Semantic Analysis

Overview

Semantic analysis moves beyond syntactic structure to understand meaning. It encompasses word-level disambiguation, sentence-level inference, structured meaning representations, and opinion mining. These tasks require models to capture nuance, context, and world knowledge.

Word Sense Disambiguation (WSD)

Determining which sense of a polysemous word is intended in context.

Example: "bank" can mean a financial institution, a river bank, or the act of banking an airplane.

Approaches

Resources

• WordNet: Lexical database organizing words into synonym sets (synsets) linked by semantic relations (hypernymy, meronymy, etc.)
• SemCor: Largest manually sense-annotated corpus (~230k annotations)
• BabelNet: Multilingual encyclopedic dictionary combining WordNet and Wikipedia

Current state: BERT-based WSD models achieve ~80% F1 on the ALL evaluation framework, approaching the estimated inter-annotator agreement ceiling.

Semantic Similarity

Measuring how similar two text units are in meaning.

Word-Level Similarity

• Path-based (WordNet): Shortest path between synsets in the taxonomy
• Information content: Probability of the least common subsumer (Lin similarity, Jiang-Conrath)
• Embedding-based: Cosine similarity of word vectors

Sentence-Level Similarity (STS)

Predict a continuous similarity score (0-5) between sentence pairs.

Cross-encoders are more accurate but O(n^2) for pairwise comparisons. Bi-encoders (SBERT) encode once and compare cheaply.

Semantic Parsing

Translates natural language into a formal meaning representation.

Abstract Meaning Representation (AMR)

A rooted, directed, acyclic graph representing sentence meaning.

"The boy wants to go"

(w / want-01
   :ARG0 (b / boy)
   :ARG1 (g / go-02
      :ARG0 b))

• Nodes: concepts (word senses, PropBank framesets)
• Edges: semantic roles (ARG0, ARG1, etc.)
• Abstracts away syntax: "the boy's desire to go" has the same AMR
• Challenges: reentrancy (nodes with multiple parents), alignment to text
• Neural AMR parsers: sequence-to-graph models achieve ~85 Smatch F1

Text-to-SQL

Converts natural language questions into SQL queries over a database.

"How many employees earn more than 50000?"
->
SELECT COUNT(*) FROM employees WHERE salary > 50000

Key challenges:

• Schema linking: mapping mentions to tables and columns
• Complex queries: joins, subqueries, aggregation, GROUP BY
• Generalization to unseen databases

Approaches:

• Grammar-based decoding: constrain output to valid SQL
• Schema encoding: represent table/column names alongside the question
• Spider benchmark: cross-database evaluation (200 databases)
• Current SOTA: LLM-based approaches with schema prompting achieve ~85% execution accuracy on Spider

Natural Language Inference (NLI)

Determine the logical relationship between a premise and hypothesis.

Datasets

Models

• Decomposable attention: Align words between premise and hypothesis, aggregate
• ESIM: BiLSTM with cross-sentence attention
• BERT cross-encoder: Concatenate [CLS] premise [SEP] hypothesis [SEP], classify
• BERT achieves ~92% on SNLI, ~90% on MultiNLI
• NLI models are used as general-purpose zero-shot classifiers (hypothesis = category description)

Sentiment Analysis

Determining the opinion, emotion, or attitude expressed in text.

Levels of Granularity

Lexicon-Based Approaches

• VADER: Rule-based, handles social media conventions (capitalization, emoticons, intensifiers)
• SentiWordNet: Assigns positivity/negativity scores to WordNet synsets
• AFINN, Opinion Lexicon: Curated word lists with polarity scores
• Fast but limited by vocabulary coverage and inability to handle negation/sarcasm robustly

Machine Learning Approaches

Classical ML:

• Features: unigrams/bigrams, POS tags, negation scope, lexicon scores
• Classifiers: SVM, logistic regression, Naive Bayes
• Effective for well-defined domains with labeled data

Deep Learning:

• CNN over word embeddings captures local n-gram patterns
• BiLSTM captures sequential context and long-range negation
• Attention mechanisms highlight relevant words

Pretrained Transformers:

• Fine-tune BERT on sentiment datasets
• State-of-the-art across benchmarks (SST-2: ~97% accuracy)

Aspect-Based Sentiment Analysis (ABSA)

Subtasks:

1. Aspect term extraction: identify aspect mentions ("food", "service")
2. Aspect category detection: map to predefined categories
3. Aspect sentiment classification: determine polarity per aspect

Approaches:

• Joint extraction with sequence labeling (BIO tags for aspects)
• Attention over aspect term to determine sentiment
• Instruction-tuned LLMs with structured output achieve strong performance

Relation Extraction

Identifying semantic relationships between entities in text.

"Steve Jobs founded Apple in 1976."
-> (Steve Jobs, founded, Apple)
-> (Apple, founded_in, 1976)

Approaches

Open Information Extraction (OpenIE)

Extracts relation triples without predefined schema.

• Input: "Einstein was born in Ulm and developed the theory of relativity"
• Output: (Einstein, was born in, Ulm), (Einstein, developed, theory of relativity)
• Systems: OpenIE 5, Stanford OpenIE, neural OpenIE
• Useful for knowledge base construction and question answering

Knowledge Graph Construction

Relation extraction feeds into knowledge graph construction:

1. Entity recognition and linking
2. Relation extraction between entity pairs
3. Triple validation and canonicalization
4. Integration into knowledge graphs (Wikidata, Freebase)

Evaluation Across Tasks

Key Takeaways

• WSD resolves lexical ambiguity using context; BERT-based models approach human performance
• Semantic parsing maps language to formal representations (AMR, SQL) enabling precise reasoning
• NLI is a versatile task used for zero-shot classification and textual reasoning
• Sentiment analysis ranges from document-level polarity to fine-grained aspect-based opinion mining
• Relation extraction bridges text understanding and knowledge graph construction
• Pretrained transformers dominate all these tasks, but task-specific design choices still matter