(SEM VII) THEORY EXAMINATION 2023-24 NATURAL LANGUAGE PROCESSING

KCS072 – NATURAL LANGUAGE PROCESSING

B.Tech (SEM VII) – Exam-Oriented Answers

SECTION A

(2 Marks × 10 = 20)

a) Evolution of NLP

NLP has evolved from rule-based systems in the 1950s to statistical methods in the 1990s and modern deep learning models today. Early systems relied on grammar rules, while current NLP uses machine learning and transformer-based models like BERT and GPT.

b) Challenges in language modeling

Challenges include data sparsity, handling unseen words, ambiguity, long-range dependencies, and selecting appropriate context size for accurate predictions.

c) Handling ambiguity in parsing

Ambiguity is handled using probabilistic parsing, semantic constraints, context-based rules, and statistical models that choose the most likely interpretation.

d) Role of Dynamic Programming in parsing

Dynamic Programming avoids repeated computations by storing intermediate results. Algorithms like CYK and Earley parsing use DP to efficiently parse sentences.

e) Limitations of supervised approaches in WSD

Supervised WSD requires large labeled datasets, is costly to build, domain-dependent, and performs poorly on unseen words or new domains.

f) Role of semantic attachments in WSD

Semantic attachments use contextual meaning and relationships to associate correct word senses, reducing ambiguity by linking words to semantic roles.

g) Applications of filter bank methods

Filter banks are used in speech recognition, speaker identification, audio compression, and feature extraction like MFCCs.

h) Contribution of filter banks in speech analysis

They decompose speech signals into frequency bands, helping analyze spectral properties relevant to human hearing.

i) Perceptual Linear Prediction (PLP)

PLP models human auditory perception by emphasizing perceptually important frequencies, improving speech recognition accuracy.

j) Role of feature extraction in speech understanding

Feature extraction converts raw speech into meaningful representations such as MFCCs, capturing patterns essential for recognition and classification.

SECTION B

(10 Marks × Any 3 = 30)

a) HMM and Maximum Entropy models in word-level analysis

Hidden Markov Models (HMMs) are probabilistic models representing sequences with hidden states. They are widely used in POS tagging and speech recognition.
Strengths: Efficient, interpretable
Weaknesses: Independence assumptions

Maximum Entropy Models use feature-based probability estimation.
Strengths: Flexible, fewer assumptions
Weaknesses: Computationally expensive

Both models improve sequence prediction and classification tasks.

b) Probabilistic CYK parsing and Lexicalized PCFGs

Probabilistic CYK parsing assigns probabilities to parse trees using PCFGs, selecting the most likely structure.
Lexicalized PCFGs include word-level dependencies, improving parsing accuracy. These methods handle ambiguity better than traditional CFG parsing.

c) First-order logic vs propositional logic

Propositional logic deals with simple true/false statements, while first-order logic includes variables, predicates, and quantifiers.
FOL is more expressive and suitable for representing complex relationships in NLP.

d) Challenges in speech sound representation

Speech sounds vary due to accents, noise, speaking rate, and coarticulation. These variations make classification difficult and impact speech recognition accuracy.

e) Likelihood distortions and feature extraction

Likelihood distortions measure mismatch between speech models and observed data. Feature extraction captures relevant speech patterns, improving model reliability and evaluation.

SECTION C

(10 Marks × Any 1 = 10)

3(a) Minimum Edit Distance (MED)

Minimum Edit Distance measures similarity between two strings using operations like insertion, deletion, and substitution.
 

Example:
“kitten” → “sitting” requires 3 edits.
MED is used in spell checking, OCR correction, and word similarity analysis.

3(b) N-grams and smoothing

N-grams are sequences of N words used in language modeling.
Unsmoothened N-grams assign zero probability to unseen sequences.
Smoothing techniques like Laplace and Good-Turing handle data sparsity and improve evaluation.

SECTION D

(10 Marks × Any 1 = 10)

Dependency Grammar vs Phrase Structure Grammar

Dependency Grammar focuses on word-to-word relationships, while Phrase Structure Grammar represents sentences as hierarchical phrases. Dependency grammar is compact; PSG is more structured.

Shallow Parsing

Shallow parsing identifies chunks like noun phrases without full syntax trees.
 

Advantages: Fast, efficient
Limitations: Limited syntactic depth
Deep parsing provides complete grammatical structure but is computationally expensive.

SECTION E

Thesaurus-based vs Distributional word similarity

Thesaurus-based methods use lexical resources like WordNet, offering precise meanings but limited coverage.
 

Distributional methods use word co-occurrence, scalable but context-sensitive.

Dictionary vs Thesaurus-based WSD

Dictionaries give definitions, while thesauri provide semantic relationships. Choice depends on domain, coverage, and application needs.

SECTION F

Acoustic phonetics

Speech acoustics involve frequency, amplitude, and duration. These properties affect how sounds are perceived and distinguished by listeners.

Linear Predictive Coding (LPC)

LPC models speech by predicting samples using past values.
 

Advantages: Efficient, low bit-rate
Used in speech synthesis and compression.

SECTION G

Dynamic Time Warping (DTW)

DTW aligns speech sequences with different speaking rates. It finds the optimal alignment path and is widely used in speech recognition.

Evaluating HMMs and Viterbi Search

The Viterbi algorithm finds the most probable state sequence (optimal state sequence) in HMMs, crucial for decoding speech and text sequences.