The standard NLP pipeline has FIVE STAGES, and this enumeration is the guaranteed exam question. Each takes the previous stage's output and adds structure. 1. LEXICAL ANALYSIS (also MORPHOLOGICAL ANALYSIS) Break the input into tokens and analyse each word's structure. ACTIVITIES: · TOKENISATION — split into words and sentences · normalisation — case folding, expanding contractions · STEMMING or LEMMATISATION · morphological decomposition into morphemes · look up each token in the LEXICON to find its possible parts of speech WORKED on "The officers issued licences." tokens: The | officers | issued | licences | . morphology: officer + -s (plural) issue + -ed (past) licence + -s (plural) lexicon: "The"→Det; "officers"→N; "issued"→V (or Adj); "licences"→N (or V) OUTPUT: a token sequence with candidate analyses. NOTE THAT AMBIGUITY IS ALREADY PRESENT — "issued" could be an adjective ("the issued licence"), so the next stage must choose. 2. SYNTACTIC ANALYSIS (PARSING) Check that the token sequence is grammatical and build a structure showing the relationships. Uses a GRAMMAR, typically a context-free grammar: S → NP VP NP → Det N VP → V NP WORKED PARSE of "The officer issued the licence": S ┌────┴────┐ NP VP ┌──┴──┐ ┌───┴───┐ Det N V NP │ │ │ ┌──┴──┐ The officer issued Det N │ │ the licence OUTPUT: a parse tree (or a dependency graph). WHAT THIS STAGE REJECTS: "Officer the licence issued the" has no valid parse, so it is caught here rather than producing nonsense downstream. WHAT IT CANNOT DO: choose between valid parses. "I saw the man with the telescope" has two, and syntax alone has no basis for preferring either. 3. SEMANTIC ANALYSIS Assign meaning to the syntactic structure, and reject structures that are grammatical but meaningless. ACTIVITIES: · WORD SENSE DISAMBIGUATION · assigning SEMANTIC ROLES — agent, patient, instrument · building a LOGICAL FORM or filling a frame · SELECTIONAL RESTRICTION checking — does the verb accept these kinds of argument? THE CLASSIC EXAMPLE OF WHAT THIS STAGE CATCHES: "Colourless green ideas sleep furiously." (Chomsky) Perfectly grammatical — the parser accepts it. Semantic analysis rejects it: ideas do not sleep, and nothing is both colourless and green. WORKED OUTPUT for our sentence: issue(agent: officer, object: licence, tense: past, definite: both) OUTPUT: a meaning representation, independent of the original wording. 4. DISCOURSE INTEGRATION Interpret the sentence in the context of the sentences around it. ACTIVITIES: · ANAPHORA / PRONOUN RESOLUTION · COREFERENCE — which mentions denote the same entity · connecting to entities introduced earlier · maintaining the topic and the focus WORKED: "The officer reviewed Ram's application. She issued the licence." Discourse integration determines that "She" = the officer, not Ram — using gender agreement here, and in harder cases requiring plausibility. WITHOUT THIS STAGE each sentence is understood in isolation, which is the dialogue-system failure noted in the previous topic. 5. PRAGMATIC ANALYSIS Determine what was actually MEANT, as opposed to what was said. ACTIVITIES: · identifying the SPEECH ACT — is this a statement, a question, a request, a promise? · recognising INDIRECT requests · applying world knowledge and conversational conventions WORKED: "Can you tell me the status of my application?" Syntactically a yes/no question about ability. Pragmatic analysis identifies it as a REQUEST FOR INFORMATION, so the correct response is the status, not "yes". THIS STAGE IS WHERE THE REPRESENTATION BECOMES AN ACTION. THE STAGES ARE OFTEN DRAWN AS A STRICT PIPELINE, and that is a simplification worth flagging: real systems need FEEDBACK. A semantic failure may indicate the wrong parse was chosen, so the parser must be able to offer alternatives — which is why probabilistic parsers rank parses rather than returning one.

PARSING DESERVES EXPANSION because it is the most algorithmically developed stage. TWO STRATEGIES, mirroring the search directions of ACtE0902: TOP-DOWN PARSING — start from S and expand rules until the input is matched. ✔ goal-directed: only builds structures the grammar permits ✘ may explore expansions inconsistent with the actual words ✘ LEFT RECURSION (NP → NP PP) causes infinite descent BOTTOM-UP PARSING — start from the words and combine them into larger constituents. ✔ never builds anything unsupported by the input ✘ builds constituents that cannot form part of any complete parse THE COMBINATION used in practice: CHART PARSING, which stores every constituent found in a table so it is never recomputed — dynamic programming applied to parsing, exactly as variable elimination is to probability. THE CYK ALGORITHM is the standard bottom-up chart parser, running in O(n³·|G|) for a sentence of n words: n = 10 words → 1,000 chart cells n = 20 words → 8,000 n = 40 words → 64,000 Cubic is entirely practical for sentence-length inputs, which is why chart parsing made parsing feasible. HANDLING AMBIGUITY — the practical answer: PROBABILISTIC CONTEXT-FREE GRAMMARS (PCFG) attach a probability to each rule, so a parse's probability is the product of its rules' probabilities and the parser can return the MOST LIKELY parse. THE PROBABILITIES ARE LEARNED FROM A TREEBANK — a corpus of hand-parsed sentences. This is where the statistical turn in NLP came from: instead of writing preference rules for PP attachment by hand, count how often each attachment occurs in real text. LEXICALISED PCFGs condition on the actual words, because "eat pizza with a fork" and "eat pizza with anchovies" attach differently and the difference is entirely lexical. DEPENDENCY PARSING — the alternative output format: Instead of nested phrases, link each word directly to its HEAD with a labelled relation: issued ──nsubj──▶ officer issued ──dobj───▶ licence officer ──det───▶ The licence ──det───▶ the ✔ compact, and directly gives the who-did-what-to-whom that semantic analysis needs ✔ better suited to languages with free word order ✔ faster: transition-based dependency parsers run in linear time → this is why most modern pipelines produce dependencies rather than constituency trees.

A FULL WORKED EXAMPLE, tracing one input through all five stages. INPUT: "Has Ram's licence been issued yet? He applied last month." STAGE 1 — LEXICAL tokens: Has | Ram | 's | licence | been | issued | yet | ? | He | applied | last | month | . morphology: "Ram's" → Ram + possessive "issued" → issue + past participle "applied" → apply + past two sentences identified, by the ? and the . STAGE 2 — SYNTACTIC sentence 1 parses as a yes/no question: Aux(Has) NP(Ram's licence) VP(been issued yet) sentence 2 parses as a declarative: NP(He) VP(applied last month) STAGE 3 — SEMANTIC sentence 1: issued(object: licence_of(Ram), state: query, time: up_to_now) sentence 2: apply(agent: X, time: last_month) NOTE THE X — the agent of "applied" is not resolved yet, because "He" is a pronoun and pronouns are the next stage's business. STAGE 4 — DISCOURSE "He" must be resolved. Candidates: Ram. Gender agrees, and Ram is the only person mentioned. X = Ram Also: "last month" is resolved against the current date to an actual month, and "licence_of(Ram)" is linked to a specific application record. STAGE 5 — PRAGMATIC Sentence 1 is syntactically a yes/no question but is pragmatically a REQUEST FOR THE STATUS. Answering "no" alone would be technically correct and unhelpful. Sentence 2 is not a bare statement of fact — it functions as JUSTIFICATION for the enquiry, implying "enough time has passed that I expect an answer". THE APPROPRIATE SYSTEM RESPONSE therefore includes the status AND an expected date, because the pragmatic analysis identified an implicit concern about delay. NOTE WHAT THE LAST STAGE ADDED: nothing about the words, and everything about what to DO. That is why pragmatics is the stage that connects understanding to action. WHERE THE PIPELINE FAILS, in order of frequency: · TOKENISATION on real-world text — URLs, code, mixed scripts, emoji, no spaces after punctuation · PARSING long sentences with several attachment ambiguities · COREFERENCE, which remains genuinely hard and is where world knowledge is unavoidable · PRAGMATICS, which is barely attempted in most deployed systems — hence assistants that answer the literal question AND THE MODERN COMPLICATION worth knowing: LARGE PRE-TRAINED MODELS DO NOT IMPLEMENT THESE STAGES EXPLICITLY. A transformer maps text to text with no identifiable parser inside, and probing studies find that syntactic information is present in its internal representations without having been designed in. THE STAGES REMAIN THE RIGHT ANALYSIS OF THE PROBLEM even where they are not the architecture of the solution — which is why they are still taught, and still how you diagnose what a system got wrong.