Natural Language Understanding (NLU)

The process of categorizing a user's utterance into a predefined action or goal. Unlike simple keyword matching, intent classification analyzes the full semantic context to determine what the user wants to do.

• Binary vs. Multi-class: Distinguishes between simple yes/no intents and complex categorical selections
• Out-of-scope detection: Identifies queries that fall outside the model's trained domain to prevent hallucinated responses
• Real-world example: Mapping "My internet is down" to a TROUBLESHOOT_CONNECTIVITY intent rather than a BILLING_INQUIRY

The task of locating and classifying atomic pieces of information in text into predefined categories such as persons, organizations, locations, dates, and custom domain-specific types.

• Span detection: Identifies the exact character offsets of each entity mention
• Nested NER: Handles entities embedded within other entities, such as "[Bank of [America]]"
• Domain adaptation: Fine-tuning on proprietary schemas like medical codes (ICD-10) or legal citations
• Example: Extracting {drug: "Metformin", dosage: "500mg", frequency: "twice daily"} from clinical notes

A complementary task to intent classification that extracts the specific parameters required to execute a user's request. Slots are the variables that populate an API call or database query.

• Schema-guided: Slots are defined by a strict ontology of required and optional fields
• Multi-turn slot collection: The system asks clarifying questions until all mandatory slots are filled
• Example: For a flight booking intent, slots include {origin_city, destination_city, departure_date, return_date, passengers}

The deep linguistic analysis that identifies the predicate-argument structure of a sentence, answering who did what to whom, when, where, and how. SRL surfaces the underlying proposition regardless of syntactic variation.

• Core roles: Agent (doer), Patient (receiver), Instrument, Beneficiary
• Frame semantics: Uses resources like FrameNet to map words to conceptual frames
• Example: In "Acme Corp acquired StartupX for $200M," SRL identifies Acme Corp as the Buyer, StartupX as the Acquired, and $200M as the Price

The process of mapping ambiguous textual mentions to their unique, canonical entries in a knowledge base such as Wikidata, DBpedia, or a proprietary enterprise graph.

• Disambiguation: Resolves "Apple" to the company vs. the fruit based on context
• Nil prediction: Identifies entities that have no corresponding entry in the target knowledge base
• Cross-document coreference: Links mentions of the same real-world entity across multiple documents
• Example: Linking "POTUS" and "Joe Biden" to the same Q6279 Wikidata identifier

The classification of affective states and subjective opinions expressed in text, ranging from basic polarity to fine-grained emotional taxonomies.

• Aspect-based sentiment: Identifies sentiment toward specific features (e.g., "battery life is great but the screen is dim")
• Emotion taxonomies: Ekman's six basic emotions (anger, disgust, fear, joy, sadness, surprise) or Plutchik's wheel
• Intensity scoring: Quantifies sentiment strength on a continuous scale rather than binary positive/negative
• Example: Detecting escalating frustration across a customer support chat to trigger a human handoff

The process of categorizing a user's utterance into a predefined action category. This is the primary decision engine behind NLU, mapping raw text like "What's the weather in London?" to a GetWeather intent.

• Binary vs. Multi-class: Simple yes/no versus complex routing
• Out-of-scope detection: Identifying queries the system cannot handle
• Confidence thresholding: Rejecting low-probability classifications to trigger fallback responses

Also known as Named Entity Recognition (NER), this task identifies and classifies atomic pieces of information within text. It transforms unstructured strings into structured key-value pairs.

• Slot filling: Extracting parameters like city: London and date: tomorrow
• Custom entity types: Domain-specific labels such as PRODUCT_NAME or ACCOUNT_NUMBER
• Span detection: Locating the exact character offsets of each entity in the original text

The deep linguistic task of identifying the predicate-argument structure of a sentence. It answers "Who did what to whom, when, where, and how?" by assigning roles like Agent, Patient, and Instrument to sentence constituents.

• Enables extraction of complex relational facts beyond flat entities
• Critical for building knowledge graph triples from unstructured text
• Uses FrameNet and PropBank as standard role ontologies

The NLP task of identifying all linguistic expressions that refer to the same real-world entity. It links pronouns, definite descriptions, and named mentions into coherent chains.

• Anaphora resolution: Linking "she" or "the device" back to a prior mention
• Cataphora resolution: Resolving forward references before the entity is named
• Without this, NLU systems fragment a single entity into multiple disconnected references, corrupting downstream reasoning

The classification of emotional polarity and affective state from text. Modern NLU systems use aspect-based sentiment analysis to extract fine-grained opinions about specific product features rather than just overall document sentiment.

• Polarity detection: Positive, negative, neutral
• Emotion classification: Joy, anger, frustration, satisfaction
• Intensity scoring: Measuring the strength of expressed sentiment
• Critical for customer experience monitoring and churn prediction