UMass CS685 S23 (Advanced NLP) #1: Introduction, overview of the state of NLP

New Section

The instructor introduces the course, outlining the focus on recent developments in large language models and logistics for the semester.

Course Introduction

• The semester will delve into explaining how models like GPT work and explore research on enhancing them.

• Students can attend classes in person or watch online via YouTube, with live-streamed lectures and flexible viewing options.

• Three homework assignments and weekly quizzes are planned, emphasizing learning over correctness in assessments.

• Assignments will be submitted via GradeScope, with the first assignment (homework zero) due next Friday to accommodate schedule adjustments.

• The course has three TAs who are NLP PhD students available to mentor students for final projects and academic support.

Logistics and Communication

Details about office hours, communication channels, and resources for students are discussed.

Office Hours and Communication

• Three TAs introduced for academic support; students advised to contact instructors through designated email accounts for queries.

• Emphasis on using Piazza Forum for public or private messages; course website crucial for updates on readings and topics.

• Course materials posted before class; interactive sessions include slide annotations during lectures.

Office Hours Information

Office hour schedules, locations, formats, and accessibility details provided.

Office Hour Logistics

• TAs offer office hours four out of five days a week in CS building; extended slots before exams or challenging assignments if needed.

Course Overview and Deadlines

In this section, the instructor provides important information about deadlines, feedback mechanisms, prerequisites for the course, and details regarding assignments and exams.

Important Deadlines and Feedback Mechanisms

• The deadline to keep in mind is February 21st for the ad drop deadline. Feedback on the course can be provided through an Anonymous Google form outside of Piazza or email.

• An Anonymous Google form is available for leaving critical or positive feedback. Submitted responses will be reviewed at the start of each class.

Course Prerequisites and Homework Assignments

• The course does not require prior knowledge of NLP but familiarity with basic machine learning, probability stats, and programming would be beneficial.

• Homework zero has been released as the first assignment. Students are advised to study independently if they face difficulties with any questions to avoid falling behind in the fast-paced class.

Programming Requirements and Exam Details

• The course involves extensive programming in Python and some homework assignments will utilize PyTorch, a deep learning library.

• Matrix calculus will be covered during discussions on backpropagation; students should be prepared or willing to learn about it along with basic concepts like taking derivatives.

• Weekly quizzes account for five percent of the total grade while there are three planned homework assignments worth ten percent each, including homework zero.

Final Project Details

• A significant portion of the grade will depend on the final project where students work in groups of four on a topic of their choice. Two deliverables include a project proposal due early on and a final report worth 30% of the grade.

New Section

In this section, the instructor discusses the use of AI assistance in assignments and emphasizes the importance of transparency when utilizing such tools.

Using AI Assistance for Assignments

• The instructor mentions that using AI like chat GPT for assignments is allowed, but students must disclose specific prompts used and describe how the AI aided them. This disclosure is required for all assignments.

• Students are responsible for ensuring the correctness of AI-generated work before submission, as these tools may make factual errors, especially with complex or specific questions.

• Despite challenges in detecting AI usage, there is no reliable defense against students using chat GPT. The instructor suggests embracing its use and focusing on understanding its workings and limitations.

Embracing AI Tools

• Analyzing AI disclosures can provide insights into effective prompts and responses, potentially forming a meta assignment where students evaluate each other's approaches to challenging questions using chat GPT.

• GitHub Copilot can be used by providing prompts for code generation tasks; however, it is essential to maintain integrity by not sharing entire incremental code progress unless necessary.

Natural Language Processing and Code Generation

The instructor introduces natural language processing (NLP) concepts, highlighting the application of NLP technologies to programming languages like English and discussing text generation's significance in NLP problem-solving.

Understanding Natural Language Processing

• Natural language refers to human-evolved languages like Spanish or English, emphasizing languages spoken naturally by humans. This semester will also explore code generation applications within NLP, bridging natural language technologies with programming languages.

Understanding Linguistic Hierarchy

In this section, the speaker delves into the linguistic hierarchy, starting from characters to sentences and beyond, exploring how words combine to form meaningful units in language.

Exploring Linguistic Units

• Characters serve as the basic building blocks of text.

• Subwords like suffixes can convey specific meanings such as past tense verbs.

• Words carry semantic meaning contributing to sentence comprehension.

Syntax and Semantic Representation

Syntax governs grammatical rules at the word level, while semantic representation focuses on the meaning of sentences beyond mere structure.

Syntax and Sentence Structure

• Syntax provides grammatical rules for language elements like verbs and prepositions.

• Parsing a sentence into a tree structure reveals how units combine for meaning.

Semantic Representation

• Semantic representations focus on the actual meaning of sentences.

• Formal specifications encode various aspects like agents, recipients, and temporal elements.

Levels of Language Comprehension

The discussion extends to semantic representations covering all sentence meanings and explores discourse analysis beyond single sentences.

Semantic Representations

• Formal representations aim to cover all sentence meanings comprehensively.

• Challenges exist in structuring complex semantic representations effectively.

Discourse Analysis

• Discourse analysis delves into understanding multiple sentences' collective meanings in paragraphs or documents.

Training Powerful NLP Models without Labeled Data

The discussion focuses on training powerful NLP models without access to large amounts of labeled data by utilizing self-supervised learning techniques.

Leveraging Unlabeled Text for Training

• Self-supervised learning in NLP emphasizes language modeling using unlabeled text data.

• Training models on vast amounts of internet text, including reviews, translations, and questions/answers, aids in task comprehension without manual curation.

Language Modeling in Self-Supervised Learning

• Language modeling involves predicting the next word given the beginning of a sentence or document.

• This simple yet powerful task requires semantic understanding and syntactic knowledge for accurate predictions.

Scaling with World Knowledge

• Predicting words necessitates understanding contexts like food items at restaurants, showcasing model complexity.

• Incorporating vast datasets like GitHub enhances the need for world knowledge and complex reasoning in predictions.

Transfer Learning in NLP

Transfer learning is discussed as a method to pre-train models on general tasks like language modeling before fine-tuning them for specific NLP tasks.

Pre-training and Fine-tuning Process

• Pre-training involves training models on general tasks before fine-tuning them for specialized NLP tasks like sentiment analysis.

• Fine-tuning adjusts model parameters based on labeled data to enhance performance on specific tasks like sentiment analysis.

Specialization through Fine-tuning

• Starting from scratch with labeled data limits model learning; transfer learning allows gradual specialization after general pre-training.

Sentiment Analysis and Prompt-Based Learning

In this section, the discussion revolves around sentiment analysis and prompt-based learning using large language models. The focus is on pre-training language models to generate sentiment scores for sentences without fine-tuning.

Pre-Training Language Models for Sentiment Analysis

• Pre-train a large language model to determine the sentiment of a sentence by inputting the sentence and expecting the model to produce a positive or negative score.

Prompt-Based Learning Requirements

• Prompt-based learning necessitates a powerful language model with basic comprehension of instructions provided. The model must reason over prompts to understand tasks without parameter updates.

Specializing Language Models for Tasks

• Questioning how prompt-based learning influences training language models compared to delivering tasks. Emphasizes that instructions given to solve tasks may differ from how the model operates.

Techniques for Specializing Language Models

• Various techniques like instruction tuning and human feedback tuning are used to adapt language models for prompt-based learning, enhancing their ability to follow instructions effectively.

Generalization in Language Modeling

This segment delves into how properties such as following instructions emerge in language models as they scale up in text volume, leading them to specialize outputs for specific tasks through natural language prompts.

Emergence of Task-Solving Abilities

• Language models trained on predicting the next word gradually develop properties like task-solving capabilities when exposed to extensive text data, enabling them to specialize outputs based on prompts.

Impact of Pre-Training Data on Model Performance

• The scale of pre-training data significantly affects a model's ability in prompt-based learning. Larger datasets contribute to emerging properties crucial for task specialization within language models.

New Section

In this section, the discussion revolves around sentiment analysis and how models can be trained to generate specific words based on sentiment.

Sentiment Analysis and Model Generation

• To solve sentiment analysis, the model needs to indicate whether a sentence is positive or negative. One approach is to make the model generate the word "positive" for positive sentiments and "negative" for negative sentiments. This simplifies sentiment analysis into a generation task.

• In supervised learning, models can only classify as positive or negative. However, treating sentiment analysis as a generation problem allows models to solve various tasks beyond sentiment analysis.

New Section

This part discusses the application of large language models in healthcare and methods to specialize these models for specific domains.

Large Language Models in Healthcare

• Large language models can be utilized in healthcare where specialized knowledge is required by adapting methods that involve specializing these models to particular domains.

• Methods include additional pre-training on domain-specific text or modifying the data that the model is trained on. For instance, Chat GPT uses prompt-based learning where inputs are prompts and outputs are generated accordingly.

New Section

In this section, the instructor discusses the importance of project selection and proposal writing for the course. Emphasis is placed on creativity and critical thinking in project design.

Project Selection Criteria

• The instructor highlights that selecting multiple GitHub repositories without deeper engagement will not suffice as a project.

Proposal Writing Guidelines

• Students are encouraged to think creatively about their projects, focusing on new tasks and research questions.

• Key questions to address in proposals include defining the research problem, reviewing prior work, outlining experiments, and explaining evaluation methods.

• Success is not solely based on achieving state-of-the-art results; strong error analysis and effort are valued.

New Section

This segment delves into accessing NLP research resources online, highlighting platforms like ACL Anthology and providing guidance on staying updated with cutting-edge developments.

Accessing NLP Research

• The best NLP research is openly accessible online, with resources like ACL Anthology offering top published papers in the field.

• Recommendations are made to engage with NLP-related papers from machine learning conferences for further insights.

New Section

The discussion shifts towards effective paper search strategies using tools like Semantic Scholar and Google Scholar. Additionally, guidance on expanding reading through citation exploration is provided.

Paper Search Strategies

• Utilize platforms such as Semantic Scholar and Google Scholar for paper searches based on keywords.

• Expand reading by exploring papers cited by or citing a selected paper to deepen understanding of the topic.

New Section

Details regarding project-related tasks being conducted via Overleaf using LaTeX are shared. Instructions for utilizing templates for project proposals and reports are outlined.

Overleaf Platform Usage

• All project-related tasks will be completed using Overleaf with LaTeX; templates will be provided for proposals and final reports.

New Section

Deadlines for key deliverables including homework assignments, group formation deadlines, and project proposal submissions are highlighted. Computational resource recommendations are also discussed.

Deadlines & Computational Resources

• Important deadlines include Homework Zero due on February 17th, group assignment selection by February 17th, and project proposal submission by March 8th.

• Recommendations for computational resources include utilizing Google Colab for small-scale training projects to access GPUs easily.

Detailed Overview of Chat GPT Abilities

In this section, the speaker discusses the capabilities and limitations of Chat GPT, emphasizing its potential applications in various tasks such as creative writing and language learning.

Exploring Chat GPT Capabilities

• Chat GPT can effectively solve diverse tasks based on prompts provided to it, showcasing its ability to generate creative content like short stories.

• The AI's proficiency in completing assignments quickly and creatively is highlighted, enabling users to focus on personal pursuits but potentially hindering genuine learning experiences.

Rewriting Text with Chat GPT

• Demonstrates how Chat GPT can rewrite text in different styles upon request, such as transforming a story into Twitter speak using emojis and slang.

• Despite its adaptability in rewriting content, the AI's output retains quality and coherence even when presented in unconventional formats.

Applications and Limitations of Chat GPT

• Discusses practical uses of Chat GPT for non-native speakers to enhance language skills through grammar correction or professional text refinement.

• Highlights a drawback where the AI lacks real-time information updates, affecting its responses' relevance to current events or specific queries.

Model Updates and Human Feedback

This segment delves into how OpenAI incorporates human feedback to refine their models periodically, enhancing accuracy while addressing potential drawbacks related to model updates.

Incorporating Human Feedback

• OpenAI utilizes human feedback from contractors and users to adjust model outputs based on positive or negative responses received.

• The periodic tuning of models through human feedback ensures continuous improvement without complete retraining, balancing consistency with potential variability in responses.

Real-Time Knowledge Challenges

• Illustrates a scenario where Chat GPT struggles with real-time knowledge queries due to pre-training limitations, leading to irrelevant or random responses despite user expectations.

Exam Question Evaluation with Chat GPT

The speaker demonstrates how complex exam questions can be inputted into Chat GPT for evaluation purposes, showcasing both the AI's capabilities and limitations in handling intricate academic tasks.

Evaluating Exam Questions

• Utilizes an exam question involving text summarization evaluation for Thai language documents as an example of testing Chat GPT's comprehension of specialized topics.

New Section

In this section, the speaker tests a model's understanding of language by inputting a question related to Thai language structure and evaluates the model's response.

Testing Model Understanding

• The model recognizes the complexity of tokenizing Thai words using white space, highlighting that it may not accurately represent the structure of the Thai language.

• When faced with more challenging tasks, such as addressing issues in Thai language structure compared to English, the model struggles to provide a valid solution within given constraints.

• The model fails to adhere to specific constraints provided in the question, suggesting solutions that do not align with the requirements.

• It demonstrates limitations in both comprehending prompts and generating appropriate responses, leading to answers that would not score well in an exam setting.

• While capable of complex reasoning, the model can easily mislead users into incorrect answers due to its limitations in understanding and generating viable responses.