NATURAL LANGUAGE UNDERSTANDING (NLU) is the task of MAPPING NATURAL LANGUAGE INPUT ONTO A MEANING REPRESENTATION the system can act upon. text → [ analysis ] → a structured representation THE DIRECTION IS MANY-TO-ONE: many surface forms map to the same meaning. "Issue a licence to Ram." "Ram should be issued a licence." "Give Ram his licence." "Please process Ram's licence issuance." ALL FOUR MUST PRODUCE THE SAME REPRESENTATION, something like issue(agent: officer, object: licence, recipient: Ram) That collapsing of variation is what "understanding" means operationally, and it is why canonical forms — Schank's conceptual dependency primitives from the frames topic — were invented. WHAT MAKES NLU HARD, beyond the ambiguity of the previous topic: · the input is UNCONSTRAINED. Users produce typos, fragments, slang, code-switching between languages, and sentences no grammar covers. · MEANING DEPENDS ON CONTEXT that may not be in the text · the system must handle what it was NOT designed for without collapsing · IMPLICIT INFORMATION must be filled in — the restaurant script from the frames topic exists for exactly this THE SUBTASKS OF NLU, in the order they are usually applied: TOKENISATION and normalisation POS TAGGING PARSING to a syntactic structure NAMED ENTITY RECOGNITION COREFERENCE RESOLUTION WORD SENSE DISAMBIGUATION SEMANTIC ROLE LABELLING → who did what to whom INTENT CLASSIFICATION and SLOT FILLING (in dialogue systems): the intent is what the user wants, the slots are the parameters "Book a flight to Delhi on Tuesday" intent: BookFlight slots: destination = Delhi, date = Tuesday MAPPING to a logical form or an action THE OUTPUT REPRESENTATIONS, and each connects to ACtE0903: A LOGICAL FORM in first-order logic A FRAME with slots filled — the FrameNet approach A SEMANTIC NETWORK fragment A DATABASE QUERY (semantic parsing: text to SQL) AN API CALL with arguments THE CHOICE DEPENDS ON WHAT THE SYSTEM WILL DO WITH IT, which is the granularity issue from the KR topic — represent no more detail than the task requires.

NATURAL LANGUAGE GENERATION (NLG) is the reverse: PRODUCING NATURAL LANGUAGE TEXT FROM A NON-LINGUISTIC REPRESENTATION. structured data → [ generation ] → text THE DIRECTION IS ONE-TO-MANY: one meaning has many acceptable expressions, and the system must CHOOSE. That choice — what to say, in what order, in what words — is the whole difficulty, and it is a difficulty of DECISION rather than of INTERPRETATION. THE CLASSICAL PIPELINE, three stages (Reiter and Dale): 1. DOCUMENT PLANNING (macro-planning) CONTENT DETERMINATION — what information to include, and what to leave out. Usually the hardest decision, because the input data contains far more than should be said. DOCUMENT STRUCTURING — the order and the rhetorical relations between the parts 2. MICROPLANNING (sentence planning) LEXICALISATION — choosing the words. "Increased", "rose", "went up", "climbed" all express the same relation with different connotations. AGGREGATION — combining facts into single sentences rather than emitting one sentence per fact REFERRING EXPRESSION GENERATION — deciding between "the applicant", "Ram", "he", and "the 21-year-old from Chitwan". Getting this wrong produces text that is either repetitive or ambiguous. 3. SURFACE REALISATION Applying grammar and morphology to produce the final string: word order, agreement, inflection, punctuation. A WORKED GENERATION, from the study tracker's own data: INPUT (structured): student: Ram topics_done: 412 of 519 subjects_complete: 8 exam_date: 2026-11-14 days_remaining: 99 weakest_subject: ACtE09 (4 of 56) CONTENT DETERMINATION — what matters? The exam is in 99 days and one subject is badly behind. The count of completed subjects is encouraging but less actionable. DECIDE: lead with the gap, mention the progress for context, omit the exact date since "99 days" conveys it. STRUCTURING — progress first (establishes credibility), then the problem, then the recommendation. LEXICALISATION — "behind" rather than "failing"; "worth prioritising" rather than "you must". AGGREGATION — combine the two counts into one clause instead of two sentences. REFERRING EXPRESSIONS — "you" throughout; name the subject once and then use "it". REALISATION: "You've completed 412 of 519 topics with 8 subjects finished — good progress with 99 days left. Artificial Intelligence is the outlier at 4 of 56 topics, so it is worth prioritising next." NOTE HOW MANY DECISIONS PRODUCED TWO SENTENCES. Every one of them could have gone differently and produced acceptable output, which is why NLG has no single correct answer and cannot be evaluated by exact match. THREE APPROACHES TO NLG, in increasing flexibility: TEMPLATE-BASED fixed strings with slots. "You have completed {n} of {total} topics." ✔ predictable, controllable, cannot produce nonsense ✔ still the right choice for most business reporting ✘ rigid, and repetitive across many outputs RULE-BASED / PIPELINE the three stages above, implemented with explicit grammars and choice rules ✔ flexible, linguistically principled ✘ laborious to build NEURAL / LEARNED a model trained to map input to text ✔ fluent and varied ✘ may HALLUCINATE facts not in the input — the decisive problem for any report that must be accurate, and the reason template-based generation persists in finance and medicine

NLU AND NLG ARE NOT MIRROR IMAGES, and the differences are examinable. NLU NLG ───────────────────────────────────────────────────────────── many surface forms → one one meaning → many acceptable meaning (many-to-one) forms (one-to-many) the input is given and the input is complete and uncontrolled controlled the problem is AMBIGUITY — the problem is CHOICE — which reading was meant? which expression to use? failure = misunderstanding failure = awkward, verbose or misleading text evaluable against a gold NO single correct output, so annotation evaluation is hard must handle input outside its generates only within its own design capability ───────────────────────────────────────────────────────────── THE EVALUATION ASYMMETRY IS THE PRACTICAL ONE. NLU output can be compared with a human-annotated answer and scored. NLG output cannot, because a different-but-good sentence would score badly. The workarounds: BLEU, ROUGE, METEOR n-gram overlap with reference texts. Cheap, and they penalise legitimate paraphrase — a serious known weakness. HUMAN EVALUATION raters score fluency, adequacy and correctness. The gold standard, and expensive. TASK-BASED does the generated text let a reader do the job? The most meaningful and least used. WHICH IS HARDER? A defensible answer, since exams ask: NLU IS HARDER IN PRINCIPLE, because it must cope with arbitrary input, unbounded ambiguity and unstated world knowledge — the trophy-and-suitcase problem has no generation counterpart. NLG IS HARDER TO EVALUATE AND TO CONTROL, because there is no target to compare against and a fluent system can be confidently wrong. IN THE ERA OF LARGE LANGUAGE MODELS THE PRACTICAL DIFFICULTY HAS SHIFTED TOWARD GENERATION, precisely because fluency became easy while FAITHFULNESS TO THE INPUT DATA did not. A COMPLETE DIALOGUE SYSTEM uses both, with state in between: 1. speech recognition (if spoken) → text 2. NLU: intent + slots → structured 3. DIALOGUE STATE TRACKING — accumulate what is known across turns, and identify what is still missing 4. DIALOGUE POLICY — decide the next action: answer, ask a clarifying question, confirm, or hand off 5. NLG → text 6. speech synthesis (if spoken) → audio THE PARTS THE SYLLABUS'S NLU/NLG SPLIT OMITS ARE 3 AND 4, and they are where most real dialogue systems fail. A system that understands each utterance perfectly and forgets the previous turn is useless — which is the discourse level of the previous topic, appearing as an architectural requirement. AND A NOTE ON WHERE THIS SITS RELATIVE TO THE REST OF THE SECTION: NLU produces exactly the representations of ACtE0903 — logical forms, frames, semantic networks — and NLG consumes them. THE KNOWLEDGE REPRESENTATION SECTION IS THE MIDDLE OF THIS PIPELINE, which is why the two sections sit together in the syllabus.