Interesting NLP Tools: PII, Sentiment Analysis...

August 3, 2024

The first time I got the chance to use a Natural Language Processing model was with detoxify, which classifies comments with PyTorch and Transformers.

Kind of sentiment analysis tool.

But there are other interesting tools.

Btw, this is how they fit in the AI Landscape:

ML 101

Name Entity Recognition

A NER (Named Entity Recognition) system is like a smart highlighter that goes through this sentence and picks out the important “things” or “entities” and tells you what category they belong to.

Imagine you have a sentence like:

“Elon Musk visited Berlin on Monday to meet with Tesla executives.”

In our example, the NER system would highlight:

Elon Musk and say “This is a PERSON”
Berlin and say “This is a GPE” (which stands for Geopolitical Entity, like a city or country)
Monday and say “This is a DATE”
Tesla and say “This is an ORG” (which stands for Organization)

So, in simple terms, NER is a way for computers to:

Identify specific words or phrases in text.
Categorize those words or phrases into predefined groups like people, places, dates, organizations, etc.

Think of it as teaching a computer to understand the key players, locations, and times mentioned in a piece of writing. This is really useful for things like:

News analysis: Quickly finding out who did what, where, and when.
Customer service: Understanding what products or issues a customer is talking about.
Information extraction: Automatically pulling out important details from documents.

Basically, NER helps computers make sense of text by recognizing the “who, what, when, and where” of it.

SpaCy

Soacy can be used to detect PII information!

https://pypi.org/project/spacy/

You need to download the model first:

python -m spacy download en_core_web_sm

Example to use SPacy for PII:

import spacy

# Load the spaCy model for English
nlp = spacy.load("en_core_web_sm")

# Define a list of strings to check
strings_to_check = [
    "John Doe",          # Person
    "john.doe@example.com",  # Email
    "123-45-6789",       # Social Security Number
    "Google",            # Organization
    "123 Main St.",      # Address
    "555-1234",          # Phone number
    "April 5, 1990",     # Date
    "Jane Smith",        # Another person
    "Credit card 4111 1111 1111 1111", # Credit card
    "Non-PII string"     # Random string
]

# Define PII entity labels that spaCy can recognize
PII_labels = ["PERSON", "ORG", "GPE", "DATE", "CARDINAL", "LOC", "MONEY"]

def check_pii(text):
    doc = nlp(text)
    entities = [(ent.text, ent.label_) for ent in doc.ents if ent.label_ in PII_labels]
    return entities if entities else "No PII found"

# Check each string in the list for PII
for string in strings_to_check:
    result = check_pii(string)
    print(f"String: '{string}' -> PII Detected: {result}")

Conclusions

Nowadays you can also use LLMs for this kind of PII detection tasks.

But…it is always great to keep handy this kind of more traditional NLP tools.

You can also have a look to Sketch, as it can make data cataloguing, like PII identification.

MIT | AI code-writing assistant that understands data content. Sketch is usable in seconds and doesn’t require adding a plugin to your IDE.

Other NERs

NER stands for Named Entity Recognition.

It is a subtask of information extraction that seeks to locate and classify named entities mentioned in text into predefined categories such as the names of persons, organizations, locations, expressions of times, quantities, monetary values, percentages, etc.

NER is not directly used for detecting Personally Identifiable Information (PII) or for sentiment analysis, but it can be an important component in those tasks. For instance:

PII Detection: While NER itself does not specifically detect PII, it can identify entities like names, addresses, or other specifics that might be considered PII. Additional rules or models are typically needed to specifically classify data as PII.
Sentiment Analysis: NER is generally not used directly in sentiment analysis. Sentiment analysis focuses on determining the attitude or emotion conveyed in a piece of text, such as positive, negative, or neutral sentiments. However, understanding what entities are being discussed can provide context that might be useful in a more nuanced analysis of sentiment.

NER is a foundational NLP tool that helps in structuring text for deeper analysis, which can be useful in tasks like PII detection and sentiment analysis, but it is not solely sufficient for these tasks.

Generalist and Lightweight Model for Named Entity Recognition (Extract any entity types from texts) @ NAACL 2024

A very simple framework for state-of-the-art Natural Language Processing (NLP)

FAQ

Interesting Offline Tools - Translations

https://github.com/LibreTranslate/LibreTranslate

agpl 3.0 | Free and Open Source Machine Translation API. Self-hosted, offline capable and easy to setup.

What about n-Grams?

In Natural Language Processing (NLP), an n-gram is a contiguous sequence of n items from a given sample of text or speech.

The items can be characters, words, or even syllables.

A breakdown of what n-grams are? 🚀

“n” represents the number of items:
- A 1-gram (or unigram) is a single item.
- A 2-gram (or bigram) is a sequence of two items.
- A 3-gram (or trigram) is a sequence of three items.
- And so on…
Items can be:
- Words: “The cat sat” would have the following word-based bigrams: “The cat”, “cat sat”.
- Characters: “hello” would have the following character-based trigrams: “hel”, “ell”, “llo”.
Purpose:
- N-grams are used for various NLP tasks, including:
  - Language modeling: Predicting the next word in a sequence.
  - Text classification: Categorizing text into different topics.
  - Spell checking: Identifying and correcting spelling errors.
  - Machine translation: Translating text from one language to another.
  - Sentiment analysis: Determining the emotional tone of text.
  - Information retrieval: Searching for relevant documents.
How they work:
- N-grams capture the statistical properties of a text.
- By analyzing the frequency of n-grams, models can learn patterns and relationships between words or characters.
- For example, in language modeling, a model might learn that the bigram “New York” occurs frequently, and therefore, when it sees “New”, it’s likely to predict “York”.

In essence, n-grams provide a simple yet powerful way to analyze and understand the structure of language.

F/OSS Gen AI Tools - Summer 2025

This document discusses various open-source tools and concepts related to generative AI.

Lobe Chat: A tool from LobeHub (https://lobehub.com/?utm_source=chat_preview&utm_content=brand_watermark, GitHub: https://github.com/lobehub/lobe-chat?tab=License-1-ov-file#readme).
GPT4All: A desktop application.
Ollama
KNIME: AI-powered data apps (https://www.youtube.com/watch?v=HN-IJdUP4kc).
node-llmatic: (https://github.com/fardjad/node-llmatic).

Key Concepts

Vector DBs: Databases optimized for storing and querying vector embeddings.
Context Window: The amount of information a model can consider when generating text.
Hugging Face Models: Models available on the Hugging Face Hub.
RAG (Retrieval Augmented Generation): A framework for enhancing LLMs with external knowledge.
Semantic Search: Search based on the meaning of the query.
Embeddings: Numerical representations of words or phrases (e.g., Word2Vec).
- Word2Vec: A neural network-based algorithm for learning word embeddings.
GGUF: A file format for storing large language models, successor to GGML.

Neural Network Architectures

CNNs (Convolutional Neural Networks):
- Used for processing structured array data like images.
- Key components: Convolutional Layers, Activation Functions (e.g., ReLU), Pooling Layers, Fully Connected Layers.
- Applications: Image classification, object detection, face recognition, etc.
RNNs (Recurrent Neural Networks):
- Designed for processing sequences of data.
- Used in time-series analysis, NLP, and speech recognition.
- Challenges: Struggle with long-range dependencies.
Transformers:
- Rely on attention mechanisms instead of recurrence.
- Handle long-range dependencies more effectively than RNNs.
- Foundation of models like BERT, GPT, and T5.
GPT (Generative Pre-trained Transformer):
- A type of neural network architecture for natural language processing.
- Pre-trained on large datasets and fine-tuned for specific tasks.
- Generates human-like responses to text prompts.
GANs (Generative Adversarial Networks):
- Two neural networks (Generator and Discriminator) contesting each other.
- Used for generating realistic images, videos, and voice recordings.
Diffusion Models:
- Focus on quality and diversity in generated outputs.
- (https://www.youtube.com/watch?v=sc_zOsnmMx0)
VAEs (Variational Autoencoders):
- Focus on speed and diversity.
Autoencoders:
- Used for unsupervised learning of efficient codings of unlabeled data.
- Structure: Encoder, Decoder, Bottleneck.
- Applications: Dimensionality reduction, feature learning, denoising, anomaly detection.
Prompting: Giving instructions to an AI system. Small changes can significantly affect the output.
- Zero-Shot: Performing a task without prior examples.

Using DJI Tello Drone with Python ChatBots with AI: Langflow and other OS Tools