4.1 Introduction to Semantic Analysis

After establishing the grammatical structure of a sentence through syntactic analysis, the next phase in the NLP pipeline is semantic analysis. This phase is concerned with drawing the exact, or dictionary, meaning from the text. While lexical analysis also involves word meanings, semantic analysis operates on a larger scale. It focuses on how the meanings of individual words combine to create meaning in larger chunks of text, such as phrases and sentences. The process can be broadly divided into two primary tasks: first, studying the meaning of individual words (lexical semantics), and second, studying how these words combine to form the meaning of a complete sentence.

4.2 Core Elements of Lexical Semantics

Lexical semantics explores the relationships between words and their meanings. Understanding these relationships is fundamental to building a computational model of language understanding.

4.2.1 Hyponymy

This describes a hierarchical relationship between a generic term (hypernym) and its specific instances (hyponyms). It is often thought of as an “is-a-kind-of” relationship.

• Example: The word “color” is a hypernym, while words like “blue,” “yellow,” and “red” are its hyponyms.

4.2.2 Homonymy

Homonyms are words that have the same spelling or pronunciation but different and unrelated meanings.

• Example: The word “bat” can refer to a flying mammal or an implement used to hit a ball. These two meanings are historically and conceptually distinct.

4.2.3 Polysemy

A polysemous word is a single word with multiple, related meanings or senses.

• Example: The word “bank” can refer to a financial institution, the building where that institution is located, or as a verb meaning “to rely on.” These meanings are all connected to the concept of security or repository.

4.2.4 Distinguishing Polysemy and Homonymy

The crucial difference between polysemy and homonymy lies in whether the multiple meanings associated with a single word form are related or not. In polysemy, the senses are related (e.g., “bank” as a financial institution and a building). In homonymy, the senses are entirely unrelated (e.g., “bat” as an animal and a piece of sporting equipment).

4.2.5 Synonymy

Synonyms are words that have different forms but share the same or a very close meaning.

• Example: “author” and “writer”; “fate” and “destiny.”

4.2.6 Antonymy

Antonyms are words that have opposite meanings. This opposition can exist across several scopes:

• Application of a property or not: Describes a binary opposition (e.g., life / death).
• Application of a scalable property: Describes opposites on a continuum (e.g., rich / poor, hot / cold).
• Application of a usage: Describes relational opposites (e.g., father / son).

4.3 Meaning Representation

A core task of semantic analysis is to convert a natural language sentence into a formal structure that captures its meaning. This is known as meaning representation.

4.3.1 Building Blocks of Semantic Systems

Any system for representing meaning is constructed from several key building blocks:

• Entities: Representations of specific individuals, such as people, locations, or organizations. (e.g., Ram, India).
• Concepts: Representations of general categories of individuals. (e.g., person, city).
• Relations: Representations of the relationships between entities and concepts. (e.g., The relation that holds between Ram and person is an “is-a” relation).
• Predicates: Representations of verb structures, often detailing semantic roles. (e.g., The action of giving involves an agent, a recipient, and an object).

4.3.2 Approaches to Meaning Representation

Various formalisms have been developed to represent meaning, including:

• First-Order Predicate Logic (FOPL)
• Semantic Nets
• Frames
• Conceptual Dependency (CD)
• Rule-based architectures
• Case Grammar
• Conceptual Graphs

4.3.3 The Necessity of Meaning Representation

Creating a formal meaning representation is essential for several reasons:

• Linking to the Non-Linguistic World: It provides a bridge between linguistic elements (words) and the non-linguistic concepts, entities, and events they refer to.
• Resolving Lexical Ambiguity: It allows a system to map various words or phrases with the same meaning to a single, unambiguous, canonical form.
• Enabling Reasoning: A formal representation can be used as input for inferential reasoning systems, allowing a machine to verify what is true or to infer new knowledge from the text.

4.4 Word Sense Disambiguation (WSD)

Word Sense Disambiguation (WSD) is the specific computational task of determining which meaning, or “sense,” of a word is activated by its use in a particular context. This is a direct attack on the problem of lexical ambiguity.

For example, consider the word bass. WSD is the process that would determine whether the sentence “I can hear bass sound” refers to a low-frequency tone or the sentence “He likes to eat grilled bass” refers to a type of fish.

4.4.1 Evaluating WSD Systems

Evaluating the performance of a WSD system requires two key inputs:

1. A Dictionary: A sense inventory (like WordNet) that defines the set of possible senses for each word to be disambiguated.
2. A Test Corpus: A collection of texts where words have been manually annotated with their correct senses. This corpus can be of two types:
   • Lexical sample: A corpus where only a small, pre-selected set of target words are annotated.
   • All-words: A corpus where every word in the text is annotated with its correct sense.

4.4.2 Methodologies for WSD

WSD methods are typically classified based on the knowledge sources they use.

Dictionary-based (Knowledge-based) Methods These methods rely primarily on lexical resources like dictionaries and thesauruses. The seminal approach is the Lesk algorithm, which works by comparing the dictionary definition of an ambiguous word’s different senses with the definitions of the other words in its context. The sense whose definition has the highest overlap with the context words is chosen as the correct one.

Supervised Methods These machine learning methods use a sense-annotated corpus as training data. The context of an ambiguous word is represented as a feature vector, and a classifier is trained to map these features to the correct sense. Supervised methods, such as those using Support Vector Machines, are often the most accurate but require large amounts of expensive, manually sense-tagged data.

Semi-supervised Methods To mitigate the need for large labeled datasets, semi-supervised methods use a small amount of labeled data along with a large amount of unlabeled data. A common technique is bootstrapping, where a classifier is first trained on the small seed set of labeled data and then used to label the unlabeled data. The most confident predictions are added to the labeled set, and the process repeats.

Unsupervised Methods These methods do not rely on any manually annotated data. They operate on the assumption that similar senses occur in similar contexts. They work by clustering the occurrences of a word based on the similarity of their contexts, thereby inducing sense distinctions directly from the raw text. This task is often called word sense induction.

4.4.3 Applications and Challenges of WSD

WSD is a foundational task that enables many other NLP applications, but it remains a significant challenge.

• Applications of WSD:
  • Machine Translation: Selecting the correct translation for a word with multiple possible translations.
  • Information Retrieval: Resolving query ambiguities to improve search relevance.
  • Text Mining & Information Extraction: Accurately identifying concepts and entities.
  • Lexicography: Aiding in the creation of dictionaries by analyzing word usage in corpora.
• Difficulties in WSD:
  • Differences between Dictionaries: Different dictionaries may divide a word’s meanings into different senses, making the “correct” answer subjective.
  • Algorithm Specificity: Different applications may require entirely different WSD algorithms.
  • Inter-Judge Variance: Humans often disagree on the correct sense of a word in a given context, making it difficult to create a “gold standard” for evaluation.
  • Word-Sense Discreteness: The assumption that word meanings can be neatly divided into discrete senses is often a simplification of the fluid nature of language.

This module has shown how semantic analysis and WSD are crucial for moving beyond syntactic structure to true comprehension. The next module will expand this view to see how meaning is handled across multiple sentences in connected discourse.