Natural Language Processing
Natural language processing has become the substrate of agentic AI: the professional work has shifted from training task-specific models toward specifying, grounding, orchestrating, and evaluating language models that call tools and act on retrieved knowledge. The course therefore keeps its linguistic and neural foundations while adding retrieval-augmented generation, tool calling and the Model Context Protocol, injection guardrails, and evaluation harnesses with LLM-as-judge scoring and cost and latency budgets.
Current description → proposed description
This course explores the computational techniques that enable machines to interpret, generate, and interact with human language. Students will examine both the foundational linguistics and the advanced algorithms that power modern Natural Language Processing (NLP) systems, from rule-based parsing to large-scale transformer models. Topics include text preprocessing, tokenization, word embeddings, part-of-speech tagging, named entity recognition, syntactic parsing, sentiment analysis, and machine translation. The course covers classic models such as Hidden Markov Models (HMMs) and n-gram language models, as well as deep learning-based approaches including Recurrent Neural Networks (RNNs), sequence-to-sequence models, and transformer architectures. Students will apply techniques in real-world scenarios such as chatbots, summarization, text classification, information retrieval, and question answering. Projects emphasize both implementation and evaluation, with attention to ethical concerns such as bias, misinformation, and the responsible use of generative language models.
This course explores the computational techniques that enable machines to interpret, generate, and interact with human language. Students will examine both the foundational linguistics and the advanced algorithms that power modern Natural Language Processing (NLP) systems, from rule-based parsing to large-scale transformer models and tool-using language agents. Topics include text preprocessing, tokenization, word and contextual embeddings, part-of-speech tagging, named entity recognition, syntactic parsing, sentiment analysis, and machine translation. The course covers classic models such as Hidden Markov Models (HMMs) and n-gram language models, as well as deep learning-based approaches including Recurrent Neural Networks (RNNs), sequence-to-sequence models, and transformer architectures, together with the fine-tuning or parameter-efficient adaptation of pretrained models. Students will apply techniques in real-world scenarios such as chatbots, summarization, text classification, information retrieval, and question answering, and will extend these into agentic applications through prompt engineering, retrieval-augmented generation over vector indexes, structured output, and tool or function calling under protocols such as the Model Context Protocol. Projects emphasize both implementation and evaluation, including agent evaluation harnesses, LLM-as-judge methods, guardrails against prompt injection, and cost and latency budgets, with continued attention to ethical concerns such as bias, misinformation, privacy, and the responsible use of generative language models.
What changes
- Retrieval-augmented generation over vector indexes
- Tool/function calling and the Model Context Protocol
- Agent evaluation harnesses and LLM-as-judge scoring
- Prompt-injection guardrails and cost/latency budgets
- Contextual embeddings and parameter-efficient adaptation
6 proposed outcomes, mapped to 11 program outcomes
Each outcome below is written to be observable and assessable, and each is mapped to the program learning outcomes for which it produces evidence.
Students will be able to differentiate the linguistic, statistical, and neural foundations of natural language processing across the rule-based, n-gram and hidden Markov, neural sequence, transformer, and agentic eras, including how each era's demonstrated capabilities support or undercut semantic, historical, and popular claims that a machine "understands" language.
Working across n-gram and hidden Markov models, formal grammars and parsing, and transformer attention is direct application of foundational AI theory, probabilistic models, and language theory (PLO 3.1), while testing "understanding" claims era by era against semantic, historical, and popular definitions is precisely the definitional evaluation required by PLO 5.2.
Students will be able to construct a reproducible text corpus pipeline that acquires documents from APIs, files, and databases and that performs cleaning, normalization, tokenization and subword segmentation, annotation, deduplication, and provenance and licensing tracking suitable for downstream model training or retrieval.
The pipeline collects, cleans, stores, and analyzes language data from multiple sources through APIs and databases for a real downstream use, which is the data acquisition outcome of PLO 2.1 stated almost term for term.
Students will be able to implement neural sequence models for core language tasks, including recurrent and sequence-to-sequence architectures, transformer encoders and decoders, and the fine-tuning or parameter-efficient adaptation of pretrained models for named entity recognition, classification, summarization, and machine translation.
Building and supervising the training of recurrent, sequence-to-sequence, and transformer models and adapting pretrained weights is development of learning-based AI systems under PLO 4.1, and the attention, embedding, and optimization mathematics those implementations require is the applied AI theory of PLO 3.1.
Students will be able to design a domain-specific language agent, implemented against a hosted model API and embedding service, that grounds its responses in a retrieval-augmented pipeline over an embedded vector index and executes tasks through tool or function calling under an interoperability protocol such as the Model Context Protocol, with guardrails, least-privilege tool scopes, and human-in-the-loop escalation that contain prompt-injection and tool supply-chain risk.
The agent is domain-scoped, executes tasks through governed tool calls held to least-privilege scopes, and integrates real-time knowledge by retrieval, which is exactly the secure and flexible agent integration named in PLO 2.2, while building it against a hosted model API and embedding service over a vector index is the use of LLMs through APIs and embeddings named in PLO 6.1.
Students will be able to engineer prompt specifications — system prompts, structured-output schemas, and tool descriptions — together with retrieval configuration and model selection, iterating each against a versioned evaluation harness that combines task-appropriate metrics, LLM-as-judge scoring calibrated against human adjudication, red-team and prompt-injection cases, and cost and latency budgets.
The prompt specification itself — system prompts, structured-output schemas, and tool descriptions, which are the alignment, formatting, and adaptability levers PLO 4.2 names — is the CLO's direct object and therefore the discrete artifact an exit assessor scores for PLO 4.2, the course's only assessed-at-exit outcome, while the versioned harness's task-appropriate metrics, human-calibrated judge scoring, and iterate-against-diagnostics cycle is the performance assessment and iterative refinement of PLO 6.2.
Students will be able to critique the deployment of a generative language system, addressing bias across dialects and speaker populations, privacy and data stewardship, misinformation and attribution, transparency, economic access, and accountability for machine-produced text, in paired deliverables written for a technical review board and for the affected non-technical stakeholders.
The critique's treatment of bias, fairness, privacy, transparency, economic access, and accountability is the ethical and legal integrity outcome that the workbook states twice, word for word, once under Christian Faith as PLO 1.1 and once under Integrated Disciplinary Knowledge as PLO 3.2, so the CLO is mapped to both; the paired deliverables, one for a technical review board and one for affected non-technical stakeholders, supply the audience-differentiated communication evidence for PLO 5.1.
Program outcomes this course reaches
Filled cells are program learning outcomes with at least one supporting course learning outcome in this course. Sparse coverage is expected — no single course carries all twelve.
The workbook maps this course too
The proposal workbook's assessment map already assigns program learning outcomes to this course. The outcomes above were written to cover it.
| Program outcome | Workbook level | In this draft |
|---|---|---|
| PLO 1.1 Ethical and Legal Integrity in AI | D | covered |
| PLO 2.1 Data Acquisition for Real Problems | I | covered |
| PLO 2.2 Domain-Specific AI Agents | D | covered |
| PLO 3.1 AI Theory and Mathematics | D | covered |
| PLO 3.2 Ethical and Legal Integrity in AI | D | covered |
| PLO 4.1 Learning-Based Systems and Agents | D | covered |
| PLO 4.2 Prompt Engineering for LLMs | D,AE | covered |
| PLO 5.1 Communicating AI to Any Audience | D | covered |
| PLO 5.2 Defining and Evaluating AI | I | covered |
| PLO 6.1 LLMs in Industrial Settings | D | covered |
| PLO 6.2 Model Evaluation and Refinement | D | covered |
I = Introduced · D = Developed · AE = Assessed at Exit