Natural Language
Note:
Participants must complete at least
3 2
tasks for the -week internship and
Processing Tasks 4
—
p
tasks for the 1-month internshi
from any level.
Level 1
Task 1: Sentiment Analysis on Product Reviews
Description:
Dataset (Recommended): IMDb Reviews or Amazon Product Reviews (Kaggle
Analyze product reviews to determine whether the sentiment is positive or negative
Clean and preprocess text (e.g., lowercasing, removing stopwords)
Convert text to numerical format using TF-IDF or CountVectorizer
Train a binary classifier (e.g., logistic regression) and evaluate its performance
Tools & Libraries:
Python Pandas NLTK/spacy Scikit-learn
Covered Topics
Text classification | Sentiment analysis
Bonus:
Visualize the most frequent positive and negative words
Try using a Naive Bayes classifier and compare accuracy
2
Task : News Category Classification
Description:
G
Dataset (Recommended): A News (Kaggle
Classify news articles into categories such as sports, business, politics, technology, etc.
Perform standard text preprocessing (tokenization, stopword removal, lemmatization)
Vectorize the text using TF-IDF or word embeddings
Train a multiclass classifier (e.g., Logistic Regression, Random Forest, or SVM)
Tools & Libraries:
Pandas Scikit-learn Optionally: XGBoost or LightGBM
Covered Topics
Multiclass classification | Text preprocessing | Feature engineering with text
Bonus:
Visualize the most frequent words per category using bar plots or word clouds
Try training the model using a neural network (e.g., simple feedforward NN with Keras)
1
Natural Language
Processing Tasks
Level 2
Task 3: Fake News Detection
Description:
Dataset (Recommended): Fake and Real News Dataset (Kaggle)
Classify news articles as real or fake based on their text content
Preprocess title and content (remove stopwords, lemmatize, vectorize)
Train a logistic regression or SVM classifier
Evaluate using accuracy and F1-score
Tools & Libraries:
Python Pandas NLTK/spacy Scikit-learn
Covered Topics
Binary classification | TF-IDF and preprocessing
Bonus:
Use a word cloud to visualize common terms in fake vs. real news
Task 4: Named Entity Recognition (NER) from
News Articles
Description:
Dataset (Recommended): CoNLL003 (Kaggle)
Identify named entities (like people, locations, and organizations) from article content
Use rule-based and model-based NER approaches
Highlight and categorize extracted entities in the text
Tools & Libraries:
Python SpaCy Pandas
Covered Topics
Sequence labeling | NER (Named Entity Recognition)
Bonus:
Visualize extracted entities with displacy
Compare results using two different spaCy models
2
Natural Language
Processing Tasks
Level 2
Task 5: Topic Modeling on News Articles
Description:
Dataset (Recommended): BBC News Dataset (Kaggle)
Discover hidden topics or themes in a collection of news articles or blog posts
Preprocess the text: tokenization, lowercasing, stopword removal
Apply Latent Dirichlet Allocation (LDA) to extract dominant topics
Display the most significant words per topic
Tools & Libraries:
Python Gensim pyLDAvis NLTK/spacy Scikit-learn
Covered Topics
Topic modeling | Unsupervised NLP
Bonus:
Compare LDA vs. NMF performance
Use pyLDAvis or word clouds to visualize topic-word distributions
3
Natural Language
Processing Tasks
Level 3
Task 6: Question Answering with Transformers
Description:
Dataset (Recommended): SQuAD v1.1 - Stanford Question Answering (Kaggle
Build a system that answers questions based on a given context or passage
Use pre-trained transformer models (e.g., BERT or DistilBERT) fine-tuned for question answering
Feed the model both the context and the question, and extract the correct answer span
Evaluate with exact match and F1 score
Tools & Libraries:
Hugging Face Transformers Tokenizers Pandas
Covered Topics
Question answering | Span extraction | Transformer-based NLP
Bonus:
Try different base models (e.g., BERT, RoBERTa, ALBERT) and compare performance
Build a simple command-line or Streamlit interface to input a passage and a question
Task 7: Text Summarization Using Pre-trained Models
Description:
Dataset (Recommended): CNN-DailyMail News (Kaggle)
Generate concise summaries from long documents using NLP models
Use an encoder-decoder architecture (T5, BART, or Pegasus)
Preprocess long texts and truncate to model input limits
Evaluate summary quality using ROUGE scores
Tools & Libraries:
ROUGE Hugging Face Transformers Pandas
Covered Topics
Abstractive summarization | NLP with deep learning
Bonus:
Try extractive summarization using TextRank or Gensim
Fine-tune a pre-trained summarizer on a custom dataset
4
Natural Language
Processing Tasks
Industry Level
Task 8: Resume Screening Using NLP
Description:
Dataset (Recommended): Resume Dataset + Job Dataset (you will need both) (Kaggle)
Develop a system to screen and rank resumes based on job descriptions
Preprocess resumes and job descriptions using embeddings
Match resumes using cosine similarity or classification
Present top-ranked resumes with brief justifications
Tools & Libraries:
Sentence Transformers Pandas Scikit-learn
Covered Topics
Document similarity | Semantic search in NLP
Bonus:
Create a simple front-end to upload a resume and get matching results
Try named entity extraction from resumes (skills, experience)
Bonus Explanation:
User uploads a resume file (usually a .txt, .pdf, or .docx)
The system reads and processes the text from the resume
It compares that resume to a job description or a set of job requirements using NLP
(e.g., cosine similarity with embeddings)
The user then sees a matching score or result like
“Match Score: 85%
“Strong match for: Junior Data Analyst
or a list of highlighted skills that matched the job description
