GROUPED BY WHAT THEY DO, with an honest note on maturity. MACHINE TRANSLATION Converting text between languages. · RULE-BASED (1950s–80s): hand-written grammars and bilingual dictionaries. Brittle. · STATISTICAL (1990s–2010s): learn phrase correspondences from PARALLEL CORPORA — the same text in two languages. · NEURAL (2014–): sequence-to-sequence models, now transformer-based. A large jump in fluency. MATURITY: excellent for high-resource language pairs and formal text; POOR for low-resource languages, which matters directly for Nepali — parallel corpora are scarce, so quality lags English–French by a wide margin. This is a DATA problem, not an algorithmic one. INFORMATION RETRIEVAL AND SEARCH Matching a query to documents. Uses tokenisation, stemming, stop-word removal, TF-IDF and now embeddings. MATURITY: mature. The shift to embeddings enabled SEMANTIC search — matching meaning rather than words, so "how do I fix a puncture" finds a document about repairing tyres. INFORMATION EXTRACTION Pulling structured facts out of unstructured text: named entities, relations, events. Populating a knowledge graph from news articles is the canonical use — which is the knowledge-acquisition bottleneck from earlier in this section, attacked automatically. MATURITY: good for well-defined entity types, weaker for relations and events. TEXT CLASSIFICATION AND SENTIMENT ANALYSIS Assigning categories: spam or not, positive or negative, which department a complaint belongs to. MATURITY: mature for clear-cut cases; SENTIMENT remains unreliable on sarcasm, mixed opinions and domain-specific idiom. "The battery life is incredible" is positive; "incredible — it died in an hour" is not, and the surface words are similar. QUESTION ANSWERING · CLOSED-DOMAIN over a fixed corpus or database · OPEN-DOMAIN over general knowledge · EXTRACTIVE (find the span in a document) versus ABSTRACTIVE (generate an answer) MATURITY: extractive over a supplied document is reliable; abstractive answers may be fluent and wrong, which is the faithfulness problem from the NLG topic. SUMMARISATION EXTRACTIVE selects existing sentences; ABSTRACTIVE writes new ones. MATURITY: extractive is safe and clunky; abstractive reads better and can introduce claims not in the source. DIALOGUE SYSTEMS AND ASSISTANTS Task-oriented (booking, support) or open-domain (chat). Needs the full pipeline plus dialogue state tracking. MATURITY: task-oriented within a narrow domain works; multi-turn reasoning and remembering context across a long conversation remain weak. SPEECH RECOGNITION AND SYNTHESIS Speech-to-text and text-to-speech. MATURITY: recognition is very good for clear speech in well-resourced languages, and degrades sharply with accent, background noise and code-switching — a real limitation in multilingual settings. SPELLING, GRAMMAR AND STYLE CHECKING MATURITY: mature and genuinely useful. OPTICAL CHARACTER RECOGNITION Images of text to text, with a language model to correct visual errors — "rn" versus "m" is decided by which word exists. APPLICATIONS PARTICULARLY RELEVANT LOCALLY: · Nepali and Devanagari OCR for digitising records · machine translation for government and legal documents · agricultural advisory systems taking spoken questions in Nepali — where speech recognition for a low-resource language is the bottleneck rather than the reasoning

THE CHALLENGES, and the examinable list. Note that AMBIGUITY from the terminology topic underlies most of them. 1. AMBIGUITY AT EVERY LEVEL lexical, syntactic, semantic, referential, pragmatic — with the readings MULTIPLYING, as the Catalan-number growth of PP attachment showed. This is the root challenge. 2. WORLD KNOWLEDGE AND COMMON SENSE The trophy-and-suitcase problem. Resolving reference and implication often requires facts about how the world works that are nowhere in the text and were never written down anywhere. 3. CONTEXT DEPENDENCE The same sentence means different things in different situations. "It's warm in here" is an observation, a complaint, or a request to open a window. 4. FIGURATIVE AND NON-LITERAL LANGUAGE Idiom ("kick the bucket"), metaphor, irony, sarcasm, humour, understatement. A literal reading is simply wrong, and detecting that a reading should not be literal is itself unsolved. 5. VARIATION AND INFORMALITY Dialects, registers, slang, typos, abbreviations, emoji, and CODE-SWITCHING — mixing languages within one sentence, which is completely normal in Nepal and breaks systems trained on monolingual text. 6. LOW-RESOURCE LANGUAGES Most of the world's roughly 7,000 languages have little digital text, so data-hungry methods do not apply. THE INEQUITY IS STRUCTURAL: performance follows corpus size, so speakers of well-documented languages get better tools. 7. MORPHOLOGICAL RICHNESS Languages with heavy inflection or agglutination produce enormous vocabularies, so any fixed word list misses much of the language. Nepali's case marking and verb agreement make this concrete: one root generates many surface forms. Subword tokenisation is the standard mitigation. 8. IMPLICIT INFORMATION AND ELLIPSIS "Ram passed. Hari too." — the second sentence omits the verb and its object, and the reader supplies them. 9. EVALUATION For generation there is no single correct output, so automatic metrics penalise good paraphrase — the asymmetry from the NLU/NLG topic. 10. BIAS AND FAIRNESS Models learn the associations present in their training text, including stereotypes, and then apply them uniformly and at scale. The AI applications topic's warning applies with force here: the appearance of objectivity is itself the harm. 11. HALLUCINATION AND FAITHFULNESS Fluent generation of false statements. Fluency is precisely what makes the errors hard to spot, and it is the dominant practical problem in deployed generation today. 12. LONG-RANGE DEPENDENCE AND DISCOURSE Tracking entities, topics and commitments across a long document or conversation. THE UNDERLYING REASON LANGUAGE IS HARD, stated once: LANGUAGE IS A COMPRESSED ENCODING THAT ASSUMES A SHARED MODEL OF THE WORLD. Speakers omit everything the listener can reconstruct, so understanding requires the reconstruction — and the shared model is exactly what a computer lacks. That is why NLP is an AI problem and not a string-processing problem.

THE APPROACHES, and their honest limits: MORE DATA AND LARGER MODELS Pre-train on enormous text corpora, then fine-tune or prompt for a task. ✔ dramatically improved fluency, translation, classification and extraction ✔ frequency and plausibility — the tractable half of human disambiguation — come for free ✘ does not supply world knowledge reliably, only statistical association that often resembles it ✘ does nothing for low-resource languages, which have no corpus to scale ✘ hallucination is not fixed by scale TRANSFER LEARNING AND MULTILINGUAL MODELS Train on many languages together so that a low-resource language benefits from a high-resource one's structure. ✔ the most promising route for languages like Nepali, and genuinely effective ✘ performance still tracks data volume GROUNDING Connect language to something other than text — images, databases, a simulated or physical environment. The argument is that meaning cannot be learned from form alone, so a system trained only on text has no referent for its words. This is an active and unsettled research direction. RETRIEVAL AND STRUCTURED KNOWLEDGE Fetch relevant facts before generating, and cite them, so the output can be checked. Addresses faithfulness by constraint rather than by hoping — the retrieval-augmented approach from the NLG topic. HYBRID SYMBOLIC AND STATISTICAL Statistical methods for perception and disambiguation, symbolic representations where explanation and guarantees are required. Exactly the neuro-symbolic direction the AI concepts topic flagged as open. HUMAN IN THE LOOP For high-stakes uses — legal, medical, financial — the system drafts and a human approves. This is the error-cost argument from the AI applications checklist, applied to language. WHAT REMAINS GENUINELY UNSOLVED, stated plainly because a graduate who can name these is more useful than one who cannot: · robust COMMON-SENSE REASONING · reliable FAITHFULNESS in generation · genuine understanding of FIGURATIVE language · equitable performance across LOW-RESOURCE languages · long-range DISCOURSE coherence and memory · EVALUATION of generated text · knowing WHAT THE SYSTEM DOES NOT KNOW — a fluent model has no reliable signal for its own ignorance, which is the confident-error problem from the applications topic THE PROFESSIONAL POSITION: NLP is now good enough that the binding constraint is usually not accuracy but VERIFIABILITY. A system that is right 95% of the time and cannot tell you which 5% is unusable for anything consequential, and useful for a great deal that is not. KNOWING WHICH SIDE OF THAT LINE AN APPLICATION FALLS ON IS THE JUDGEMENT THAT MATTERS.