Korean Phone Scam Corpus

• Korean Phone Scam Corpus is a large-scale dataset featuring de-identified, annotated transcriptions of Korean voice phishing calls for social engineering research.
• It includes detailed crime script labels and intent annotations across multiple scam stages, with balanced scam and benign instances to ensure training robustness.
• The dataset employs rigorous pre-processing, including anonymization, numerical normalization, and morphological segmentation using tools like MeCab-ko and SentencePiece.

The Korean Phone Scam Corpus is a structured, large-scale dataset of annotated Korean voice phishing (전화금융사기) transcriptions, specifically designed to support research in social engineering scam detection and natural language understanding for LLMs. It provides meticulously de-identified, turn-level, intent-annotated transcripts of actual scam calls and matched benign controls, with precise labeling for crime script inference and pragmatic intent (Kim et al., 20 Jan 2026).

1. Data Collection, Sources, and Inclusion Criteria

The corpus is sourced from the LAW ORDER benchmark (Police Science Institute PSI 2025), which comprises transcriptions of voice-phishing calls intercepted and recorded by the Korean National Police Agency between 2015 and 2019. Only phone‐based "prosecutor impersonation" scams are included, reflecting the most prevalent and financially damaging modus operandi in the Korean threat landscape. The dataset comprises 571 complete call-record cases, with each case manually transcribed and de-identified. Personal health information (PHI) such as names, account numbers, addresses, and ages is replaced by standardized placeholders (e.g., [Name], [Bank Name], [Address], [Age]).

To establish robust negative controls, legitimate Korean police summons calls were sampled from public-domain transcripts, with compatible anonymization applied. Each transcript undergoes cleaning: whitespace and punctuation normalization, dates and numerals replaced with <NUM>, and honorific endings unified via a canonical mapping table. Pragmatic details such as honorifics are preserved, and neither stemming nor stop-word removal is used to avoid loss of potentially salient cues.

The dataset is formally defined as

$$D_o = \{ C^{(i)} \}_{i=1}^{571}, \quad C^{(i)} = \left\{ \left(U^{\text{usr}}_t, U^{\text{scm}}_t\right) \right\}_{t=1}^{T_i}$$

where $U^{\text{usr}}_t$ and $U^{\text{scm}}_t$ denote the user and scammer utterances at turn $t$.

2. Dataset Structure, Statistics, and Class Balancing

The dataset construction pipeline results in several key stages, summarized below. The final corpus maintains a strict 1:1 class balance of scam and benign instances for model training stability.

$\begin{tabular}{lrr} \hline Dataset Stage & \# Cases & \# Instances / Utterances \ \hline Original voice-phishing corpus %%%%3%%%% & 571 & 48,229 total utterances \ Intent-Annotated Subset %%%%4%%%% & – & 23,771 scammer–utterance/intents \ Crime Script Inference Dataset (CSID) & 571 & 22,712 context–prediction instances \   – scam segments & – & 11,356 \   – benign segments & – & 11,356 \ \hline \end{tabular}$

The corpus exhibits notable imbalance at the crime script stage level. The “Case Involvement” stage alone accounts for 43.3% of all annotated scammer utterances:

$\begin{tabular}{lrr} \hline Script Stage & \# Intents & \# Utterances (\%) \ \hline 1. Identity Confirmation & 3 & 1,924 (8.1\%) \ 2. Case Introduction & 7 & 4,112 (17.3\%) \ 3. Case Involvement & 15 & 10,298 (43.3\%) \ 4. Prep. for Voice Inv. & 8 & 3,255 (13.7\%) \ 5. Voice Investigation & 11 & 4,182 (17.6\%) \ Total & 44 & 23,771 (100\%) \ \hline \end{tabular}$

A plausible implication is that models trained on this corpus may require explicit strategies for addressing class balance during crime script stage classification.

3. Annotation Protocol and Schema

Annotation is performed by professional crime analysts and follows a two-level schema:

• Script-Stage Label ($s$): Each scammer utterance is tagged with one of five discourse stages, reflecting a granular crime script typology:
  1. Identity Confirmation & Introduction
  2. Case Introduction (fabrication of a crime)
  3. Case Involvement (vetting/verifying the victim’s ostensible role)
  4. Preparation for Voice Investigation (technical instructions, e.g., setting up call recording)
  5. Voice Investigation (solicitation of sensitive or financial data)

• Intent Label ($y$): Fine-grained intent, drawn from a finite set $\mathcal{Y}$ of 44 categories. Example intents in Stage 3 (“Case Involvement”) include:

  • Explaining link of bank account to crime
  • Warning of legal penalty for account sale

Formally, for the Intent-Annotated Subset: $D_I = \{ (U^{\text{scm}}_t, (s_t, y_t)) \}$ Inter-annotator agreement is quantified by Cohen’s Kappa with a reported value of 0.91, indicating high reliability of the labeling procedure.

4. Release Format and File Specification

Two core files are released for downstream use:

• raw_intent_data.json (utterance-level annotations):

$U^{\text{usr}}_t$7

• csid_data.json (context-prediction instances for model training):

$U^{\text{usr}}_t$8

Each record includes:

• case_id: Unique conversation identifier
• context: List of ordered preceding utterances
• is_scam: Binary class label (1 = scam, 0 = benign)
• next_scam_utterance: Gold next utterance (scammer)
• intent_explanation: Natural-language rationale of intent

5. Linguistic Processing and Normalization

The text processing pipeline includes several domain- and language-specific operations:

1. Anonymization: Human-identifying information is mapped to placeholders ([Name], [Bank Name], [Address], [Age]).
2. Numerical Normalization: All digit sequences are standardized to <NUM>:

$U^{\text{usr}}_t$0

1. Tokenization: Performed using MeCab-ko for morphological segmentation, followed by SentencePiece (v0.1.96) to obtain word pieces with a 32,000-type vocabulary.
2. Honorific Unification: Variant forms of Korean honorifics are collapsed into a canonical representation for consistency.

Whitespace, punctuation, and honorific endings are standardized, but no stemming or stop-word removal is performed in order to preserve function words and pragmatic markers that may be predictive of scam intent.

6. Data Augmentation and Contextual Instance Generation

To create a training dataset suitable for conversational modeling, a sliding-window segmentation strategy is used. Let $U^{\text{usr}}_t$1 be a conversation of length $U^{\text{usr}}_t$2 turns, with sequences $U^{\text{usr}}_t$3. For each prefix $U^{\text{usr}}_t$4, $U^{\text{usr}}_t$5, an instance is generated for predicting the next scammer utterance $U^{\text{usr}}_t$6 and its intent.

Each scam instance is mirrored with a matched benign police summons script, resulting in balanced counts: 11,356 scam and 11,356 benign segments, yielding a total of 22,712 CSID context–prediction instances. The original corpus thus serves as the seed for systematic context–target pairs at all intermediate conversational positions.

This suggests the dataset is highly suitable for sequential reasoning and next-utterance prediction tasks, as well as joint scam classification and intent explanation, under settings that reflect real-world conversation dynamics.

7. Research Applications and Public Availability

By combining de-identification, dual-level crime script annotation, morphological normalization, and exhaustive context segmentation, the Korean Phone Scam Corpus provides a foundation for cognitively adaptive, human-centered LLMs in social engineering defense. It is published in a format ready for fine-tuning and evaluation on joint tasks: scam detection, next-utterance prediction, and intent explanation in the Korean language. The dataset supports empirical evaluation of model reasoning over multi-turn, personalized deception scenarios and is integral to frameworks such as ScriptMind, enabling accurate simulation and assessment of both automated and human-in-the-loop scam defense efficacy (Kim et al., 20 Jan 2026).