Syllabus Map

• Study map: Syllabus Study Map

Overview

• NLP tasks turn unstructured text into predictions, labels, or generated text.
• Different tasks use different output formats: class labels, spans, sequences, or rankings.
• A strong baseline pipeline is usually: preprocessing, representation, model, evaluation, and error analysis.

Classification Tasks

Sentiment Analysis

• Predict sentiment label such as positive, negative, or neutral.
• Common outputs: single-label class probabilities.

Topic Classification

• Assign text to predefined categories such as sports, finance, or politics.
• Often solved with TF-IDF + linear models or fine-tuned transformers.

Intent Classification

• Predict user intent in assistants and chatbots (for example: book flight, reset password).
• Usually paired with slot filling in task-oriented dialogue systems.

Sequence Labeling Tasks

Named Entity Recognition (NER)

• Tag tokens with entity types (for example PER, ORG, LOC).
• Common labeling format: BIO/BIOES tag schemes.

Part-of-Speech Tagging

• Predict syntactic category for each token (noun, verb, adjective, etc.).
• Useful as an intermediate linguistic feature in some pipelines.

Slot Filling

• Extract structured fields from user utterances (date, city, account ID).
• Frequently used with intent classification in assistants.

Span and Pair Tasks

Question Answering (Extractive)

• Predict answer span start/end positions inside a context passage.
• Typical datasets provide passage-question-answer triples.

Natural Language Inference (NLI)

• Classify relation between premise and hypothesis: entailment, contradiction, or neutral.
• Common benchmark task for sentence understanding.

Paraphrase and Semantic Similarity

• Predict whether two texts are semantically equivalent or how similar they are.
• Used in duplicate detection, search, and retrieval reranking.

Generation Tasks

Summarization

• Produce concise summary of a document.
• Can be extractive (select spans) or abstractive (generate new phrasing).

Machine Translation

• Translate text from source language to target language.
• Usually trained with parallel corpora and sequence-to-sequence models.

Text Generation

• Generate continuations, responses, or structured outputs from prompts.
• Quality depends heavily on prompting, decoding strategy, and safety constraints.

Typical Task Workflow

Step 1: Define output format

• Choose label, span, sequence, or ranking outputs based on product need.
• Write precise annotation and evaluation guidelines.

Step 2: Build a baseline

• Start with simple models and preprocessing to validate data quality.
• Establish a reproducible baseline before complex architectures.

Step 3: Improve model and data

• Use pretrained encoders or language models for better transfer.
• Focus on data cleaning, class balance, and edge-case coverage.

Step 4: Evaluate and analyze errors

• Evaluate on held-out data and representative slices.
• Inspect systematic failures and iterate on data/model choices.

Practical Notes

Match metrics to task goal

• Classification tasks often emphasize precision/recall tradeoffs.
• Sequence/span tasks need token-level or span-level evaluation.
• Retrieval and ranking tasks usually need ranking-aware metrics.

Data quality drives results

• Inconsistent labels can dominate model error.
• Domain-specific terminology often requires targeted data collection.
• Balanced, representative test sets are essential for reliable conclusions.

Deployment constraints matter early

• Latency and memory limits affect model choice and serving strategy.
• Monitoring for drift and regressions should be planned before launch.
• Human-in-the-loop review is useful for high-risk failure modes.