Note 1: Datasets are provided for all projects
Note 2: Not all datasets contain train/dev/test splits. It is up to the group members to decide a suitable split in those cases (or cross-validation).
NATURAL LANGUAGE PROCESSING ———————————————————————-
All datasets (including references) for the Natural Language Processing projects can be downloaded from: https://github.com/cardiffnlp/tweeteval
The main theme of the NLP projects is social media, and in particular Twitter.
1. Emotion recognition
This task consists of recognizing the emotion evoked by a tweet. The dataset is based on the SemEval2018 task “Affects in Tweets” (Mohammad et al. 2018).
2. Emoji prediction
The goal of this task is to predict an emoji (e.g. 🤔) given a tweet. The dataset based on the SemEval 2018 task on emoji prediction (Barbieri et al. 2018).
3. Hate speech detection
Given a tweet or a piece of comment, the task of hate speech consists of predicting whether the given text represents hate speech or not, and classify it accordingly. This task is based on the SemEval 2019 task on detection of hate speech against immigrants and women in Twitter (Basile et al. 2019).
4. Irony detection
This task consists of recognizing whether a tweet includes ironic intents or not. We use the Subtask A dataset of the SemEval2018 Irony Detection challenge (Van Hee et al., 2018).
5. Offensive language identification
This task consists in identifying whether some form of offensive language is present in a tweet. The task is based on the SemEval2019 OffensEval dataset (Zampieri et al., 2019).
6. Sentiment analysis
The goal for the sentiment analysis task is to recognize if a tweet is positive, negative or neutral. We use the Semeval-2017 dataset for Subtask A (Rosenthal et al., 2019).
7. Stance detection
Stance detection is the task to determine, given a piece of text, whether the author has a favourable, neutral, or negative position towards a proposition or target. We use the SemEval-2016 shared task on Detecting Stance inTweets (Mohammad et al., 2016). In the original task, five target domains are given: abortion, atheism, climate change, feminism and Hillary Clinton. Unlike the other tasks, training is provided separately for each target domain.
COMPUTER VISION ———————————————————————-
1. Fine-grained image classification (Birds)
The goal of this task is to develop an algorithm to learn to classify images containing objects of the same category (e.g. birds, dogs) into specific sub-categories, i.e. specific species. The dataset is available:
Caltech-UCSD Birds-200-2011: 200 categories of birds
http://www.vision.caltech.edu/visipedia/CUB-200-2011.html
Stanford dogs: 120 categories of dogs
http://vision.stanford.edu/aditya86/ImageNetDogs/
2. Fine-grained image classification (Dogs)
The goal of this task is to develop an algorithm to learn to classify images containing objects of the same category (e.g. birds, dogs) into specific sub-categories, i.e. specific species. The dataset is available:
Stanford dogs: 120 categories of dogs
http://vision.stanford.edu/aditya86/ImageNetDogs/
3. Image emotion recognition
Images often trigger different emotions to the viewer. The goal of this task is to develop an automatic algorithm to recognise the emotion that a specific image exhibits.
Dataset available at: https://cf-my.sharepoint.com/:f:/g/personal/laiy4_cardiff_ac_uk/ElhtEyhDRa1GqeD_Y43ZSL8BcBRiKZ8N 3kNgdZa79ZNaaw?e=xlvJPE
Some explanation for the dataset: https://arxiv.org/pdf/1605.02677.pdf
4. Detection and recognition of traffic signs
Detection and recognition of traffic signs is an important task for autonomous driving. The task is to identify traffic signs in images and recognise them.
Dataset available at https://sid.erda.dk/public/archives/daaeac0d7ce1152aea9b61d9f1e19370/published-archive.html (German Traffic Sign Recognition Benchmark)
5. Object localisation
It is straightforward for people to locate objects in an image, but can you develop a system to learn to do this? Given an image, the task is to find all the instances of relevant objects (as bounding boxes).
Dataset available at http://host.robots.ox.ac.uk/pascal/VOC/voc2012/index.html
It contains multiple datasets, and you have the flexibility to choose the ones that satisfy your project needs. So the following are just suggestions, and you may come up with your own approach, as long as it is reasonable and well justified.
For object detection/localisation, the most relevant dataset is the one used for Detection: Predicting the bounding box and label of each object from the twenty target classes in the test image.
You can see example images for this task at
http://host.robots.ox.ac.uk/pascal/VOC/voc2012/examples/index.html