Sociolinguistic research on Korean spoken and written language variation across two projects.
Key word: Style shift
Summary: Investigates stylistic variation and vowel lengthening in written chat/messaging data. Examines how speakers mark vowel lengthening (e.g., tilde ~) in text and relates this to speaker identity and social network context. Uses Press member network chat data with manually collected messaging; members are ranked by network counts, active membership, and network type counts.
Data information and processing steps:
- Data: Press member network chat data; manually collected written messaging; 12 members (HJ, MC, J, SM, ME, JR, SH, UM, OY, EM, SY, DY); network metrics per member.
- Processing: Reshape raw array to DataFrame with columns (total network counts, active member, network type counts, sum, rank); index by member ID; identify lengthening markers (e.g., tilde ~) in chat messages; aggregate metrics per member for ranking.
Core analysis and technique:
- Network ranking: extraction and aggregation of network metrics (total network counts, network type counts, active member, sum, rank) per member.
- Style variation: identification of lengthening markers in chat messages and comparison across members.
- Technique: descriptive statistics, network metrics aggregation, tabular comparison.
Core libraries: pandas, numpy, seaborn, matplotlib, collections, re, os
Key word: Intonational phrase
Summary: Two analyses on the NIKL (National Institute of Korean Language) Dialogue corpus 2023. (1) IP length: Relationship between speaker sex and intonational phrase length (token count, duration per phrase) in spoken Korean dialogue; modeled with SVR. (2) Vowel lengthening in transcriptions: How vowel lengthening and phonetic variants (e.g., tilde ~, spelling variants) are marked in NIKL transcriptions; compares original_form vs. form to detect deviation.
Data information and processing steps:
- Data: NIKL Dialogue 2023 corpus; JSON files (~1974 files); each file:
document,metadata,utterance; speaker metadata (id, sex, age, occupation, birthplace, residence, education); utterance fields:form,original_form,speaker_id,start,end; 2–4 speakers per dialogue. - Processing: Load JSON; extract speakers and map sex (여성/남성 to M/F); extract utterances; compute token_count, syll_count, duration; derive audience_sex and sex_combi for dyadic vs. group contexts; one-hot encode sex, audience_sex, sex_combi; for deviation analysis: extract_speech per speaker; compare original_form vs. form to build deviation pairs.
Core analysis and technique:
- IP length modeling: SVR with RBF kernel; GridSearchCV for hyperparameters (C, epsilon); negative MSE scoring; StandardScaler for feature scaling.
- Deviation detection: extract_speech extracts per-speaker utterances; pairs where original_form != form capture lengthening markers and spelling variants.
- EDA: mean, variance of token count and IP length; distribution analysis per speaker and per dialogue.
- Technique: Support Vector Regression, grid search, deviation comparison, descriptive statistics.
Core libraries: pandas, numpy, seaborn, matplotlib, json, os, collections, scikit-learn (SVR, GridSearchCV, StandardScaler, mean_squared_error, train_test_split)