Apppublic

equinox013/hikari-complaint-classifier

sourceHugging Facemitupdated 5mo agoView on Hugging Face

0likes

App README

HIKARI · HIMARI · MACHI

Complaint Driver Classification Pipeline

An end-to-end NLP pipeline that reads raw customer verbatims, strips PII, and multi-label classifies each complaint into one or more of 10 complaint driver categories — built for utility sector CX analytics.

How to use

1.Paste a raw customer complaint (email, web form, or phone log transcript) into the text area.
2.Select the Case Origin so HIMARI applies the correct channel-extraction rules.
3.Adjust the confidence threshold if needed (default 0.5).
4.Click Classify.

The app runs the full three-stage pipeline:

—MACHI redacts PII before any text is analysed
—HIMARI normalises and tokenises the cleaned text
—HIKARI predicts complaint driver labels and confidence scores

Model Details

Component	Architecture	Training Data
MACHI (PII redaction)	CRF tagger (sklearn-crfsuite) with hand-engineered lexical features	ai4privacy/pii-masking-200k
HIKARI (classifier)	Bidirectional LSTM + GloVe 6B 100d (frozen) + GlobalMaxPool1D	Internal utility sector verbatims labelled via Qualtrics TextIQ

Complaint Driver Labels

Customer Service · Digital · Online Experience · Outages and Faults · Process · Reputation · Sustainability · Trust · Value for Money · Vulnerability & FDV

Notes

—PII patterns are AU-specific (mobile formats, postcodes, street types).
—GloVe weights are frozen inside hikari_v2.keras — no external embedding download required.
—This demo is for portfolio/research purposes. Do not submit real customer PII.