This page records what the course slides said at the time, including the policies below. Use Brightspace for current announcements, releases, and deadlines.
- Sep 3, 11:59 PM: the original deadline for forming project teams. Lecture 5 noted that a few groups were still arranging their membership.
- Sep 10, end of day: each team was to submit one project proposal on Brightspace. The proposal counted for 5% of the course grade. Lecture 6 reminded students that it was due that day.
- Homework 1, due Oct 13 at 2:59 PM Pacific: Lecture 7 announced the release for Sep 15, with a TA announcement and submission on Brightspace. Details are in the Homework 1 section below. Earlier instructions set homework submission times at 2:59 PM on the due date and other deliverables at 11:59 PM PT.
- Project proposals: Lecture 7 reported that all 46 teams had submitted, with grading expected in about a week.
Where to attend class and get help Lec 1–3
- Instructor: the slides introduce Xiang Ren as an Associate Professor of Computer Science at USC, with a PhD from UIUC and connections to AI2, Snapchat, Stanford, and Sahara AI. His research covers NLP; reasoning and generalization in large language models; safety and trust; how models attribute, remember, and trace information; and ways to make inference more efficient as systems scale. His website is seanre.com.
- Classes and materials: the course schedule lists Tuesday and Thursday, 4:00–5:50 PM, in SAL 101. Lectures and readings are on the course website. Announcements and class recordings are on Brightspace.
- Office hours: Lecture 3 updated the time to Friday, 8:45–9:30 AM, at GCS SB5.
- Questions: use Piazza for course questions. Post publicly when the answer could help other students; reserve private posts for personal circumstances. The course asks students not to email the instructor directly. Enrollment and D-clearance are handled by the CS department through myViterbi, so do not email the instructor about those either. The instructor cannot reserve a place in the class.
- DEN students: attend online through the Brightspace link. Slides are posted after class. Live attendance is not required for DEN students, but the quiz requirements still apply.
- Computing resources: at the time of these slides, CARC access was being arranged. Students were asked to watch the course website for updates.
How the grade is divided Lec 1 · Aug 25
| Work | Share of final grade | Details from the slides |
|---|---|---|
| Two homework assignments, 10% each | 20% | Programming and written questions, with 3–4 weeks per assignment. The possible modules, subject to change, were n-gram language models; word embeddings and RNN language models; attention in Transformers; and advanced LLM topics. |
| Class project | 55% | Proposal: 5%; midterm presentation: 5%; midterm report: 15%; final presentation: 10%; final report: 20%. |
| Paper review | 10% | |
| Participation | 15% | Six in-class quizzes, attendance, and participation on Piazza. |
The stated grading schedule was 2–2.5 weeks after submission. Students had one week after grades were announced to request a regrade.
How late-day allowances work
A late-day token lets a student cover one day of late submission on eligible work. The slides give each student six tokens for the semester, with these rules:
- Tokens can cover homework, the paper review, and eligible project submissions. They cannot cover quizzes, presentations, or the final project report.
- A single assignment can use at most three late days. Days are counted as whole days, with no partial-day allowance.
- For team work, tokens are tracked separately for each person. Each teammate spends one of their own tokens for every late day they want covered.
- Each late day that is not covered by a token carries a 20% penalty. Leftover tokens do not become extra credit.
Homework 1: sentiment analysis of product reviews Lec 7 · PDF pp. 5–7
Lecture 7 presented the first homework. It is an individual assignment worth 100 points, built on Amazon Office Products reviews: 200,000 reviews with an 80/20 train/test split. It is due October 13, 2026, at 2:59 PM Pacific. The three parts draw directly on the first six lectures:
| Part | Points | What to do | Notes on this site |
|---|---|---|---|
| Prepare reviews and extract features | 50 | Clean the text, remove stop words, lemmatize, and build TF-IDF features. | Preparing text, TF-IDF |
| Train and evaluate three classifiers | 30 | Perceptron, logistic regression, and multinomial Naive Bayes. Report train and test accuracy, precision, recall, and F1. | Logistic regression |
| Build a bigram language model | 20 | Handle unknown words and compare add-k smoothing. Evaluate corpus perplexity for \(k=1\), \(0.1\), and \(0.01\). | n-grams, unknown words, add-k, perplexity |
- Submission: one file,
sentiment.py, through Brightspace → Activities → Assignments → HW1. Use Python 3.12. No written report is required. - Perceptron and Naive Bayes: the assignment supplies background and references. Use the specified scikit-learn implementations; no algorithm derivations are required. Neither classifier has been derived in lecture, so these notes do not cover them.
Quizzes, attendance, and absences Lec 1, 2, 4, 7
- The course lists six quizzes with no makeups. They encourage students to do the readings and are released and submitted during class. DEN students take them through Brightspace.
- The attendance policy calls for random checks, with possible grade reductions for non-attendance. Quizzes also serve as attendance checks, and Piazza activity contributes to the participation grade.
- Lecture 4's quiz procedure: enter the access code given in class to take the quiz on Brightspace, and check in using the dynamic QR code shown at the same time. These are two separate steps. Missing the QR check-in makes the final quiz score zero, even if Brightspace shows a passing score.
- Quiz 1 attendance (Lec 7): counting excused and DEN students, 254 students took the quiz and 23 did not.
- DEN quiz procedure (Lec 7): the quiz is published after class. DEN students have a 72-hour window in which to start it, the quiz itself takes 10 minutes, and it must be taken in the Respondus lockdown browser. The slide spells the browser name as “Rsoondus”; Respondus is the product Brightspace uses.
- To request an absence: submit the form under Brightspace → Activities → Assignments at least 24 hours before class. Include evidence, such as a doctor's or advisor's note, conference registration, or explicit instructor approval. Reasons outside medical or academic needs require the instructor's approval in advance.
Planning the class project Lec 1–7
The project addresses a question in NLP and serves learning objective O4. It can take the form of a research project or a working demonstration. The slides allow two directions:
- Research-focused: develop models or analyze data for an existing problem, or define a new problem to investigate.
- Application-focused: train a model, possibly by fine-tuning an existing one, and deploy it in a new application area.
The standard team size is six students, and teams can mix DEN and in-person students. Everyone must contribute. Lecture 5 mentioned a few exceptions and membership changes, so actual team assignments should follow Brightspace and confirmation from the course staff. Lecture 7 reported the final count: 46 full teams, numbered 1–47 with no project 41, and 46 proposals submitted. The staff were also looking for two teams willing to add a seventh member; interested students were told to email hjarai@usc.edu.
TAs act as project mentors, and teams can seek feedback on Piazza. Past examples include reducing harmful language through context distillation; generating haiku; Legal-SBERT; using prompts to produce more varied responses; making LLMs more truthful; generating recipes; identifying an author from limited text; and detecting insults in social commentary. The slides also point to the Stanford CS224N and CS229 project pages for ideas.
What the proposal needs to explain (Sep 10 deadline, 5%)
The stated format is one page, single-spaced, in 10-point font. Start by naming the project type, then clearly answer the following questions. Each answer can be a sentence or a few paragraphs, within the overall page limit.
- What do you want to do? Explain the goal without jargon.
- How do people handle this problem now, and where do those approaches fall short?
- What is new about your approach? Why do you expect it to work?
- Who would care about the result?
- If the project succeeds, what difference will it make?
- What might go wrong? How would you change direction early if that happens?
- How much time and computing will you need? Is that realistic within one semester and the resources you can access?
- What two milestones will mark progress toward the finished project?
How the proposal is scored Lec 5 · PDF pp. 4–5
Each team submits one proposal using the one-page, single-spaced, 10-point format. An application-focused project does not have to introduce a new research question. The rubric has 20 points; the proposal itself is worth 5% of the course grade.
| Area | Points | What the reader needs to understand |
|---|---|---|
| Idea and objectives | 4 | What you aim to do and whether the project focuses on research or an application. |
| Related work | 3 | How others approach the problem and what their methods cannot yet do. |
| Novelty and impact | 5 | What is new, why it might work, who would benefit, and what would change. |
| Planning and assessment | 6 | Risks and an alternative plan: 2 points; time and computing resources: 2 points; two milestones: 2 points. |
| Clarity and presentation | 2 | Whether the proposal is clear, organized, and easy to follow. |
Sources: Lecture 5, PDF pp. 4–5; Lecture 6, PDF p. 2.
Working with others and using generative AI Lec 1 · Aug 25
- The slides allow discussion of homework, but require students to write their own code and reports. You are also responsible for keeping your solution from being shared or copied.
- The homework policy requires independent work rather than answers produced with generative AI. The slides state that AI use and overlap between submissions will be checked, with zero tolerance for plagiarism.
Textbooks listed on the course website
The slides say that all three books are available free through the course website.
- Jurafsky & Martin, Speech and Language Processing, 3rd edition: the main source for fundamentals. The readings listed so far are Chapter 3 for n-grams in Lectures 2–3 and Chapter 5 for logistic regression in Lectures 3–4.
- Eisenstein, Natural Language Processing: explains approaches based on machine learning.
- Goldberg, Neural Network Methods for NLP: develops the subject from a deep-learning perspective.