Davide Modolo 229297

Intent classification is a text classification task in which the objective is to assign an intent for a given sentence or utterance.

Utterance: Can you help me find out about flights?

Intent: InfoRequest

Slot filling is a sequence labelling task where the objective is to map a given sentence or utterance to a sequence of domain-slot labels.

Utterance: I want to travel from nashville to tacoma

Concepts: O O O O O B-fromloc.city_name O B-toloc.city_name

Implement a neural network that predicts intents and slots in a multitask learning setting.

Since the datasets are tiny, you have to train and test your model from scratch at least 5 times. Report average and standard deviation.

Datasets: ATIS and SNIPS

Goal: Improve baseline results by at least 2/3%:

• ATIS -> Slot F1: 92.0%, Intent Acc.: 94.0%

• SNIPS -> Slot F1: 80.0%, Intent Acc.: 96.0%

PROJECT TODO:

1. Implement baseline methods

2. Build different architectures (Seq2Seq, Bi-LSTM + CRF, etc.)

3. Try to use pre-trained models (e.g. BERT, GPT2, T5, etc.)

MY TODO:

• download and import datasets

• prepare validation dataset for ATIS (since only SNIPS has it) - ~10% of the train but intents with only one instance remain in training

• implement baseline methods

• implement architectures from scratch (PyTorch)

• implement pre-trained models (PyTorch) BERT & ERNIE

• data visualization

• write paper

project
│   README.md
│   NLU_Report_Modolo.pdf: report on this project
│   conll.py: script to evaluate results
│   modolo_davide.ipynb: python notebook containing the baseline model, the bi-directional one and ED
│   pretrainedBERT.ipynb: python notebook containing the BERT model
│   pretrainedERNIE.ipynb: python notebook containing the ERNIE model
│
└───data
    └───ATIS
    │   test.json
    │   train_full.json
    │   train.json
    │   valid.json
    │
    └───SNIPS
        test.json
        train.json
        valid.json