Beam search decoding in NLP - Model Pipeline Trace

Model Pipeline - Beam search decoding

Beam search decoding is a method used in language models to find the most likely sequence of words by exploring multiple options at each step, keeping only the best few choices to balance quality and speed.

Data Flow - 5 Stages

1Input sequence

1 sequence x 1 token→Start with initial token (e.g., <start>)→1 sequence x 1 token

"<start>" token

↓

2Model prediction

1 sequence x 1 token→Model predicts probabilities for next tokens→1 sequence x vocabulary size probabilities

Probabilities for next words like {"the":0.3, "a":0.2, "cat":0.1, ...}

↓

3Beam expansion

beam width sequences x current length tokens→Expand each sequence by all possible next tokens, score them→beam width * vocabulary size sequences x (current length + 1) tokens

From 3 sequences, each expanded by 5 tokens → 15 sequences

↓

4Beam pruning

beam width * vocabulary size sequences x (current length + 1) tokens→Keep top beam width sequences with highest scores→beam width sequences x (current length + 1) tokens

Keep top 3 sequences out of 15

↓

5Repeat until end token

beam width sequences x tokens→Repeat expansion and pruning until <end> token or max length→beam width sequences x final length tokens

Final 3 sequences like ["<start> the cat <end>", "<start> a dog <end>", ...]

Training Trace - Epoch by Epoch

Loss
2.3 |****
1.8 |***
1.4 |**
1.1 |*
0.9 |

Epoch	Loss ↓	Accuracy ↑	Observation
1	2.3	0.25	High loss and low accuracy as model starts learning
2	1.8	0.40	Loss decreases and accuracy improves
3	1.4	0.55	Model learns better word predictions
4	1.1	0.65	Loss continues to decrease steadily
5	0.9	0.72	Model converges with improved accuracy

Prediction Trace - 5 Layers

Layer 1: Initial token input

Layer 2: Model predicts next token probabilities

Layer 3: Beam expansion with beam width=2

Layer 4: Beam pruning

Layer 5: Repeat until <end> token

Model Quiz - 3 Questions

What does beam width control in beam search decoding?

A Number of sequences kept at each step

B Length of the output sequence

C Size of the vocabulary

D Number of training epochs

Key Insight

Beam search balances exploring multiple possible sequences and focusing on the best ones, improving prediction quality compared to greedy search without excessive computation.

Practice

1. What is the main purpose of beam search decoding in natural language processing?

easy

A. To keep track of multiple best candidate sequences during prediction

B. To randomly select words for output generation

C. To generate only one possible output sequence

D. To speed up training by skipping steps

Solution

1. Step 1: Understand beam search goal

   Beam search keeps multiple candidate sequences to explore more options than greedy search.

2. Step 2: Compare options

   Only To keep track of multiple best candidate sequences during prediction describes keeping multiple best guesses; others describe random choice, single output, or unrelated speed-up.

3. Final Answer:

   To keep track of multiple best candidate sequences during prediction -> Option A

4. Quick Check:

   Beam search = multiple best sequences [OK]

Hint: Beam search tracks several top guesses, not just one [OK]

Common Mistakes:

• Confusing beam search with random sampling
• Thinking beam search outputs only one sequence
• Assuming beam search speeds up training

2. Which of the following is the correct way to describe the beam width in beam search decoding?

easy

A. The size of the vocabulary used for prediction

B. The number of candidate sequences kept at each decoding step

C. The length of the output sequence generated

D. The number of layers in the neural network

Solution

1. Step 1: Define beam width

   Beam width is how many top sequences the algorithm keeps at each step to explore.

2. Step 2: Eliminate incorrect options

   Output length, vocabulary size, and network layers are unrelated to beam width.

3. Final Answer:

   The number of candidate sequences kept at each decoding step -> Option B

4. Quick Check:

   Beam width = candidate count per step [OK]

Hint: Beam width = how many sequences you keep each step [OK]

Common Mistakes:

• Mixing beam width with output length
• Confusing beam width with vocabulary size
• Thinking beam width relates to model architecture

3. Consider a beam search with beam width 2 decoding a sequence. At step 1, the top 2 tokens have scores [0.6, 0.4]. At step 2, each token expands to two tokens with scores: from first token [0.5, 0.3], from second token [0.7, 0.2]. Which two sequences will beam search keep after step 2?

medium

A. [First token + second expansion (0.6*0.3), Second token + second expansion (0.4*0.2)]

B. [First token + first expansion (0.6*0.5), First token + second expansion (0.6*0.3)]

C. [Second token + first expansion (0.4*0.7), Second token + second expansion (0.4*0.2)]

D. [First token + first expansion (0.6*0.5), Second token + first expansion (0.4*0.7)]

Solution

1. Step 1: Calculate scores for all expansions

   Calculate combined scores: 0.6*0.5=0.3, 0.6*0.3=0.18, 0.4*0.7=0.28, 0.4*0.2=0.08.

2. Step 2: Select top 2 sequences by score

   Top two scores are 0.3 and 0.28, corresponding to first token + first expansion and second token + first expansion.

3. Final Answer:

   [First token + first expansion (0.6*0.5), Second token + first expansion (0.4*0.7)] -> Option D

4. Quick Check:

   Top scores = 0.3 and 0.28 [OK]

Hint: Multiply scores, pick top beam width sequences [OK]

Common Mistakes:

• Choosing expansions only from one token
• Not multiplying scores correctly
• Picking lower scoring sequences

4. You implemented beam search decoding but notice it always returns the same output sequence regardless of input. What is the most likely bug?

medium

A. The vocabulary size is too large

B. The model is not trained

C. Beam width is set to 1, making it greedy search

D. The beam search is not normalizing scores

Solution

1. Step 1: Analyze symptom of identical outputs

   Always same output suggests no exploration of multiple sequences.

2. Step 2: Identify beam width effect

   If beam width = 1, beam search reduces to greedy search, always picking highest scoring token only.

3. Final Answer:

   Beam width is set to 1, making it greedy search -> Option C

4. Quick Check:

   Beam width 1 = greedy search [OK]

Hint: Check beam width; 1 means no beam search [OK]

Common Mistakes:

• Blaming vocabulary size for output sameness
• Ignoring beam width setting
• Assuming model training causes identical outputs

5. In a machine translation task, you want to balance output quality and decoding speed. You have a beam search decoder with beam width 5. What happens if you increase the beam width to 20?

hard

A. Output quality may improve but decoding will be slower

B. Output quality will decrease and decoding will be faster

C. Output quality and decoding speed remain the same

D. Decoding speed improves but output quality is unpredictable

Solution

1. Step 1: Understand beam width effect on quality

   Larger beam width explores more sequences, often improving output quality.

2. Step 2: Understand beam width effect on speed

   More sequences to track means more computation, slowing decoding speed.

3. Final Answer:

   Output quality may improve but decoding will be slower -> Option A

4. Quick Check:

   Higher beam width = better quality, slower speed [OK]

Hint: Bigger beam = better results but slower decoding [OK]

Common Mistakes:

• Assuming bigger beam always speeds decoding
• Thinking quality decreases with bigger beam
• Believing beam width doesn't affect speed