InternshipAll 30 tasks, level by level
The real graded Natural Language Processing tasks the portal assigns, grouped into three levels.
The portal assigns these Natural Language Processing tasks. Finish Level 1 before Level 2 and Level 2 before Level 3. Your mentor may mark some tasks as required for the certificate.
Level 1: Foundations (12 tasks)
| Task | What to do |
|---|---|
| What NLP is | In a repo named nlp-internship, explain NLP and 6 real uses (search, chatbots, spam filters, translation, sentiment, summarising) with an Indian example for each. |
| Text cleaning | Clean 100 messy social-media comments in Python: lowercase, remove URLs, emojis and punctuation, and show before and after. |
| Tokenization | Tokenize English and Tamil or Hindi sentences with NLTK or spaCy and a subword tokenizer, and compare the tokens. |
| Stop words, stemming and lemmatization | Apply stop-word removal, stemming and lemmatization to 20 sentences and explain when each helps or hurts. |
| Word frequencies and n-grams | Find the top words, bigrams and trigrams in a set of product reviews and plot them as a bar chart and a word cloud. |
| Regular expressions for text | Extract emails, Indian phone numbers, PIN codes and dates from 50 lines of sample text using regular expressions. |
| Bag of words and TF-IDF | Turn 200 documents into bag-of-words and TF-IDF vectors with scikit-learn and show the top TF-IDF words per document. |
| Spam classifier | Train a Naive Bayes spam classifier on the free SMS Spam dataset and report accuracy, precision, recall and the confusion matrix. |
| Part-of-speech and named entities | Run spaCy on 10 news paragraphs and list the people, organisations, places and dates it finds. Mark its mistakes. |
| Word embeddings | Load pretrained word vectors (GloVe or fastText), find similar words and solve 5 analogies like king - man + woman. |
| Sentiment with a pretrained model | Use a Hugging Face sentiment pipeline on 50 reviews and compare its labels with your own. |
| Language detection | Detect the language of 30 sentences mixing English, Tamil, Hindi and Tanglish, and note where detection fails. |
Level 2: Building real features (10 tasks)
| Task | What to do |
|---|---|
| Fine-tune a text classifier | Fine-tune DistilBERT or a small multilingual model on a labelled dataset (news topics or reviews) and compare it with your TF-IDF model. |
| Named entity recognition model | Label 200 sentences with your own entity types (for example COURSE and COLLEGE) and train a spaCy NER model. |
| Text summarisation | Summarise 10 long articles with an extractive method and a transformer model, and rate both summaries. |
| Semantic search | Build search over 500 FAQs using sentence embeddings so users find answers even with different words. |
| Topic modelling | Find topics in 1,000 customer complaints with LDA or BERTopic and name each topic. |
| Indian languages | Classify or translate text in one Indian language using IndicNLP resources or a multilingual model, and report the results. |
| Question answering | Build an extractive question-answering demo over your college prospectus using a pretrained QA model. |
| Evaluation metrics | Explain and compute accuracy, F1, BLEU and ROUGE on your models' outputs, and say which metric fits which task. |
| Text API | Serve your best classifier as a FastAPI endpoint with input length limits and clear error messages. |
| Chatbot intents | Build an intent classifier for a college help bot with 10 intents and fallback replies, and test it with 50 messages. |
Level 3: Professional work and capstone (8 tasks)
| Task | What to do |
|---|---|
| Speech to text | Transcribe 10 audio clips with Whisper (including Indian-English accents), measure the word error rate and clean the transcripts. |
| Bias check | Test your sentiment model for bias with sentences that differ only by name, gender or region, and report what you find. |
| Error analysis | Find 30 mistakes your classifier makes, group them by cause (sarcasm, negation, mixed language) and improve one group. |
| Model optimisation | Make your transformer model faster with distillation, quantization or ONNX export and compare speed and accuracy. |
| Web app | Build a Streamlit or Gradio app for your NLP model and host it free on Hugging Face Spaces. |
| Data and privacy | Document how your dataset was collected, its licence and how personal data (names, phones) is removed before training. |
| Model card | Write a model card: data, training, metrics, limits, languages supported and intended use. |
| Capstone: NLP product | Ship a complete NLP app (review analyser, complaint router, multilingual FAQ bot or resume screener) with a live link, repo and a 2-minute demo video. |