NLP & Explainable AI (XAI) — Quick Reference Notes

AI generated - These are quick reference notes meant to help reconstruct the concepts later rather than serve as a deep dive.

NLP Techniques — Quick Reference

TechniqueConceptOne-liner
Bag of Words (BoW)Sparse text representationRepresents text using word occurrence/counts while largely ignoring word order and context.
Naive BayesProbabilistic classifierUses Bayes’ theorem with a strong feature-independence assumption; historically common for text classification.
VADERLexicon-based sentiment analysisUses a predefined sentiment lexicon and rules to estimate positive, negative, neutral, and compound sentiment.
BERT EmbeddingsContextual representationRepresents words using surrounding context, so the same word can have different representations depending on usage.
ABSAAspect-Based Sentiment AnalysisDetermines sentiment toward individual aspects rather than assigning one sentiment to the entire text.
Zero-shot ClassificationNo task-specific examplesUses a pretrained/foundation model to perform classification without providing labeled examples.
Few-shot ClassificationIn-context examplesProvides a small number of examples in the prompt to demonstrate the desired classification behavior.

Evolution at a Glance

Lexicon / Rules
VADER
Classical ML
BoW → Naive Bayes
Contextual Language Models
BERT / Embeddings
Fine-grained NLP
ABSA
Foundation Models
Zero-shot / Few-shot

XAI — Post-Hoc Explainability

Post-hoc explainability means explaining or interpreting a model’s behavior after the model has already produced its prediction.

TechniqueCore QuestionOne-liner
Attention WeightsWhat did the model attend to?Inspects how strongly different tokens or regions attend to one another; attention should not automatically be treated as a causal explanation.
Saliency / GradientsWhich inputs most influenced the output?Measures how sensitive the model’s output is to particular input features, often using gradients.
LIMEWhat explains this particular prediction locally?Perturbs the input, observes prediction changes, and fits a simple interpretable model around that specific example.
SHAPHow much did each feature contribute?Uses Shapley-value ideas from game theory to attribute portions of a prediction to individual input features.

LIME vs SHAP — Example

Consider a sentiment model receiving:

"The camera is excellent but the battery is terrible."

Suppose the model predicts:

Negative: 0.82

Both LIME and SHAP can help explain this prediction, but they approach the problem differently.

LIME — Local Interpretable Model-agnostic Explanations

LIME tries to understand the model’s behavior around this particular input.

It creates many slightly modified versions of the input and observes how the model’s prediction changes.

For example:

Original:
"The camera is excellent but the battery is terrible."
→ Negative: 0.82

Remove "terrible":
"The camera is excellent but the battery is."
→ Negative: 0.18

Remove "excellent":
"The camera is but the battery is terrible."
→ Negative: 0.94

Remove "battery":
"The camera is excellent but the is terrible."
→ Negative: 0.61

From many such perturbations, LIME might observe:

terrible   → strongly pushes toward NEGATIVE
battery    → somewhat important
excellent  → pushes toward POSITIVE
camera     → relatively small effect

LIME then fits a simple interpretable model, often something like a linear model, around these observations.

The simple model is not trying to reproduce the behavior of the original model everywhere. It only tries to approximate the model near this particular input.

Hence the name:

  • Local — explains the model around one particular prediction.
  • Interpretable — uses a simpler model that humans can understand.
  • Model-agnostic — does not need to understand the internals of the original model; it only needs inputs and outputs.

Mental model:

Original input
Perturb the input many ways
Run each variation through the model
Observe how predictions change
Fit a simple local model
Identify locally important features

LIME = perturb the input and approximate the model’s behavior locally.


SHAP — SHapley Additive exPlanations

SHAP approaches explanation as a credit-allocation problem.

It is based on Shapley values from cooperative game theory. Imagine several players cooperating to produce an outcome. Shapley values try to answer:

How much credit should each player receive for the final result?

In machine learning:

Game Theory                 Machine Learning

Players                     Input features
   ↓                              ↓
Work together               Combine together
   ↓                              ↓
Produce outcome             Produce prediction
   ↓                              ↓
Shapley values              SHAP values

Return to the sentiment example.

Suppose the model’s baseline negative score is:

Baseline = 0.40

For our sentence it predicts:

Negative = 0.82

The features collectively moved the prediction from 0.40 → 0.82.

SHAP tries to determine how much each feature contributed to that movement.

Conceptually:

Baseline                         0.40

"terrible"                      +0.38
"battery"                       +0.12
"excellent"                     -0.14
"camera"                        +0.01
other features                  +0.05
                               ─────
Prediction                      0.82

The important question is: how do we decide that "terrible" deserves +0.38?

SHAP considers the contribution of the feature across different combinations of the other features:

How much does "terrible" contribute alone?

How much does "terrible" contribute
when "battery" is already present?

How much does "terrible" contribute
when "excellent" is present?

How much does "terrible" contribute
when "battery" and "camera" are present?

...

The Shapley-value calculation aggregates these marginal contributions across possible feature combinations to assign credit to each feature.

This also explains the word Additive in SHAP:

Baseline
    +
Feature contribution 1
    +
Feature contribution 2
    +
Feature contribution 3
    +
...
Explained prediction

Hence:

SHapley Additive exPlanations.

Mental model:

Prediction
Treat features as cooperating players
Consider different feature combinations
Measure each feature's marginal contribution
Calculate Shapley values
Assign contribution to each feature

SHAP = distribute the prediction’s credit among features using Shapley values.


LIME vs SHAP — The Distinction to Remember

                    MODEL
                  Prediction
              POST-HOC EXPLANATION
       ┌──────────────┼──────────────┐
       │              │              │
   Attention       Saliency       LIME / SHAP
       │              │              │
What was the      What strongly    Why this
model attending   influenced the   particular
to?               output?          prediction?
LIME
Perturb the input
Observe model predictions
Fit a simple LOCAL approximation
"Which features seem important
around this particular example?"


SHAP
Consider features in different combinations
Measure marginal contributions
Calculate Shapley values
"How much did each feature contribute
to this prediction relative to a baseline?"

The simplest distinction:

LIME = perturb and approximate locally.

SHAP = distribute prediction credit among features using Shapley values.

Four XAI One-Liners

  • Attention: What was the model looking at?
  • Saliency: What parts of the input strongly influenced the output?
  • LIME: Perturb the input and approximate this prediction locally with a simple model.
  • SHAP: Attribute the prediction among its input features using Shapley values.

Important: These techniques provide interpretations or attributions of model behavior. They do not necessarily reveal the model’s true internal reasoning.