Instructions for ICML 2026 SectionPersonaVanilla Self-BLEU Intra-Cos Self-BLEU Intra-Cos 10.3210.7850.4630.818 20.3130.7780.4290.803 30.3000.7740.4210.795 40.3000.7700.4100.782 50.2990.7710.3950.781 60.2900.7690.3910.786 70.2820.7650.3790.779 Table 2. Diversity metrics by document length on COLIEE. Docu- ments are divided into seven sections of 1,000, ordered by token count (Section 1 = shortest). Lower Self-BLEU and Intra-Cosine scores indicate greater diversity. 4. Experiments & Analysis 4.1. Various Perspectives in seed data (CLERC vs COLIEE) We hypothesize that each persona maintains its own stand- point on a case, and that when there is insufficient content, the persona cannot generate diverse perspectives. To verify this hypothesis, we conduct augmentation experiments on two datasets: CLERC and COLIEE. And we use gpt-5-nano model for augmentation. A CLERC query contains only a small portion of the legal case passage, focusing particu- larly on the citation section. In contrast, a COLIEE query comprises the full case statement, encompassing various perspectives and providing richer context to the LLM. We select 1,000 queries from each dataset and generate five aug- mentations per query. The average token length of CLERC queries is 338 tokens, whereas COLIEE queries average 6,380 tokensâapproximately 19 times longer (Table 1). Lexical Diversity. Contrary to our hypothesis, persona- based augmentation consistently produces more lexically di- verse outputs than vanilla augmentation across both datasets, as measured by Self-BLEU scores (lower indicates greater diversity) (Table 1). On CLERC,P P achieves a Self-BLEU of 0.368 compared toP V âs 0.435, representing a 15% im- provement in diversity. On COLIEE, this gap widens further, with Persona at 0.318 versusP V at 0.414âa 23% improve- ment. This consistency across datasets with vastly different source text lengths (338 vs. 6,380 tokens) suggests that the diversity gains stem primarily from the introduction of distinct professional perspectives rather than from character- istics of the input text. Legal personas naturally emphasize different aspects of case documents, including lexical vari- ation thatP V prompting cannot achieve through simple paraphrasing. ApproachR@1R@5R@10R@20R@50 Original0.05400.17840.25720.33480.4481 Vanilla0.04380.14060.20920.28940.4035 Persona0.04100.13600.19730.26900.3729 Table 3. BM25 recall on COLIEE. 1,000 original queries are sampled, with 5 augmentations generated per query (5,000 total per method). ApproachR@1R@5R@10R@20R@50 Original0.02710.08870.12800.18270.2665 Vanilla0.02500.07080.10220.14380.2149 Persona0.02080.06040.08840.12410.1890 Table 4. Retrieval performance of BGE-base-en-v1.5 (non-fine- tuned) on COLIEE. Dataset sizes are identical to Table 4. Semantic Diversity.Intra-Cosine Similarity measures se- mantic variation among augmented queries, where lower scores indicate greater semantic diversity (Table 2). On CLERC,P P achieves a higher Intra-Cosine Similarity (0.855) thanP V (0.819), indicating that persona-based aug- mentations are semantically more similar to one another despite being lexically more diverse. On COLIEE, however, this pattern reverses:P P produces a lower Intra-Cosine Similarity (0.778) thanP V (0.799), demonstrating greater semantic diversity. We attribute this difference to the length of the source texts. Shorter CLERC queries provide limited semantic content for personas to differentiate, resulting in augmentations that vary in wording but converge in meaning. In contrast, longer COLIEE documents offer sufficient mate- rial for each persona to emphasize genuinely distinct aspects of the case, enabling both lexical and semantic diversifica- tion. This finding partially supports our hypothesis: input text richness influences semantic diversity, even though lex- ical diversity improvements remain consistent regardless of source length. 4.2. Token Distribution We further investigate the initial hypothesis that longer input texts can generate more diverse augmentations. Notably, a pattern emerges in the results of CLERC versus COLIEE: both Self-BLEU and Intra-Cosine Similarity scores for both P V and P P on COLIEE are better than on CLERC. The COLIEE dataset exhibits a wide token distribution across 7,350 texts. The 1,000th document contains 2,089 tokens, while the 7,000th document contains 15,798 tokens. We divide the dataset into seven sections, each containing 1,000 texts. Each section index represents a ranking range: Section 1 corresponds to documents ranked 1-1,000; Section 2 corresponds to ranks 1,001-2,000; and so forth. The more tokens the input texts contain, the lower the Self- 6 Submission and Formatting Instructions for ICML 2026 Table 5. Semantic similarity metrics for COLIEE augmentations. DatasetApproachAvgMean COLIEEPersona0.6970.843 Vanilla0.6820.840 BLEU scores, regardless of approach (Table 2). Lexical diversity steadily improves when the original text has am- ple material for augmentation. Additionally, all sections demonstrate that the Persona approach can generate more diverse augmentations by producing more unique tokens and exploiting a wider variety of words. 4.3. Augmentation Quality Our methodology transforms an original query into short augmentations that can be utilized by compact retrieval models such as DPR and BERT-based encoders. This is advantageous for IR because compact models remain widely adopted in this field due to their computational efficiency. However, we must demonstrate that the generated queries are sufficiently representative of the original query. This raises the following research questions: 1. Do the augmentations retain an appropriate degree of lexical overlap with the original queryâneither too narrow (under-retrieving relevant passages) nor too broad (over- retrieving irrelevant ones)? 2. Do the augmentations pre- serve semantic proximity to the original queryâs represen- tation? Do their distributional shifts effectively enhance retrieval performance? We conduct experiments to validate the quality of our aug- mentations on COLIEE with respect to these questions. 1.BM25 retrieval scores: We evaluate whether the aug- mentations preserve sufficient lexical overlap with pos- itive passages without excessive keyword dilution. 2.BGE-base retrieval scores: We assess whether the augmentations retain semantic proximity to the original query in the embedding space. 3. Fine-tuning and evaluation: We verify that the se- mantic shifts induced by the augmentations lead to meaningful improvements in retrieval performance. Lexical overlap decreases for both augmentation methods, with a more pronounced reduction in the Persona approach due to its generation of more lexically diverse terms (Table 5). Nevertheless, the difference from the original queries remains within approximately 20%, indicating that the aug- mentations preserve key lexical features adequately. This suggests that the augmented data can serve as a viable legal IR dataset with (augmented query, positive passage) pairs that are sufficiently challenging for model training. We analyze two aspects: BGE-base-en-v1.5 recall scores and semantic similarity metrics. Although the Persona and Vanilla methods exhibit similar semantic proximity to the original queries, the dense retrieval results reveal that their augmentations occupy distinct regions in the embedding space, suggesting that semantic similarity alone may not pre- dict retrieval performance (Table 5). To investigate whether the semantic space of Persona-augmented queries aligns bet- ter with the retrieverâs learned representations, we further examine fine-tuning performance. DPR and BGE-base-en-v1.5 achieve better performance (Table 9 in Appendix C) with Persona augmentations, even though the evaluation setup favors the Original method (i.e., chunking original case sentences and using them directly for retrieval). In contrast, LegalBERT, BGE-M3, and E5-base with Persona do not outperform the other methods. We hy- pothesize that LegalBERT, having been pre-trained on legal texts such as legislation, court cases, and contracts, is less compatible with the more informal and naturally phrased augmentations, which may interfere with its domain-specific representations. Additionally, BGE-M3 has a larger context window, which allows it to incorporate more content from the original cases; this may explain its higher recall scores compared to those of the Persona method. Finally, E5-base is a general-purpose model that may not be as effective for the legal domain as the specialized models, which could explain why there are no significant differences among the methods. Based on the above analysis, we conclude that the Per- sona method is not always the best option for fine-tuning and retrieval performance. However, the downstream re- trieval results demonstrate that the Persona method does not cause distribution shift in the models, except for LegalBERT, while successfully generating more diverse augmentations. We anticipate that our approach can be beneficial when com- bined with supplementary methodologies for low-resource domains. Full retrieval results are provided in Table 9 in Appendix C. 5. Ablation 5.1. How many personas are needed? We conduct an experiment to determine the optimal number of personas using the COLIEE dataset. We evaluate four persona set sizes (3, 5, 7, and 10), where each smaller set is a nested subset of the full 10-persona set. The composition of each set is detailed in the Appendix A.5. Several clear trends emerge as the number of personas varies (Figure 5). With more personas, Self-BLEU scores rise (indicating reduced lexical diversity), while both cosine similarity with original queries and intra-augmentation co- sine similarity decrease (indicating greater semantic diver- sity). Interestingly, the 3-persona configuration exhibits a 7 Submission and Formatting Instructions for ICML 2026 Figure 4. Distribution of augmented query lengths on COLIEE for 3, 5, 7, and 10 personas. Persona-based prompting yields a broader token distribution than vanilla prompting regardless of persona count. trade-off: it achieves the lowest Self-BLEU score and the largest Vanilla-Persona gap, yet its intra-augmentation co- sine similarity remains relatively high. This suggests that fewer personas produce lexically varied augmentations that nonetheless occupy a more constrained semantic region. The results do not reveal a single optimal number of per- sonas. We therefore select five personas as a practical com- promise based on the following reasoning: ⢠Substantial semantic diversity gain from 3 to 5 per- sonas â˘Sufficient Self-BLEU gap between Vanilla-Persona methods ⢠Using 5 personas is computationally cheaper than 7 or 10 while achieving competitive fine-tuning perfor- mance. 6. Limitations & Future Work Despite the growing availability of legal datasets in various languages, this work focuses exclusively on English datasets. Future research should investigate whether our methodology generalizes to other languages. Additionally, as we are not legal experts and could not afford professional legal annotators, we cannot verify that our augmentations are legally sound and reliable; however, retrieval performance remains competitive, suggesting that the augmentations are at least functionally useful. To mitigate this concern, we introduced the concept of âEssentialsâ to guide LLM-based augmentation and prevent alteration of key legal facts in the arguments. Furthermore, we did not explore how different combinations of personas with a fixed set size affect the metrics. Our current findings suggest that the number of Figure 5. Effect of persona count on augmentation diversity (COL- IEE). Top: Self-BLEU scores (lower = more lexically diverse). Middle: cosine similarity with original queries. Bottom: intra- augmentation cosine similarity (lower = more semantically di- verse). Increasing persona count reduces lexical diversity but improves semantic diversity. Five personas offer a balanced trade- off. personas influences lexical and semantic diversity; however, future work should examine which specific personas are most effective in the legal domain. 7. Conclusion In this work, we presented DALDALL, a data augmentation methodology that enhances lexical and semantic diversity in the legal IR domain by leveraging the generative capabil- ities of LLMs. We demonstrated that LLM-based personas significantly affect augmentation diversity, with particular effectiveness in the legal domain. Additionally, we eval- uated the effectiveness of our augmentation approach by fine-tuning retrievers on the augmented data. We believe this work provides a practical guide for generating augmen- tations in low-resource domains. 8 Submission and Formatting Instructions for ICML 2026 References Anaby-Tavor, A., Carmeli, B., Goldbraich, E., Kantor, A., Kour, G., Shlomov, S., Stern, A., and Zwerdling, N. Do not have enough data? Deep learning to the rescue! In Proceedings of the AAAI Conference on Artificial Intelli- gence, volume 34, p. 7383â7390, 2020. Bhattacharya, P., Paul, S., Ghosh, K., Ghosh, S., and Wyner, A. Identification of rhetorical roles of sentences in Indian legal judgments. In Araszkiewicz, M. and Rodr Ě Äąguez- Doncel, V. (eds.), Legal Knowledge and Information Sys- tems: JURIX 2019, volume 322 of Frontiers in Artificial Intelligence and Applications, p. 3â12. IOS Press, 2019. doi: 10.3233/FAIA190297. Bonifacio, L., Abonizio, H., Fadaee, M., and Nogueira, R. InPars: Data augmentation for information retrieval using large language models. arXiv preprint arXiv:2202.05144, 2022. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., and Amodei, D. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, p. 1877â 1901, 2020. Dai, H., Liu, Z., Liao, W., Huang, X., Cao, Y., Wu, Z., Zhao, L., Xu, S., Liu, W., Liu, N., Li, S., Zhu, D., Cai, H., Sun, L., Li, Q., Shen, D., Liu, T., and Li, X. AugGPT: Leveraging ChatGPT for text data augmentation. IEEE Transactions on Big Data, 2025. Dai, Z., Zhao, V. Y., Ma, J., Luan, Y., Ni, J., Lu, J., Bakalov, A., Guu, K., Hall, K. B., and Chang, M.-W. Promptagator: Few-shot dense retrieval from 8 examples. arXiv preprint arXiv:2209.11755, 2022. Fadaee, M., Bisazza, A., and Monz, C. Data augmenta- tion for low-resource neural machine translation. arXiv preprint arXiv:1705.00440, 2017. Fang, J., Li, X., and Liu, Y. Low-resource similar case matching in legal domain. In Pimenidis, E., Angelov, P., Jayne, C., Papaleonidas, A., and Aydin, M. (eds.), Arti- ficial Neural Networks and Machine Learning â ICANN 2022, volume 13530 of Lecture Notes in Computer Science, p. 571â582, Cham, 2022. Springer.doi: 10.1007/978-3-031-15931-247. Ge, T., Chan, X., Wang, X., Yu, D., Mi, H., and Yu, D. Scal- ing synthetic data creation with 1,000,000,000 personas. arXiv preprint arXiv:2406.20094, 2024. Hong, Z., Zhou, Q., Zhang, R., Li, W., and Mo, T. Legal feature enhanced semantic matching network for similar case matching. In 2020 International Joint Conference on Neural Networks (IJCNN), p. 1â8, Glasgow, UK, July 2020. IEEE. doi: 10.1109/IJCNN48605.2020.9207528. Hou, A. B., Weller, O., Qin, G., Yang, E., Lawrie, D., Holzenberger, N., Zotkina, E., and Van Durme, B. CLERC: A dataset for legal case retrieval and retrieval- augmented analysis generation. In Findings of the Asso- ciation for Computational Linguistics: NAACL 2025, p. 7898â7913, April 2025. Hu, Z., Chan, H. P., Li, J., and Yin, Y. Debate-to-write: A persona-driven multi-agent framework for diverse ar- gument generation. In Proceedings of the 31st Inter- national Conference on Computational Linguistics, p. 4689â4703, January 2025. Jeong, H. S., Ko, H. K., Park, S. Y., and Kim, T. H. LLM- based persona-driven text data augmentation. IEEE Ac- cess, 2025. Jeronymo, V., Bonifacio, L., Abonizio, H., Fadaee, M., Lotufo, R., Zavrel, J., and Nogueira, R. InPars-v2: Large language models as efficient dataset generators for in- formation retrieval. arXiv preprint arXiv:2301.01820, 2023. Kim, D., Kang, D., Kim, J., Ryu, S., and Lee, G. GuRE: Generative query rewriter for legal passage retrieval. In Proceedings of the Natural Legal Language Processing Workshop 2025, p. 424â438, Suzhou, China, November 2025. Association for Computational Linguistics. Krastev, M., Hamar, M., Toapanta, D., Brouwers, J., and Lei, Y. InPars+: Supercharging synthetic data genera- tion for information retrieval systems. arXiv preprint arXiv:2508.13930, 2025. Luo, J., Chen, L., Luo, R., Zhu, L., Ao, C., Li, J., Ren, Z., and Yang, M. PersonaMath: Boosting mathematical reasoning via persona-driven data augmentation. arXiv preprint arXiv:2410.01504, 2024. Malik, V., Sanjay, R., Guha, S. K., Hazarika, A., Nigam, S. K., Bhattacharya, A., and Modi, A. Semantic segmen- tation of legal documents via rhetorical roles. In Proceed- ings of the Natural Legal Language Processing Workshop 2022, p. 153â171, Abu Dhabi, UAE, December 2022. Association for Computational Linguistics. URLhttps: //aclanthology.org/2022.nllp-1.13. Nguyen, H.-T., Nguyen, T.-M., Le, X.-B., Le, T.-K., Nguyen, K.-H., Nguyen, H.-T., Vuong, T.-H.-Y., and Nguyen, L.-M. NOWJ@COLIEE 2025: A multi-stage 9 Submission and Formatting Instructions for ICML 2026 framework integrating embedding models and large lan- guage models for legal retrieval and entailment. In Pro- ceedings of the Twentieth International Conference on Artificial Intelligence and Law (ICAIL â25), p. 506â515, 2025. doi: 10.1145/3769126.3785016. Rachmat, B. K., Gerald, T., Slb, Z. Z., and Grouin, C. QA analysis in medical and legal domains: A survey of data augmentation in low-resource settings. In Proceedings of the 63rd Annual Meeting of the Association for Compu- tational Linguistics (Volume 4: Student Research Work- shop), p. 1132â1144, July 2025. Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, p. 8748â 8763. PMLR, July 2021. Seo, M., Baek, J., Thorne, J., and Hwang, S. J. Retrieval- augmented data augmentation for low-resource domain tasks. arXiv preprint arXiv:2402.13482, 2024. Shorten, C., Khoshgoftaar, T. M., and Furht, B. Text data augmentation for deep learning. Journal of Big Data, 8 (1):101, 2021. doi: 10.1186/s40537-021-00492-0. Tran, V., Le Nguyen, M., Tojo, S., and Satoh, K. En- coded summarization: Summarizing documents into continuous vector space for legal case retrieval. Ar- tificial Intelligence and Law, 28:441â467, 2020. doi: 10.1007/s10506-020-09262-4. Zhu, Y., Lu, S., Zheng, L., Guo, J., Zhang, W., Wang, J., and Yu, Y. Texygen: A benchmarking platform for text generation models. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, p. 1097â1100, June 2018. 10 Submission and Formatting Instructions for ICML 2026 A. Prompt Templates This appendix provides the complete prompt templates used in our data augmentation pipeline. Section A.1 presents the prompt for extracting legal essentials, Section A.2 shows the base augmentation template shared by both vanilla and per- sona prompts, and Section A.3 details the persona-specific extensions. Section A.4 describes each persona, and Sec- tion A.5 specifies the persona combinations used in our experiments. A.1. Extract Essentials Prompt The following prompt (Figure 6) instructs the model to extract the invariant semantic core from legal text, which serves as the foundation for semantically consistent augmen- tation. A.2. Augmentation Prompt (Base Template) Both vanilla and persona-based augmentation use the base prompt template shown in Figure 7. The persona prompt extends this template with additional persona-conditioning instructions (Section A.3). A.3. Persona-Specific Extensions For persona-based augmentation, we extend the base tem- plate with the modifications shown in Figure 8. The system instruction is updated to indicate persona conditioning, a Persona Rules section is inserted after the task description, and the closing instruction is modified to emphasize persona- based diversity. A.4. Persona Descriptions We define ten legal personas, each characterized by distinct voice, orientation, style, and constraints. Detailed descrip- tions follow. Appellate Judge (Majority).Voice/Tone: formal, author- itative, doctrinally precise. Orientation: emphasizes legal standards, precedent coherence, and institutional stability. Style: neutral judicial prose, structured reasoning, succinct case citations. Key features: frames the rule clearly; applies the legal test systematically; expresses conclusions with institutional confidence; avoids emotional or argumentative rhetoric. Appellate Judge (Dissenting).Voice/Tone: assertive, crit- ical, more rhetorical than majority. Orientation: challenges the majorityâs application of law; stresses fairness or doc- trinal risk. Style: sharper transitions, explicit disagreement, highlights consequences of the rule. Key features: points out flaws in reasoning; stresses competing precedent or al- ternative interpretations; uses expressive but still judicial language. Prosecutor. Voice/Tone:assertive,confident, enforcement-focused.Orientation:public safety, rule-of-law, strong interpretation of statutes and precedent. Style: highlights factual elements that justify government action; emphasizes culpability or legitimacy of state conduct. Key features: stresses why legal standards support the governmentâs position; uses persuasive tone within legal boundaries; underscores societal interest or procedural integrity. Defense Attorney. Voice/Tone:protective, rights- centered, adversarial where necessary. Orientation: fairness, burden of proof, procedural safeguards, statutory protec- tion. Style: emphasizes mitigating facts, narrow readings of precedent, constitutional concerns. Key features: high- lights overreach, improper inference, or government burden; frames facts in a defendant-favorable way; uses careful legal language. Law Professor. Voice/Tone: analytical, conceptual, ex- planatory. Orientation: doctrine, theory, policy implications. Style: abstract reasoning, comparative references to broader jurisprudence. Key features: explains legal standards in a teaching tone; frames issues in terms of doctrinal evolu- tion; uses academic transitions (âconceptually,â âdoctrinally,â âhistoricallyâ). Trial Judge. Voice/Tone: pragmatic, procedural, fact- sensitive. Orientation: case management, evidentiary suffi- ciency, application of law to record facts. Style: grounded judicial prose, attentive to procedural posture and standards of review. Key features: begins from the concrete facts or procedural setting; emphasizes admissibility, burdens, trial- level reasoning; avoids broad doctrinal exposition unless necessary. Concurring Judge. Voice/Tone: formal, reflective, an- alytically distinct. Orientation: agrees with the outcome but through different reasoning. Style: judicial prose that reframes or narrows the doctrinal basis. Key features: explic- itly aligns with the judgment, not necessarily the reasoning; highlights alternative legal rationale or limiting principles; maintains institutional tone without dissenting rhetoric. Public Defender.Voice/Tone: empathetic, rights-focused, institutionally critical. Orientation: systemic fairness, in- equality of resources, constitutional protection.Style: advocacy-oriented but restrained; foregrounds procedural justice. Key features: emphasizes power imbalance and due process; frames legal standards defensively and narrowly; stresses safeguards against overreach. 11 Submission and Formatting Instructions for ICML 2026 You are a legal-analysis model. Your task is to extract the invariant semantic core of a legal text. The invariant core is the minimal set of legally essential propositions that must remain unchanged across all persona-based rewrites. STRICT CONSTRAINTS: - Use only information explicitly present or unambiguously implied by the text. - Do not infer motives, procedural posture, missing facts, or unstated legal theories. - Do not add legal doctrine beyond what the text names or quotes. - Be maximally concise; remove narrative, rhetoric, or stylistic detail. - When uncertain, use neutral phrasing ("the court indicates...", "the text states...") rather than assumptions. - Exclude material not necessary for legal invariance, such as: - dicta not tied to the holding - background facts irrelevant to the legal issue - commentary or speculation - redundant procedural descriptions - Before outputting JSON, internally verify that every element is text-supported and legally essential. Essentials to extract: These are the only fields that matter for semantic invariance: - legal_issue: the central legal question explicitly addressed. - legal_test_or_standard: any doctrinal test or rule stated or quoted. - key_precedents: cited cases only (no summaries unless text gives one). - key_statutes_or_rules: cited statutes or rules only. Below are examples. (few-shot examples omitted for brevity) Figure 6. Extract Essentials prompt template. Legal Realist Scholar. Voice/Tone: skeptical, analyti- cal, outcome-aware. Orientation: practical effects, judicial behavior, law-in-action. Style: academic but critical; de- emphasizes formal doctrine in favor of consequences. Key features: questions how rules operate in practice; highlights incentives, discretion, and institutional behavior; reframes issues in functional rather than formal terms. Judicial Clerk. Voice/Tone: neutral, precise, synthesis- oriented. Orientation: issue-spotting, clarity, internal con- sistency. Style: clean, structured summaries; balanced pre- sentation of arguments. Key features: reframes the ques- tion crisply; organizes issues logically; avoids advocacy or rhetorical flourish. A.5. Persona Combinations Table 6 shows which personas are included in each set size used in our ablation study. B. Fine-tuning Configurations Table 7 summarizes the number of training samples per configuration. Table 8 presents the hyperparameters used Table 6. Persona inclusion by set size. PersonaSet 3Set 5Set 7Set 10 Defense Attorneyâ Prosecutorâ Appellate Judge (Majority) â Appellate Judge (Dissenting) â Law Professorâ Trial Judgeâ Public Defenderâ Legal Realist Scholarâ Judicial Clerkâ Concurring Judgeâ for fine-tuning and the chunking configurations for each retrieval model. C. Retrieval Results Table 9 presents the complete recall scores for all fine-tuned retrievers on the COLIEE and CLERC datasets. 12 Submission and Formatting Instructions for ICML 2026 Table 7. Number of training samples per configuration. ConfigurationCLERCCOLIEE Original1,000 âź5,000 (chunked) Vanilla only5,0005,000 Vanilla mix5,0005,000 Persona only5,0005,000 Persona mix5,0005,000 Table 8. Training hyperparameters and chunking configurations. CategoryParameterValue Training Seed42 Epochs3 Learning rate2e-5 Warmup ratio0.1 Temperature0.05 Chunking BGE-base512 / 80 BGE-m38192 / 512 E5-base512 / 80 LegalBERT512 / 80 DPR512 / 80 Note: Chunking values are shown as chunk size / overlap (in tokens). 13 Submission and Formatting Instructions for ICML 2026 ModelApproachCOLIEECLERC R@1R@5 R@10 R@20 R@1R@5 R@10 R@20 BGE-base-en-v1.5 Baseline0.076 0.2280.3160.4130.586 0.7820.8260.880 Original0.0750.2470.3610.4700.718 0.9000.9240.962 Persona Only 0.0750.2630.3750.4970.688 0.8800.9160.952 Persona Mix0.0750.2630.3750.4980.718 0.9020.9260.964 Vanilla Only0.074 0.2550.3640.4750.702 0.8960.9260.962 Vanilla Mix0.074 0.2560.3720.4770.7160.8980.9280.962 DPRBaseline0.058 0.1890.2640.3300.548 0.7620.8220.874 Original0.064 0.2140.3220.4310.6040.8200.8760.906 Persona Only 0.072 0.226 0.3370.4390.614 0.8080.8800.924 Persona Mix0.0700.2260.3370.4390.594 0.7960.8660.916 Vanilla Only0.063 0.2280.3230.4280.6040.8260.8800.916 Vanilla Mix0.0700.2160.3190.4290.596 0.8080.8700.918 LegalBERTBaseline0.028 0.0830.1250.1840.322 0.5000.5880.680 Original0.062 0.2070.2820.3780.488 0.6960.7780.846 Persona Only 0.044 0.1430.2110.2880.458 0.6740.7380.804 Persona Mix0.0610.1740.2550.3390.394 0.5940.6760.740 Vanilla Only0.041 0.1270.1660.2300.442 0.6640.7420.808 Vanilla Mix0.040 0.1240.1710.2360.4780.6780.7600.828 E5-baseBaseline0.077 0.2350.3170.4050.688 0.8420.8940.932 Original0.091 0.2740.4010.5230.732 0.9020.9280.952 Persona Only 0.0920.2780.4130.5430.710 0.8920.9240.954 Persona Mix0.093 0.2780.4120.5420.7240.8940.9220.946 Vanilla Only0.0920.2870.4190.5430.722 0.9060.9300.952 Vanilla Mix0.0920.2930.4170.5440.718 0.8980.9220.944 BGE-M3Baseline0.044 0.1550.2440.3240.744 0.8920.9200.946 Original0.064 0.217 0.3010.4280.806 0.9340.9480.956 Persona Only 0.057 0.2160.310 0.4120.8000.9340.9500.962 Persona Mix0.055 0.2120.3180.4190.8000.9320.9540.966 Vanilla Only0.063 0.2180.2970.4090.794 0.9320.9500.958 Vanilla Mix0.0630.2100.3100.4230.798 0.9380.9560.962 Table 9. Recall scores of fine-tuned retrievers on COLIEE and CLERC. Best scores are in bold, and second-best scores are underlined. 14 Submission and Formatting Instructions for ICML 2026 You are an expert legal reasoning model generating COUNTERFACTUAL TEXT REWRITES for a legal information retrieval dataset. Your task is to produce augmentation_count COUNTERFACTUAL TEXT REWRITES: - Be an alternative way of asking about the * same * underlying legal situation. - Remain semantically equivalent with respect to Essentials (E). - Differ substantially from the Original Text in wording, syntax, and rhetorical framing. - Differ from the other rewritten texts to promote lexical and structural diversity. Essentials (E) are the invariant semantic core. They MUST remain unchanged. ### Diversity Requirements Across All augmentation_count Rewrites Across the augmentation_count augmented texts, you MUST: - Use at least three different sentence structures: - e.g., a single complex sentence; two short sentences; a "Under what circumstances..." style question. - NOT start more than one augmented text with the same first three words. - Re-order clauses or issues differently across rewrites (some standard-first, some fact-first). Lexical Diversity Rules: - Avoid reusing long spans (more than 5 consecutive words) from the Original Text, except for legally indispensable terms (case names, statutes, doctrine labels). - For each augmented text, use at least 3 paraphrases or synonyms for non-technical phrases in the Original Text (e.g., "affect the outcome" -> "alter the result"). ### Strict Requirements You MUST: - Preserve all Essentials (E) exactly; do NOT change their meaning. - Ensure each rewritten text would retrieve the same case/passage as the original. - Treat this as COUNTERFACTUAL DATA GENERATION: - Maximize lexical and structural difference from the Original Text. - Change word choice, sentence structure, and rhetorical focus. - Avoid copying phrases from the Original Text unless they are legally indispensable terms. - Maintain legal correctness; do NOT invent new facts, issues, or rules. You MAY: - Re-order clauses or issues. - Foreground or background different parts of Essentials (E). - Compress or paraphrase non-essential details, so long as you do not contradict Essentials (E). ### NOW PERFORM THE TASK Essentials (E): essentials Original Text: text Generate augmentation_count COUNTERFACTUAL TEXT REWRITES that preserve Essentials (E) exactly, use substantially different phrasing, and also differ meaningfully from one another. Figure 7. Base augmentation prompt template used by both vanilla and persona-based methods. 15 Submission and Formatting Instructions for ICML 2026 # Modified system instruction: You are an expert legal reasoning model generating persona-conditioned COUNTERFACTUAL TEXT REWRITES for a legal information retrieval dataset. Your task is to produce augmentation_count COUNTERFACTUAL TEXT REWRITES, each written from the perspective of a different TARGET PERSONA. [...] # New section inserted after task description: ### Persona Rules You will produce one rewritten text per persona listed below. Each persona must have a clearly unique tone, rhetorical style, and framing. PERSONAS: persona_dict For each persona P: - Write ONLY in the style of persona P. - Avoid copying tone, rhetorical patterns, or stylistic decisions from any other persona. - Ensure strong stylistic separation across personas. - Still preserve Essentials (E) exactly. # Modified closing instruction: Produce augmentation_count COUNTERFACTUAL TEXT REWRITES, one for each persona, ensuring strong stylistic diversity across personas while preserving Essentials (E) exactly. Figure 8. Persona-specific extensions to the base prompt. Comments (lines starting with #) indicate where each modification is applied. 16