dhaka-kacchi-review-intelligence

Dhaka Kacchi Review Intelligence — Aspect-Based Sentiment Model

A multi-head DistilBERT model that predicts sentiment for six independent aspects of a restaurant review, rather than collapsing it into a single score.

Full project write-up, data pipeline, and evaluation: GitHub — dhaka-kacchi-review-intelligence. Live demo: https://reviews.dhakakacchi.com

Model description

• Base: distilbert-base-uncased, with six linear classification heads sharing one encoder.
• Aspects: food_taste, service, price, portion_size, authenticity, ambiance.
• Classes per aspect: positive, negative, neutral, not_mentioned.
• Training: weak-label pretraining (~18.5k LLM-labeled Yelp reviews) followed by fine-tuning on 207 human-labeled gold reviews.

Evaluation

Measured on 65 gold reviews held out entirely from training (never seen in weak-label pretraining or gold fine-tuning):

See the error analysis in the project README for known failure modes (over-triggering on not_mentioned, a 256-token truncation limit, and a single-annotator gold set).

Usage

import torch
from transformers import AutoTokenizer, AutoConfig, AutoModel
from huggingface_hub import hf_hub_download

REPO = "shadmanArko/dhaka-kacchi-review-intelligence"
ASPECTS = ["food_taste", "service", "price", "portion_size", "authenticity", "ambiance"]
LABELS = ["positive", "negative", "neutral", "not_mentioned"]

tokenizer = AutoTokenizer.from_pretrained(REPO)
config = AutoConfig.from_pretrained(REPO)

class MultiAspectDistilBert(torch.nn.Module):
    def __init__(self, config, n_labels, aspects):
        super().__init__()
        self.encoder = AutoModel.from_config(config)
        self.heads = torch.nn.ModuleDict({
            a: torch.nn.Linear(config.dim, n_labels) for a in aspects
        })

    def forward(self, input_ids, attention_mask):
        pooled = self.encoder(input_ids, attention_mask).last_hidden_state[:, 0]
        return {a: head(pooled) for a, head in self.heads.items()}

model = MultiAspectDistilBert(config, len(LABELS), ASPECTS)
state_dict = torch.load(hf_hub_download(REPO, "model.pt"), map_location="cpu")
model.load_state_dict(state_dict)
model.eval()

text = "The biryani tastes good but the price is very high. Service was not good."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    logits = model(inputs["input_ids"], inputs["attention_mask"])

for aspect, out in logits.items():
    pred = LABELS[out.argmax(-1).item()]
    print(f"{aspect}: {pred}")

Limitations

• 256-token truncation drops content in ~11% of reviews.
• Trained predominantly on US/Canadian Yelp reviews; underrepresented vocabulary for Bangladeshi-cuisine-specific terms (e.g. "biryani").
• Single-annotator gold set; neutral remains the weakest class across all aspects.

See the project README for the full list of known limitations and next steps.