Introduction to Natural Language Processing (i-NLP) · End Term · Sep 2024 · FN

“sleep” in “sleeping”

“un-” in “undo”

“book” in “booked”

“-s” in “books”

Computing Shapley values requires evaluating an exponential number of feature coalitions.

Shapley values assume independence between features, which may not hold in real-world data.

Shapley values are computationally efficient for deep neural networks.

Shapley values provide additive explanations, ensuring the contributions of all features sum up to the model's output.

FastText can generate embeddings for unseen words by summing the embeddings of their subword n-grams.

FastText considers only complete words and does not split them into subword n-grams.

FastText embeddings for morphologically similar words (e.g., "run" and "running") will be highly similar because they share many n-grams.

FastText cannot distinguish between semantically unrelated words like "mouse" and "house" because they share common n-grams(e.g., "ouse").

Only (1)

Only (2)

Both (1) and (2)

The model will completely ignore the future tokens during training.

The model will stop attending to any tokens in the sequence.

The model will incorrectly attend to future tokens, leading to information leakage and breaking the causality constraint.

The softmax function will fail to compute attention scores for any position.

RAG, Fine-tuning, Prompt engineering, Pre-trained model

Prompt engineering, RAG, Fine-tuning, Pre-trained model

RAG, Prompt engineering, Fine-tuning, Pre-trained model

Fine-tuning, RAG, Prompt engineering, Pre-trained model

Few-shot prompting

Chain-of-thought prompting

Zero-shot prompting

Instruction prompting

A-4, B-3, C-1, D-2

A-2, B-3, C-1, D-4

A-2, B-4, C-3, D-1

A-1, B-3, C-2, D-4

Part of speech tagging

Named entity recognition

Machine translation

Sentiment analysis

Beam search with beam width of 4

Top-k sampling

Nucleus sampling

Greedy sampling

Pre-processing mitigation

In-training mitigation

Intra-processing mitigation

Post-processing mitigation

Shapley Values

Integrated Gradients

Principal Component Analysis

Attention Mechanisms

It requires retraining the model for every subset of features.

It depends on the dataset size, which scales quadratically.

It evaluates the contribution of all possible feature subsets, which grows exponentially.

It computes gradients for every input sample.       **SPG** **Section Id :** 64065375426 **Section Number :** 10 **Section type :** Online **Mandatory or Optional :** Mandatory **Number of Questions :** 19 **Number of Questions to be attempted :** 19 **Section Marks :** 35 **Display Number Panel :** Yes **Section Negative Marks :** 0 **Group All Questions :** No **Enable Mark as Answered Mark for Review and** No **Clear Response :** **Section Maximum Duration :** 0 **Section Minimum Duration :** 0 **Section Time In :** Minutes