ScientificTextClassification_ResearchField

📚 Overview

This is a RoBERTa-base model fine-tuned for the complex task of multi-class classification of scientific article abstracts. The model predicts the primary research field (e.g., Physics, Biology, Computer Science) based solely on the abstract text, serving as a powerful tool for automated journal indexing and literature review organization.

🧠 Model Architecture

The choice of RoBERTa ensures enhanced robustness and better handling of long-range dependencies common in technical and scientific prose.

• Base Model: roberta-base (an optimized BERT approach without the next-sentence prediction objective).
• Classification Head: Outputs 8 distinct categories (num_labels: 8).
• Input Data: Detailed scientific abstracts from diverse journals.
• Output: A probability distribution over the 8 classes: Physics, Chemistry, Medicine, Computer Science, Biology, Geoscience, Materials Science, and Engineering.
• Training Dataset: ScientificArticleAbstract_Classification, providing abstracts linked to their high-level research disciplines.

🎯 Intended Use

The model offers utility in several scientific and information retrieval contexts:

1. Automated Library and Repository Indexing: Rapidly and accurately tagging new publications with their correct discipline.
2. Literature Review Automation: Filtering large databases of articles to focus on specific fields.
3. Grant Proposal Routing: Assisting research institutions in routing incoming proposals to the appropriate review panel or expert based on the summary.
4. Trend Analysis: Tracking the volume and convergence of research across different fields.

⚠️ Limitations

1. Interdisciplinary Papers: The model performs single-label classification. It may struggle with highly interdisciplinary abstracts that bridge two or more distinct fields (e.g., computational chemistry or bio-engineering).
2. Vocabulary Drift: Scientific terminology evolves quickly. New sub-disciplines or extremely novel concepts may not be classified correctly until the model is retrained.
3. Class Imbalance: If the underlying distribution of the eight fields in the real world shifts significantly from the training set, performance may vary.

MODEL 3: EcommerceAspectSentiment_BART

This model is a BART-large sequence-to-sequence model fine-tuned for abstractive multi-aspect sentiment summarization based on Dataset 3 (EcommerceCustomerReview_MultiAspectRating).

config.json

{
  "_name_or_path": "facebook/bart-large",
  "architectures": [
    "BartForConditionalGeneration"
  ],
  "model_type": "bart",
  "vocab_size": 50265,
  "d_model": 1024,
  "encoder_layers": 12,
  "decoder_layers": 12,
  "encoder_attention_heads": 16,
  "decoder_attention_heads": 16,
  "encoder_ffn_dim": 4096,
  "decoder_ffn_dim": 4096,
  "dropout": 0.1,
  "activation_function": "gelu",
  "init_std": 0.02,
  "num_labels": 3,
  "max_position_embeddings": 1024,
  "eos_token_id": 2,
  "bos_token_id": 0,
  "pad_token_id": 1,
  "is_encoder_decoder": true,
  "scale_embedding": false,
  "forced_eos_token_id": 2,
  "transformers_version": "4.35.2"
}