Shubhra Kanti Karmaker’s Post

✨ We are pleased to announce that one of our papers, "ALIGN-SIM: A Task-Free Test Bed for Evaluating and Interpreting Sentence Embeddings through Semantic Similarity Alignment," has been published at EMNLP 2024! 💥   Our paper introduces ALIGN-SIM, a novel task-free framework for evaluating and interpreting sentence embeddings based on various semantic similarity alignment criteria. A big thanks to Dr. Eduardo Blanco for his collaboration and significant contributions to this work.🤝   Read the full paper here: https://lnkd.in/gsqUiyJh #EMNLP #LLM #NLP

Excited to share our latest paper co-authored with Alessandra Zarcone at the UncertaiNLP workshop EACL 2024: "Aligning Uncertainty: Leveraging LLMs to Analyze Uncertainty Transfer in Text Summarization." Check out the full paper here: https://lnkd.in/egtG_2XK #NLP #Research #Uncertainty #TextSummarization #llms #llm

Excited to share my latest blog on Medium: "DSPy to Simplify Prompt Optimization". Dive into the DSPy framework by Stanford NLP, designed to change the way we approach programmatic prompt optimization. Learn how you can precisely tailor responses to your needs. DSPy has a bit of a learning curve, but well worth it if you actively develop. https://lnkd.in/g-84XZcf #NLP #PromptOptimization #DSPy #LLM

Exciting breakthrough in NLP: Graph Language Models (GLMs) 🚀 Just discovered GLMs - a powerful fusion of traditional Language Models and Knowledge Graphs. This innovation offers: • Rich text understanding from pre-trained LMs • Enhanced relationship representation via graph structures • Improved performance in complex tasks (e.g., question answering, information retrieval) • Potential for more context-aware and accurate AI systems Key benefits: 1. Faster development cycles by leveraging existing LM capabilities 2. More robust and interpretable models 3. Simplified integration of domain-specific knowledge 4. Scalable architecture for handling large-scale data GLMs could revolutionize how we build intelligent systems, offering a more holistic approach to language understanding and knowledge representation. Have you experimented with Knowledge Graphs or Language Models? What potential applications do you see for GLMs in your work? #NLP #AI #MachineLearning #KnowledgeGraphs #TechInnovation

The inference method is crucial for NLP models in subword tokenization. Methods such as BPE, WordPiece, and UnigramLM offer different mappings, but their performance differences need to be better.... #Greedy #inference #lead #models #NLP #SAGE #Strategies #tokenization #Unlocking

🚀 Imagine creating a “social network” for your favorite fiction worlds! 🚀 With a fine-tuned language model, it’s very possible! 👉 Dive into the details of my paper, published at EMNLP 2024’s NLP4DH workshop: https://lnkd.in/eVrqXk5g 👉 Or check out the model here: https://lnkd.in/eeDciEf9 Curious about the design choices behind the project? Let’s connect! 🌐 #EMNLP #MLOps #NLP #LLM

New Blog Post! 🚀 Curious about how tokenization powers NLP models like GPT and BERT? In my latest post, I explore key tokenization techniques from Whitespace and BPE to Hugging Face Tokenizers and SentencePiece. Learn how these methods help models handle complex language and rare words. 🔍 What’s inside: - Tokenization basics - Subword tokenization with BPE and SentencePiece - Code examples and practical tips Check it out here: https://lnkd.in/gsvPXmxW GitHub repo for the tokenizers: https://lnkd.in/gyMYxuPC NotebookLM podcast version: https://lnkd.in/gyhhh8kE #NLP #AI #Tokenization #DeepLearning #MachineLearning

I'm excited to share that my research paper, "SLTK: A Comprehensive Tokenizer for Sinhala Language," has been published as a preprint on ResearchGate. The objectives of this research are to identify the key challenges and limitations of existing tokenization methods for Sinhala language, propose a language-specific tokenizer for Sinhala, and evaluate the proposed tokenizer comparing with existing tokenization methods. The proposed tokenizer combines word-level and character-level tokenization with rule-based and dictionary-based text preprocessing to improve the performance of NLP tasks in Sinhala. Let me know if you have any questions or want to discuss the research further! #NLP #Sinhala #Tokenizer #Research #LLM https://lnkd.in/gPGRmBpf

Word Embeddings vs. Positional Encodings: How Transformers Understand Language Have you ever wondered how transformer models like BERT or GPT process text without reading it sequentially, like humans do? The answer lies in the combination of word embeddings and positional encodings, two critical components that enable these models to excel at NLP tasks. Here’s a quick breakdown of their roles: 1. Word Embeddings: Capturing Meaning Word embeddings are dense vector representations of words or tokens that encode semantic meaning. They ensure that words with similar meanings—like “cat” and “dog”—are placed closer together in a high-dimensional vector space. • Example: The word “bank” will have the same embedding, whether it’s about a river or a financial institution. Context comes later in the transformer layers. These embeddings give the model the what, i.e., the meaning of the words in the input sequence. 2. Positional Encodings: Adding Order Transformers process all tokens in parallel, so they lack a natural sense of word order. This is where positional encodings come in—they inject position information into the model, helping it understand the sequence of words. • Example: In “The cat chased the mouse” vs. “The mouse chased the cat,” positional encodings ensure the model knows which animal is doing the chasing. These encodings give the model the where, i.e., the position of each word in the sequence. How They Work Together Word embeddings and positional encodings are combined before being passed to the transformer layers. This fused representation allows the model to simultaneously understand the meaning of words and their order in the sequence. Why It Matters Without embeddings, the model can’t understand word meanings. Without positional encodings, it can’t distinguish between “The cat chased the mouse” and “The mouse chased the cat.” Together, they empower transformers to excel in NLP tasks like translation, summarization, and question-answering. Word embeddings bring the semantic understanding, while positional encodings add the structural context. It’s this synergy that has revolutionized language models. #NLP #Embeddings #transformes #DL #ML

Interested in optimizing the output of a large language model, so it references an authoritative knowledge base outside of its training data sources before generating a response? This article by EQengineered's Technical Consultant, Seth Carney, is worth reading! #techtrends #techleaders #LLMs #CIOs #CDOs #CDAOs #CTOs #CMOs #CFOs #CEOs