Enriching Prompts with Definitions and Guidelines for Improved Zero-Shot Named Entity Recognition
Providing definition and guidelines in the prompt can improve the performance and robustness of instruction-tuned language models for zero-shot named entity recognition, especially on unseen entity types.
Interpreting Cosine Similarity via Normalized Independent Component Analysis-transformed Embeddings
Cosine similarity can be interpreted as the sum of semantic similarities along the interpretable axes of normalized ICA-transformed embeddings.
Equipping Language Models with Calibrated Linguistic Expressions of Uncertainty
Language models can be fine-tuned to generate well-calibrated linguistic expressions of uncertainty that accurately reflect the likelihood of their predictions being correct.
BERT-VBD: A Novel Vietnamese Multi-Document Summarization Framework Combining Extractive and Abstractive Techniques
The proposed framework leverages a two-component pipeline architecture that integrates extractive and abstractive summarization techniques to generate high-quality Vietnamese multi-document summaries.
Multilingual Skill Matching for Efficient Freelancer-Project Alignment at Scale
A novel neural retriever architecture that leverages pre-trained multilingual language models to effectively match freelancer profiles and project descriptions, enabling efficient and scalable candidate retrieval in a multilingual setting.
A Comprehensive Benchmark for Evaluating Paraphrase Detection Models
PARAPHRASUS is a comprehensive benchmark designed to assess the performance of paraphrase detection models across a diverse range of paraphrase phenomena.
Enhancing Large Language Models with Simultaneous Thinking and Speaking Capabilities
A novel training approach called TaS that enables large language models to first generate reasonable thoughts and then express corresponding responses, mimicking the human cognitive process.
Unifying Extractive and Abstractive Text Summarization within a Single Encoder-Decoder Framework
A novel extract-and-abstract paradigm, EXTABS, that jointly and seamlessly performs extractive and abstractive summarization tasks within a single encoder-decoder model, reducing error accumulation and improving performance.
Enabling Real-Time Conversational Interactions with Minimal Training Costs
A novel channel-based parallel decoding approach, DUO, that equips large language models with duplex capabilities, enabling simultaneous input processing and output generation, while requiring minimal additional training.
Evaluating Large Language Models' Ability to Detect Character Knowledge Errors in Role-Playing
Even the latest large language models struggle to effectively detect known knowledge errors and unknown knowledge errors when playing roles, particularly for familiar knowledge.
© 2024 by Linnk AI