Constructing a Regional Slang Lexicon for Electricity Demands via Soft-Label Bidirectional Anchoring and Active Learning
Organizations: Yuxi Power Supply Bureau, Yunnan Power Grid Co., Ltd., Yuxi, 653100, China. · Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming, 650504, China. · Key Laboratory of Artificial Intelligence in Yunnan Province, Kunming University of Science and Technology, Kunming, 650504, China.
Abstract
Power-grid customer service increasingly relies on community social media, where customers describe outages, billing, and equipment faults in informal regional dialect. Such messages exhibit non-standard morphology, semantic drift, and out-of-vocabulary expressions that make manual lexicon construction costly and degrade conventional NLP tools. We propose a weak-supervision framework constructing a regional slang lexicon from curated dialect anchors and power-domain seed terms. Candidates are generated via Pointwise Mutual Information with soft-label bidirectional anchoring, retaining dialect-anchored terms while recording power-domain relevance as separate metadata, and are scored by combining Sentence-BERT semantic similarity with a jieba-based wordhood measure. Uncertainty-aware active learning with rule-based and retrieval-augmented LLM oracles is applied under three query strategies, and the resulting lexicon is evaluated by a higher-capability LLM calibrated against human expert annotations via Cohen's . On 11,177 real WeChat messages from Yuxi (Yunnan, China), the framework achieves an evaluator-estimated dialect precision of DP~~1.0 with zero false positives. An ablation shows that soft-label layering retains 751% more candidates than strict power-gating at identical precision, expanding the candidate pool from 41 to 349 terms. Human calibration () yields and . The resulting 25--33-term lexicon, 4--5 of which are power-relevant, provides a preliminary resource for dialect-aware power-grid service NLP.
Figures & tables
| Chinese term (pinyin) | Literal meaning | Power-grid meaning | Translated example |
|---|---|---|---|
| 扎实 (zhashi) | Sturdy, solid, firm | Severe, very, intense | “雷打得扎实, 变压器都跳了” (The thunder was intense; the transformer tripped). |
| 鬼火 (guihuo) | Phosphorescence, will-o’-the-wisp | Anger, frustration, annoyance | “三天两头停电, 哪个不鬼火?” (Outages every few days; who would not be frustrated?). |
| 给有 (gei you) | To give, to have | Is there? Does it exist? (question marker) | “供电局呢领导, 给有电了嘛?” (Power bureau leaders, is there electricity yet?). |
| 着炸 (zhao zha) | Affected + to explode | Damaged by electrical surge, shorted | “表箱着炸掉了, 快点来看哈” (The meter box was damaged by a surge; please come and check it). |
| 老火 (laohuo) | Old fire | Severe, terrible, critical (intensifier) | “线路老火, 快派人来修” (The line fault is severe; please send someone to repair it). |
| Parameter | Value | Description |
|---|---|---|
| 0.78 | Auto-accept threshold | |
| 0.45 | Rejection threshold | |
| 0.60 | Decision boundary | |
| / | 0.6 / 0.4 | SBERT / wordhood weights |
| 5 | Candidates per AL round | |
| 3 | Maximum AL iterations |
| Model | DP | PDP | Lexicon | Queries | ||
|---|---|---|---|---|---|---|
| Margin + Rule (Soft-Label) | 1.0 | 0.160 | 25 | 4 | 25 | 15 |
| TD-CAL + Rule | 1.0 | 0.148 | 27 | 4 | 27 | 15 |
| Margin + LLM | 1.0 | 0.125 | 32 | 4 | 32 | 15 |
| Proposed (TD-CAL + LLM) | 1.0 | 0.125 | 32 | 4 | 32 | 15 |
| Random + LLM | 1.0 | 0.121 | 33 | 4 | 33 | 15 |
| Margin + Rule (Hard Gate) | 1.0 | 0.625 | 8 | 5 | 8 | 5 |
| Pipeline Stage | Hard Gate | Soft Label | |
|---|---|---|---|
| Candidates extracted | 41 | 349 | +751% |
| After fragment filter | 9 | 58 | +544% |
| Uncertain pool | 5 | 39 | +680% |
| Final lexicon | 8 | 25 | +213% |
| DP | 1.0 | 1.0 | — |
| Dimension | 95% CI | Verdict | |
|---|---|---|---|
| has_dialect_features | 1.0 | — | Degenerate (all True) |
| power_domain_relevance | 0.85 | [0.64, 1.0] | Strong agreement |
| Configuration | DP | PDP | Lexicon | Candidates | Unigram Noise |
|---|---|---|---|---|---|
| NPMI, no bidirectional anchoring | 0.17–0.28 | 0.01–0.05 | 101–148 | 347–496 | 72–75% |
| PMI, strict power gate | 1.0 | 0.63 | 8 | 41 | 0% |
| PMI, soft-label layering | 1.0 | 0.12–0.16 | 25–33 | 349 | 0% |
Weakly Supervised Learning
Weakly supervised learning trains models from incomplete, noisy, coarse, or indirect labels instead of dense ground truth.
See all 163 papers