Model Card for qxswy/roberta-card-ner

Model Details

Model Description

This model performs Named Entity Recognition specialized for analyzing product reviews of board games, particularly Cards Against Humanity. It identifies five custom entity types corresponding to different components of perceived value: PRODUCT (functional value), ATTRIBUTE (emotional value), PEOPLE (social value), EVENT (contextual value), and OTHER_GAME (comparative value).

• Model type: Named Entity Recognition (Token Classification)
• Language(s) (NLP): English
• Finetuned from model: roberta-large

Uses

This model is designed for extracting and categorizing value components from product reviews to:

• Analyze perceived value propositions
• Identify key product attributes that resonate with consumers
• Segment customers based on value perception profiles
• Develop targeted marketing strategies

Bias, Risks, and Limitations

• Entity recognition performs better for some categories (OTHER_GAME, PRODUCT) than others (ATTRIBUTE, PEOPLE)
• The model was trained on a relatively small dataset (500 reviews), which may limit its generalizability
• Domain-specific training means performance will be best for board/card game reviews

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import pipeline

ner_pipe = pipeline(
    "token-classification",
    model="qxswy/roberta-card-ner",
    tokenizer="qxswy/roberta-card-ner",
    aggregation_strategy="simple"
)

text = "This fun game is perfect for parties with friends!"
entities = ner_pipe(text)
print(entities)

Training Details

Training Data

The model was trained on 500 manually annotated Amazon reviews with domain-specific entity labels. The data was labeled using Label Studio, with stratified sampling to ensure representation across review ratings and lengths. Inter-annotator agreement (Cohen's kappa) for the annotation process was 0.78.

Training Procedure

• Preprocessing: BIO (Begin-Inside-Outside) tagging scheme

• Training Hyperparameters:

  • Learning rate: 3e-5
  • Batch size: 4
  • Epochs: 8
  • Max sequence length: 256 tokens
  • Special techniques: R-Drop for consistency, oversampling for rare categories