(SEM VII) THEORY EXAMINATION 2023-24 TEXT ANALYTICS AND NATURAL LANGUGAE PROCESSING

SECTION A – Very Short Answer Type

(2 × 10 = 20 marks)

a) Key challenges in processing human language

Human language is ambiguous, context-dependent, and constantly evolving. Challenges include ambiguity, sarcasm, polysemy, idioms, grammar variations, and lack of structured data.

b) Linguistic principles as foundation of NLP

Linguistics provides rules for syntax, semantics, phonetics, and pragmatics, which help NLP systems understand sentence structure, meaning, sound, and context in text analytics.

c) How MaxEnt overcomes traditional limitations

Maximum Entropy (MaxEnt) models handle multiple overlapping features without independence assumptions, making them more flexible and accurate than rule-based or naive probabilistic models.

d) Transformation-based tagging vs other tagging

Transformation-based tagging starts with an initial tagging and iteratively applies learned rules to correct errors, unlike HMM or statistical taggers that rely purely on probabilities.

e) Semantic attachments

Semantic attachments link words or phrases to meanings based on context, helping NLP systems resolve ambiguity and understand relationships between sentence components.

f) Syntax contribution to semantic representation

Syntax defines sentence structure, which helps determine who did what to whom, enabling accurate semantic interpretation and knowledge representation.

g) Effect of vocal tract shape on speech

The shape of the vocal tract influences formant frequencies, affecting pitch, tone, and clarity of speech sounds.

h) Role of articulators in speech

Lips produce bilabial sounds, tongue shapes vowels and consonants, and the soft palate controls nasal sounds like “m” and “n”.

i) Warped frequency scale significance

Warped frequency scales (like Mel scale) match human hearing perception, improving accuracy in speech feature extraction such as MFCCs.

j) DTW and time misalignment

Dynamic Time Warping aligns speech signals of different lengths to compare patterns, but it is computationally expensive and sensitive to noise.

SECTION B – Long Answer Type

(Attempt any three – 10 marks each)

2(a) Role of Stop Words in Text Analytics and NLP

Stop words are common words like is, the, and, of that add little semantic value. Removing them:

Reduces dimensionality

Improves processing speed

Enhances model focus on meaningful terms

However, removing stop words blindly can distort meaning in tasks like sentiment analysis.

2(b) Interpolation and Backoff in Language Modeling

Interpolation combines probabilities from multiple n-gram models.
Backoff uses lower-order n-grams when higher-order data is missing.

These techniques:

Handle data sparsity

Improve prediction accuracy

Enhance language model robustness

2(c) Word Sense Relations and Semantic Networks

Relations include:

Synonymy (car – automobile)

Antonymy (hot – cold)

Hypernymy (animal → dog)

Hyponymy (rose → flower)

Meronymy (wheel → car)

These relations form semantic networks like WordNet, enabling meaning-based NLP tasks.

2(d) Short-Time Fourier Transform (STFT)

STFT analyzes speech by dividing it into small time windows and applying Fourier Transform to each.

Advantages:

Captures time-varying speech features

Suitable for speech recognition

It balances time and frequency resolution effectively.

2(e) Cepstral Distance Techniques

Cepstral Distance: Measures similarity between speech frames

Weighted Cepstral Distance: Prioritizes important features

Filtering: Removes noise before feature extraction

Each method suits different noise and accuracy requirements.

SECTION C – Descriptive Answer Type

(Attempt one from each question set)

3(a) Language Variations and Their Impact on NLP

Regional dialects, slang, and cultural expressions introduce ambiguity and vocabulary diversity. NLP systems struggle with:

Informal language

Code-switching

Cultural references

Solutions include large datasets, multilingual models, and contextual embeddings.

3(b) Significance of Syntactic Parsing

Syntactic parsing identifies grammatical structure using:

Constituency parsing

Dependency parsing

It helps extract relationships, improves machine translation, and supports information extraction.

4(a) HMMs in POS Tagging

Hidden Markov Models use:

Transition probabilities (tag sequences)

Emission probabilities (word-tag relations)

They handle uncertainty and sequential data better than rule-based systems.

4(b) Rule-based vs Stochastic Tagging

Rule-based systems are interpretable but rigid.
Stochastic methods are adaptive and scalable but require large datasets.

Hybrid systems often give best results.

5(a) Supervised Word Sense Disambiguation

Uses labeled data to classify word meanings.
Challenges include:

Data scarcity

Domain dependency

Used effectively in search engines and chatbots.

5(b) Bootstrapping in WSD

Bootstrapping starts with small labeled data and expands using confident predictions, improving accuracy with minimal supervision.

6(a) Linear Predictive Coding (LPC)

LPC models speech using past samples to predict current signals.

Used in:

Speech compression

Speech synthesis

LPC coefficients capture vocal tract characteristics.

6(b) Articulatory vs Acoustic Phonetics

Articulatory phonetics studies speech production, while acoustic phonetics analyzes sound waves. Articulator movement directly affects acoustic features like frequency and amplitude.

7(a) Multiple Time-Alignment Paths

Considering multiple alignment paths improves robustness in noisy speech recognition by reducing alignment errors.

7(b) LPC vs PLP vs MFCC