scripts/gen_corpus.py

"""Generate the synthetic source corpus (simulating tiered internet sources)
and four test submissions with known ground truth:
  clean.txt  -- original paper on a topic absent from the corpus
  clone.txt  -- lit review + methodology copied verbatim from a tier-1 source
  mosaic.txt -- lit review woven from two IoT sources with synonym tweaks
  idea.txt   -- same study as a source, fully re-rendered (idea plagiarism)
"""
import json
import os
import sys

import numpy as np

sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from plagdetect.textutils import mosaic_mix, sentences, synonymize  # noqa: E402

ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
CORPUS_DIR = os.path.join(ROOT, "data", "corpus")
SUB_DIR = os.path.join(ROOT, "data", "submissions")

TOPICS = {
    "bh": dict(tech="consortium blockchain", domain="healthcare data management",
               entity="electronic health records", problem="patient privacy leakage",
               method="smart-contract based access control", tool="Hyperledger Fabric",
               metric="data breach incidence", adj="patient-centric"),
    "io": dict(tech="lightweight intrusion detection", domain="industrial IoT networks",
               entity="sensor telemetry streams", problem="botnet infiltration",
               method="federated anomaly detection", tool="edge gateway accelerators",
               metric="attack detection latency", adj="energy-aware"),
    "cr": dict(tech="multispectral deep learning", domain="precision agriculture",
               entity="UAV-captured field imagery", problem="late blight outbreaks",
               method="attention-based disease classification", tool="ResNet-50 backbones",
               metric="seasonal yield loss", adj="smallholder-friendly"),
    "ed": dict(tech="immersive virtual reality", domain="engineering education",
               entity="laboratory simulations", problem="declining skill transfer",
               method="haptic-guided procedural training", tool="Unity XR toolkit",
               metric="skill retention", adj="curriculum-aligned"),
    "tg": dict(tech="spatio-temporal graph networks", domain="urban traffic management",
               entity="loop detector measurements", problem="recurrent congestion spillback",
               method="graph attention flow forecasting", tool="adaptive signal controllers",
               metric="average commute delay", adj="city-scale"),
}
CLEAN_TOPIC = dict(tech="autonomous drone logistics", domain="last-mile parcel delivery",
                   entity="aerial delivery corridors", problem="payload routing inefficiency",
                   method="multi-agent route negotiation", tool="ROS2 flight stacks",
                   metric="delivery turnaround time", adj="weather-resilient")

SURNAMES = ["Sharma", "Chen", "Gupta", "Okafor", "Martinez", "Kim", "Patel",
            "Nguyen", "Silva", "Hassan", "Tanaka", "Kowalski", "Mehta", "Johansson"]
VENUES = ["IEEE Access", "ACM Computing Surveys", "Future Generation Computer Systems",
          "Journal of Network and Computer Applications", "Sensors",
          "IEEE Transactions on Industrial Informatics",
          "Expert Systems with Applications", "IEEE Internet of Things Journal"]

SECTIONS = ["abstract", "introduction", "literature_review", "methodology",
            "results", "conclusion"]

# Two disjoint template sets: P renders corpus papers, A renders the
# "idea plagiarism" re-write (same facts, different surface form).
P = {
    "abstract": [
        "This paper proposes a {adj} {tech} framework for {domain}.",
        "We address {problem} by introducing {method} built on {tool}.",
        "Experiments on {n} cases show a {imp}% improvement in {metric}.",
        "The framework demonstrates an F1 score of {f1} with p = {p}.",
        "Our findings establish {tech} as a practical remedy for {problem}.",
    ],
    "introduction": [
        "The rapid growth of {entity} has made {problem} a central concern for {domain} [{r1}].",
        "Conventional safeguards struggle to keep pace with the scale of modern {entity} [{r2}].",
        "Recent advances in {tech} offer a promising direction for {domain}.",
        "However, existing deployments rarely integrate {method} in an end-to-end manner.",
        "In this paper we design, implement, and evaluate a {adj} pipeline that couples {tech} with {method}.",
        "Our contributions include a reference architecture on {tool} and an evaluation across {n} cases.",
        "The remainder of this paper reviews prior work, details the methodology, and reports results.",
    ],
    "literature_review": [
        "Early studies of {tech} focused primarily on feasibility within {domain} [{r1}].",
        "Subsequent work between {y0} and {y1} shifted attention toward {method} [{r2}].",
        "Several authors report that {problem} remains the dominant failure mode in production systems [{r3}].",
        "Benchmark efforts using {tool} demonstrate significant gains in {metric} [{r4}].",
        "A recurring limitation across these studies is the absence of {adj} evaluation protocols.",
        "Surveys of {domain} consistently identify {entity} as the most vulnerable asset class [{r1}].",
        "Despite this progress, no prior work unifies {tech} and {method} under realistic workloads.",
    ],
    "methodology": [
        "Our research design follows a controlled experimental protocol over {n} cases collected from {y0} to {y1}.",
        "The core of the system is {method} implemented on {tool}.",
        "Each instance of {entity} is normalized, segmented, and assigned a cryptographic provenance tag.",
        "We configure the {tech} layer with a three-stage validation pipeline to suppress {problem}.",
        "Hyper-parameters were selected by grid search, holding out 20% of cases for validation.",
        "Statistical significance is assessed with paired tests at alpha = 0.05.",
        "All experiments are repeated five times and we report the mean with confidence intervals.",
    ],
    "results": [
        "The proposed framework improves {metric} by {imp}% relative to the strongest baseline.",
        "Across {n} cases the system attains an F1 score of {f1} with p = {p}.",
        "Ablation shows that removing {method} degrades performance by {imp2}%.",
        "Latency overhead introduced by the {tech} layer remains below acceptable operational thresholds.",
        "These results confirm that {adj} integration of {tool} is feasible at scale.",
        "Error analysis attributes most residual failures to noisy {entity}.",
    ],
    "conclusion": [
        "We presented a {adj} {tech} framework that mitigates {problem} in {domain}.",
        "Evaluation over {n} cases demonstrated a {imp}% improvement in {metric}.",
        "Future work will extend {method} to cross-organizational settings.",
        "We will also study the long-term governance of {entity} under this architecture.",
    ],
}

A = {
    "abstract": [
        "A {adj} architecture leveraging {tech} is introduced for {domain}.",
        "To counter {problem}, the design embeds {method} on top of {tool}.",
        "An evaluation spanning {n} cases records {imp}% better {metric}.",
        "F1 reaches {f1} and the effect is significant at p = {p}.",
        "The evidence positions {tech} as a credible answer to {problem}.",
    ],
    "introduction": [
        "{problem} has become unavoidable as {entity} proliferate across {domain} [{r1}].",
        "Legacy protections were never engineered for today's volume of {entity} [{r2}].",
        "{tech} has recently matured into a viable foundation for {domain}.",
        "Yet end-to-end adoption of {method} remains rare in deployed systems.",
        "Here we build and assess a {adj} stack uniting {tech} with {method}.",
        "We contribute a {tool} reference design plus a study of {n} cases.",
        "The paper proceeds through related work, methods, and findings.",
    ],
    "literature_review": [
        "Initial investigations asked whether {tech} could work at all inside {domain} [{r1}].",
        "From {y0} to {y1}, the community pivoted toward {method} [{r2}].",
        "Multiple groups observe that {problem} still dominates operational incidents [{r3}].",
        "Evaluations built on {tool} report meaningful movement in {metric} [{r4}].",
        "What these efforts share is a lack of {adj} testing regimes.",
        "Review articles repeatedly single out {entity} as the weakest link [{r1}].",
        "Still missing is a treatment that joins {tech} and {method} under production load.",
    ],
    "methodology": [
        "We run a controlled experiment on {n} cases gathered between {y0} and {y1}.",
        "{method} sits at the heart of the platform, realized with {tool}.",
        "Every {entity} record is cleaned, partitioned, and stamped with provenance metadata.",
        "A three-phase validation chain in the {tech} layer guards against {problem}.",
        "Grid search over a 20% held-out split fixes all hyper-parameters.",
        "Paired significance testing uses alpha = 0.05 throughout.",
        "Five repetitions per condition yield means and confidence bounds.",
    ],
    "results": [
        "{metric} improves by {imp}% over the best competing baseline.",
        "On {n} cases the platform reaches F1 = {f1} at p = {p}.",
        "Dropping {method} in ablation costs {imp2}% of performance.",
        "Overhead from the {tech} layer stays within operational budgets.",
        "{adj} deployment of {tool} therefore scales in practice.",
        "Most remaining errors trace back to corrupted {entity}.",
    ],
    "conclusion": [
        "This work delivered a {adj} {tech} stack that curbs {problem} in {domain}.",
        "A study of {n} cases yielded {imp}% gains in {metric}.",
        "Extending {method} across organizations is the next step.",
        "Governance of {entity} over long horizons also merits study.",
    ],
}


def sample_vals(rng):
    return dict(imp=int(rng.randint(38, 82)), imp2=int(rng.randint(9, 36)),
                n=int(rng.randint(120, 950)),
                p=round(float(rng.uniform(0.001, 0.04)), 3),
                f1=round(float(rng.uniform(0.81, 0.96)), 2),
                y0=int(rng.randint(2015, 2019)), y1=int(rng.randint(2019, 2023)),
                r1=int(rng.randint(1, 11)), r2=int(rng.randint(1, 11)),
                r3=int(rng.randint(1, 11)), r4=int(rng.randint(1, 11)))


def render_sections(terms, vals, rng, alt=False):
    tpl = A if alt else P
    mapping = {**terms, **vals}
    secs = {}
    for sec in SECTIONS:
        pool = tpl[sec]
        k = min(len(pool), 6)
        idx = rng.choice(len(pool), size=k, replace=False)
        secs[sec] = " ".join(pool[i].format(**mapping) for i in sorted(idx))
    return secs


def make_refs(terms, rng):
    titles = [
        "A survey of {tech} in {domain}", "{method}: opportunities and challenges",
        "Securing {entity} with {tech}", "Towards {adj} {domain}",
        "Mitigating {problem} using {method}", "{tool} in practice",
        "Benchmarking {metric} under {tech}", "Rethinking {problem}",
        "On the limits of {tech} for {entity}", "{domain}: a systematic review",
        "Adaptive {method} for dynamic environments",
        "Privacy and {entity}: an empirical study",
        "A framework for {adj} deployment of {tool}",
        "Deep evaluation of {method} at scale",
    ]
    refs = []
    for t in titles:
        a = str(rng.choice(SURNAMES))
        v = str(rng.choice(VENUES))
        y = int(rng.randint(2015, 2024))
        refs.append(f'{a[0]}. {a}, "{t.format(**terms).capitalize()}," {v}, {y}')
    return refs


def make_title(terms, rng, alt=False):
    forms = (["{adj} {tech} for {domain}: an empirical study",
              "Mitigating {problem} with {method}",
              "{tech} meets {domain}: design and evaluation"] if not alt else
             ["Rethinking {domain} through {adj} {tech}",
              "{method} against {problem}: a new architecture"])
    return str(rng.choice(forms)).format(**terms).capitalize()


def paper_txt(title, secs, refs):
    head = {"abstract": "Abstract", "introduction": "1. Introduction",
            "literature_review": "2. Literature Review",
            "methodology": "3. Methodology", "results": "4. Results",
            "conclusion": "5. Conclusion"}
    lines = [title, ""]
    for k in SECTIONS:
        lines += [head[k], secs[k], ""]
    lines.append("References")
    lines += [f"[{i + 1}] {r}." for i, r in enumerate(refs)]
    return "\n".join(lines)


def main():
    os.makedirs(CORPUS_DIR, exist_ok=True)
    os.makedirs(SUB_DIR, exist_ok=True)
    rng = np.random.RandomState(123)
    topic_refs = {k: make_refs(t, rng) for k, t in TOPICS.items()}
    forced_tiers = {"bh02": 1, "io01": 1, "io05": 2, "tg02": 1}
    corpus = {}

    for key, terms in TOPICS.items():
        for i in range(12):
            doc_id = f"{key}{i:02d}"
            vals = sample_vals(rng)
            secs = render_sections(terms, vals, rng, alt=False)
            refs = [str(r) for r in rng.choice(topic_refs[key], 9, replace=False)]
            other = str(rng.choice([k for k in TOPICS if k != key]))
            refs += [str(r) for r in rng.choice(topic_refs[other], 2, replace=False)]
            tier = forced_tiers.get(doc_id,
                                    int(rng.choice([1, 2, 3, 4],
                                                   p=[0.4, 0.3, 0.2, 0.1])))
            doc = {"doc_id": doc_id, "title": make_title(terms, rng),
                   "sections": secs, "references": refs, "tier": tier,
                   "year": int(rng.randint(2016, 2024)),
                   "topic": key, "vals": vals}
            corpus[doc_id] = doc
            with open(os.path.join(CORPUS_DIR, f"{doc_id}.json"), "w",
                      encoding="utf-8") as f:
                json.dump(doc, f, indent=1)

    # ---- submissions -------------------------------------------------------
    subs = {}

    vals = sample_vals(rng)
    secs = render_sections(CLEAN_TOPIC, vals, rng)
    subs["clean"] = paper_txt(make_title(CLEAN_TOPIC, rng), secs,
                              make_refs(CLEAN_TOPIC, rng)[:9])

    src = corpus["bh02"]
    vals = sample_vals(rng)
    secs = render_sections(TOPICS["bh"], vals, rng)
    secs["literature_review"] = src["sections"]["literature_review"]
    secs["methodology"] = src["sections"]["methodology"]
    refs = [str(r) for r in rng.choice(topic_refs["bh"], 8, replace=False)]
    subs["clone"] = paper_txt(make_title(TOPICS["bh"], rng), secs, refs)

    s1, s2 = corpus["io01"], corpus["io05"]
    vals = sample_vals(rng)
    secs = render_sections(TOPICS["io"], vals, rng)
    secs["literature_review"] = mosaic_mix(
        sentences(s1["sections"]["literature_review"]),
        sentences(s2["sections"]["literature_review"]), rng)
    secs["results"] = synonymize(s1["sections"]["results"], rng, p=0.5)
    refs = [str(r) for r in rng.choice(topic_refs["io"], 8, replace=False)]
    subs["mosaic"] = paper_txt(make_title(TOPICS["io"], rng), secs, refs)

    src = corpus["tg02"]
    rng_idea = np.random.RandomState(777)
    secs = render_sections(TOPICS["tg"], src["vals"], rng_idea, alt=True)
    refs = [str(r) for r in rng_idea.choice(src["references"], 6, replace=False)]
    subs["idea"] = paper_txt(make_title(TOPICS["tg"], rng_idea, alt=True),
                             secs, refs)

    for name, txt in subs.items():
        with open(os.path.join(SUB_DIR, f"{name}.txt"), "w", encoding="utf-8") as f:
            f.write(txt)

    truth = {"clean": ["CLEAN"], "clone": ["PLAGIARIZED"],
             "mosaic": ["PLAGIARIZED", "SUSPICIOUS"],
             "idea": ["SUSPICIOUS", "PLAGIARIZED"]}
    with open(os.path.join(SUB_DIR, "truth.json"), "w", encoding="utf-8") as f:
        json.dump(truth, f, indent=1)
    print(f"corpus: {len(corpus)} docs -> {CORPUS_DIR}")
    print(f"submissions: {list(subs)} -> {SUB_DIR}")


if __name__ == "__main__":
    main()