Category: Neural Networks

Author: Atharv Katkar

Directed by: Sandeep Giri

OVERVIEW:

In Natural Language Processing (NLP), embeddings transform human language into numerical vectors. These are usually arrays of multiple dimensions & have schematic meaning based on their previous training text corpus The quality of these embeddings directly affects the performance of search engines, recommendation systems, chatbots, and more.

But here’s the problem:

Not all embeddings are created equal.

So how do we measure their quality?

To Identify the quality of embeddings i conducted one experiment:

I took 3 leading (Free) Text → Embedding pretrained models which worked differently & provided a set of triplets and found the triplets loss to compare the contextual  importance of each one.


1) Sentence-BERT (SBERT)

Transformer-based Captures deep sentence-level semantics:

from sentence_transformers import SentenceTransformer

model = SentenceTransformer(‘all-MiniLM-L6-v2’)

2) Universal Sentence Encoder (USE)

TensorFlow Encoder Good general-purpose semantic encoding:

import tensorflow_hub as hub

model2 = hub.load(“https://tfhub.dev/google/universal-sentence-encoder/4”)

embeddings = model2([“Cardiac arrest”])

3) FastText (by Facebook AI)

Word-based Lightweight, fast, but lacks context:

import fasttext.util

fasttext.util.download_model(‘en’, if_exists=’ignore’)

ft3 = fasttext.load_model(‘cc.en.300.bin’)

vec = ft3.get_word_vector(“Cardiac arrest”)

when i compared the sizes of output produced by them are different for each one

(384,), (1, 512), (300,)

GOALS

1. To compare them using a triplet-based evaluation approach using triplet loss.
2. Identify the Understanding of these around medical terminologies

CONCEPTS

What is Triplet Loss?

Triplet loss works with a 3-part input:

Anchor: The base sentence or phrase

Positive: A semantically similar phrase

Negative: A semantically absurd or unrelated phrase

Anchor	Positive	Negative
tuberculosis	Lung infection	test tube accident
cardiac arrest	heart attack	cardi b arrest
asthma	respiratory condition	Spiritual awakening

The goal is to push the anchor close to the positive and far from the negative in embedding space.

TripletLoss = max (d (a , p) − d (a , n) + margin , 0 )

a = anchor vector

p = positive vector (should be close to

anchor)

n = negative vector (should be far from

anchor)

d(x,y) = cosine distance

margin = a buffer that forces the negative to be not just farther, but significantly farther

What is Cosine Similarity?

Cosine similarity is a measure of how similar two vectors are — based on the angle between them rather than their magnitude. In the context of NLP, vectors represent words or sentences as embeddings.

  Cosine Similarity (A.B) =   A.B / ||A||.||B||

CosineDistance(A,B) = 1 − CosineSimilarity(A,B)

What is Margin?

The margin is a safety cushion.

If margin = 0.2, then even if the negative is slightly farther than the positive, the model still gets a penalty unless it’s at least 0.2 farther.

Testing The accuracy

TEST-SET(click on test set see set)

We ran each model over a set of ~50 curated triplets

Calculated:

Anchor–Positive distance (AP)

Anchor–Negative distance (AN)

Triplet loss

Visualized both individual performance per triplet and overall averages

(“asthma”, “respiratory condition”, “spiritual awakening”),

(“pneumonia”, “lung infection”, “foggy window”),

# General & Internal Medicine

(“diabetes”, “high blood sugar”, “candy addiction”),

(“arthritis”, “joint inflammation”, “rusty hinge”),

50+ such examples

Results:

Using PCA, we visualized where each model placed the anchor, positive, and negative in space.
Insert 2D scatter plot
You can actually see the anchor and positive clustering together especially in SBERT’s case while the negative floats far away. In below image

Triplet Example wise Score :

The SBERT & USE performed good as we can see using few interpretations and loss tracking of triplets

CONCLUSION:

What We Learned

SBERT is highly reliable for understanding sentence-level meaning

USE performs reasonably well and is easy to use with TensorFlow

FastText, while fast, struggles with context and full sentences

Visual Results

Triplet Loss (Lower = Better)

SBERT     : 0.0381

USE       : 0.0320

FastText  : 0.2175

If you’re building Search engines , Recommendation systems , Chatbots …or anything involving meaning, good embeddings are key. Triplet loss is a simple yet powerful way to test how smart your model really is. I insist all of the preprocessing stages of making NLP models or context based systems the Triplet Loss needs to be used to select optimal pretrained or trained model.