Module 7 Book Prose#
Evaluation for NLP systems#
Why are output quality and factuality hard to measure?
🧑‍🌾 SAMWISE — Student note
Pause before you run the notebook. In your own words:
Whose decision does the essential question above affect?
What baseline and result do you predict before seeing the output?
Which observation would change or strengthen your current view?
What will remain uncertain, and what would you check next?
SAMWISE is a reflection guide, not an answer key or grader. Record your own reasoning; the Populi instructions and published rubric remain authoritative.
Professional Scenario#
You are advising a product team evaluating an NLP workflow before using it in customer-facing communication. The immediate task is to decide what evidence would make a recommendation credible, what risks remain unresolved, and what should happen next. The module’s work product is: NLP evaluation packet with task framing, retrieval/evaluation design, and deployment guardrails focused on evaluation for nlp systems: Create an evaluation set with rubrics and automated checks..
The available lab data is deliberately limited: synthetic support messages, retrieval snippets, intent labels, and factuality checks. Treat it as a proxy for reasoning and method practice, not as proof that a real deployment is ready. A graduate-level submission must distinguish between what the proxy exercise demonstrates and what would still require institutional data, stakeholder review, and operational testing.
Core Concepts#
Problem framing: define the decision, population, workflow, or system boundary before choosing a method.
Baseline discipline: compare the proposed AI-enabled approach with an existing process, simple rule, or manual review pattern.
Evidence quality: separate measured results from assumptions, anecdotes, vendor claims, and synthetic-data artifacts.
Failure modes: identify where the system can fail technically, operationally, legally, ethically, or socially.
Deployment readiness: connect metrics to decision thresholds, monitoring, escalation, and rollback.
Why This Module Matters#
In AINS6004: Natural Language Processing, this module contributes to the larger course arc by requiring students to turn a domain problem into an inspectable technical artifact. The standard is not “the notebook ran.” The standard is that another reviewer can understand the decision, reproduce the reasoning, and challenge the assumptions.
Method Pattern#
State the stakeholder decision in one sentence.
Identify the evidence source and why it is adequate or inadequate.
Produce a baseline result using the lab or an equivalent transparent method.
Compare one alternative design, threshold, policy, or model.
Document false positives, false negatives, unintended incentives, and operational constraints.
Recommend a next action: continue research, run a controlled pilot, redesign the system, or stop.
Failure Modes To Check#
Measurement mismatch: the metric optimizes something adjacent to, but not identical with, the real decision.
Context loss: important operational or human factors are absent from the data.
Automation bias: users may over-trust a score, classification, or recommendation.
Equity and access risk: affected groups may experience different error rates or burdens.
Governance gap: no one owns monitoring, escalation, or rollback after launch.
Study Questions#
What decision does the module artifact support?
What does the proxy lab evidence prove, and what does it not prove?
Which baseline or manual process should the AI-enabled approach be compared against?
Which stakeholder would object to the recommendation, and on what grounds?
What monitoring signal would tell you the system is failing after deployment?
Worked Example: From Evidence to a Decision#
Return to the professional situation for this module: You are advising a product team evaluating an NLP workflow before using it in customer-facing communication. The immediate task is to decide what evidence would make a recommendation credible, what risks remain unresolved, and what should happen next. The module’s work product is: NLP evaluation packet with task framing, retrieval/evaluation design, and deployment guardrails focused on evaluation for nlp systems: Create an evaluation set with rubrics and automated checks.. The team should not begin by selecting the most sophisticated tool. First, rewrite the situation as a decision: what must be decided, by whom, using which evidence, and under which constraints? That sentence establishes the boundary of the analysis.
Next, create an inspectable baseline. For this module, a useful baseline should make Problem framing: define the decision, population, workflow, or system boundary before choosing a method. visible rather than hiding it inside an unsupported conclusion. Preserve the starting data or case facts, record the initial result, and identify the assumption most likely to change the recommendation. Then make one controlled comparison using Baseline discipline: compare the proposed AI-enabled approach with an existing process, simple rule, or manual review pattern.. Holding the other conditions fixed is what lets a reviewer interpret the difference.
Finally, connect the evidence to action. Use Evidence quality: separate measured results from assumptions, anecdotes, vendor claims, and synthetic-data artifacts. to explain why the observed result matters in the scenario, then state a limitation. The appropriate conclusion is conditional: recommend a next step only if the evidence clears a named threshold or review gate. This pattern—decision, baseline, controlled comparison, limitation, next gate—is the same structure expected in the assignment and rubric.
Comprehension Check#
Before continuing, be able to answer: What is the baseline? What single factor changes? Which evidence would reverse the recommendation? What does the exercise leave unknown?
Subject-Matter Lesson#
NLP evaluation must match the task and acceptable variation. Exact match is strict and useful for canonical fields but penalizes equivalent wording. Token F1 gives partial overlap but can reward verbose or semantically wrong responses that share words. Semantic similarity may tolerate paraphrase but inherit evaluator-model biases. Human review remains necessary for nuanced quality and harm.
For retrieval or generation, correctness is not enough: claims should be supported by cited evidence, and abstention should be evaluated when evidence is absent. Aggregate scores can hide language, dialect, topic, or length slices with poor performance. Confidence intervals and inter-rater agreement make uncertainty visible.
The lab computes exact match, token F1, a simple citation-support indicator, and slice results for short and long prompts. Students should replace one correct paraphrase with an unsupported answer that shares many tokens and compare metrics. A release decision should predefine acceptance thresholds, critical slices, severity-weighted error categories, evaluator instructions, and an adjudication process rather than selecting whichever metric looks most favorable.