7.1.3 Word Embeddings and Semantic Representation

One Sentence

Tokenization gives the model discrete IDs. Embedding turns those IDs into vectors so the model can compare, combine, and move meaning through layers.

The Mental Model

One-hot IDs can tell words apart, but they cannot tell which words are related. Dense embeddings place tokens in a vector space:

token id→embedding table lookup→dense vector

In that space:

• nearby vectors often mean related usage;
• cosine similarity measures direction similarity;
• sentence vectors are usually produced by pooling token vectors;
• contextual models can make the same token move depending on nearby words.

From One-Hot to Dense Vectors

With one-hot vectors, every different word is equally different:

refund   -> [1, 0, 0, 0]
return   -> [0, 1, 0, 0]
password -> [0, 0, 1, 0]
coupon   -> [0, 0, 0, 1]

Dense vectors can encode useful geometry:

refund  and return   -> close
password and reset   -> close
refund  and password -> far

This geometry is learned from data, not hand-written. Words that appear in similar contexts tend to get similar vectors.

Lab 1: Compare Word Similarity

Run this tiny embedding table. The numbers are hand-made for learning, but the operations are the same as real embedding retrieval.

from math import sqrt

embeddings = {
    "refund": [0.90, 0.80, 0.10],
    "return": [0.88, 0.78, 0.12],
    "password": [0.10, 0.20, 0.95],
    "reset": [0.12, 0.18, 0.92],
    "order": [0.75, 0.70, 0.15],
    "coupon": [0.05, 0.95, 0.10],
    "policy": [0.82, 0.74, 0.18],
}


def cosine(a, b):
    dot = sum(x * y for x, y in zip(a, b))
    norm_a = sqrt(sum(x * x for x in a))
    norm_b = sqrt(sum(x * x for x in b))
    return dot / (norm_a * norm_b)


print("refund vs return  :", round(cosine(embeddings["refund"], embeddings["return"]), 3))
print("refund vs password:", round(cosine(embeddings["refund"], embeddings["password"]), 3))
print("password vs reset :", round(cosine(embeddings["password"], embeddings["reset"]), 3))

Expected output:

refund vs return  : 1.0
refund vs password: 0.293
password vs reset : 1.0

Interpretation:

• high cosine means similar direction, not identical meaning;
• refund and return are close because this toy table puts them in the same customer-service region;
• password and reset are close for the same reason;
• refund and password are far because they serve different intents.

Lab 2: Build a Tiny Semantic Retriever

Now average token vectors to create sentence vectors, then rank three documents for a query.

def mean_embedding(tokens):
    vectors = [embeddings[token] for token in tokens if token in embeddings]
    dim = len(vectors[0])
    return [sum(vector[i] for vector in vectors) / len(vectors) for i in range(dim)]


query = mean_embedding(["reset", "password"])
documents = {
    "A refund policy": ["refund", "policy"],
    "B password reset": ["password", "reset"],
    "C coupon return": ["coupon", "return"],
}

ranked = sorted(
    (
        (name, cosine(query, mean_embedding(tokens)))
        for name, tokens in documents.items()
    ),
    key=lambda item: item[1],
    reverse=True,
)

for name, score in ranked:
    print(f"{name}: {score:.3f}")

Expected output:

B password reset: 1.000
C coupon return: 0.335
A refund policy: 0.333

This is the core of vector retrieval:

query text→query vector→compare with document vectors→top-k results

Real RAG systems use stronger embedding models and vector databases, but the logic is still similarity ranking.

Why Averaging Is Useful but Limited

Mean pooling is easy to understand, but it loses important information:

• word order;
• negation;
• emphasis;
• long-range dependency;
• which token should matter most.

For example, reset password and password reset become identical in the toy retriever. That is acceptable for a first intuition, but not enough for reasoning-heavy tasks.

Contextual Representations

Static embeddings usually give one word one vector. Contextual models make the vector depend on surrounding words:

bank account -> bank moves toward finance
river bank   -> bank moves toward geography

Run this small simulation:

base_bank = [0.50, 0.50, 0.50]
finance_context = [0.30, -0.10, 0.20]
river_context = [-0.20, 0.25, -0.10]

bank_in_finance = [a + b for a, b in zip(base_bank, finance_context)]
bank_in_river = [a + b for a, b in zip(base_bank, river_context)]

print("bank in finance:", [round(x, 2) for x in bank_in_finance])
print("bank in river  :", [round(x, 2) for x in bank_in_river])

Expected output:

bank in finance: [0.8, 0.4, 0.7]
bank in river  : [0.3, 0.75, 0.4]

This is not a real Transformer. It is a memory hook: the same token can end up with different representations after context is mixed in.

Project Uses

Use case	What embedding provides	Watch out for
RAG retrieval	find semantically related chunks	bad chunks or stale metadata still hurt answers
FAQ clustering	merge similar questions	close does not always mean duplicate
Deduplication	find near-duplicate content	paraphrases and templates can confuse scores
Classification	turn text into features	labels and calibration still matter
Recommendation	compare users, items, or queries	popularity bias can dominate similarity

Debugging Checklist

• Normalize vectors before cosine similarity if your library does not do it.
• Print top-k scores, not only top-1; a weak margin means retrieval is uncertain.
• Inspect false positives: related terms are not always correct answers.
• Compare static, sentence, and contextual embeddings for the same data.
• For multilingual projects, test cross-language pairs before assuming the embedding model aligns languages well.

Evidence to Keep

Keep this page’s proof of learning as a small evidence card:

Vectors

at least three text embeddings or toy vectors

Similarity Check

closest pair and score

Retrieval Result

top match for one query

Limitation

averaging or similarity misses context/negation/order

Next Use

this becomes retrieval evidence in Chapter 8

Exercises

1. Move coupon closer to password in the toy table. How does retrieval break?
2. Add a document D recover account and create vectors for recover and account.
3. Make a query refund order. Which document should rank first?
4. Explain why doctor and hospital may be close even though they are not synonyms.
5. In a RAG project, what evidence would you collect to prove your embedding model is good enough?

Project reference and review notes

1. If coupon moves close to password, similarity search may retrieve unrelated promotion or return-policy text for account-recovery queries. The failure is not random; it comes from bad geometry.
2. recover and account should be placed near password/account-support concepts, not near unrelated promotion or return-policy concepts. The added document should become a plausible match for account-recovery queries.
3. refund order should rank the refund/order document first if the embedding space captures both commerce and refund intent.
4. doctor and hospital are close because they often appear in the same domain. Similarity can mean topical relation, not strict synonymy.
5. Useful evidence includes a fixed query set, expected top-k documents, retrieval scores, known failure cases, latency, cost, and examples where wording changes but intent stays the same.

Summary

Embedding turns discrete token IDs into geometry:

identity→vector→distance→retrieval / clustering / model input

The deeper idea is not the formula. It is that meaning becomes something you can compare, rank, and pass through a neural network.