A meta-learning approach to text emotion distribution learning from small sample

Load data

• load_data.py

Embedding

• bert.py: BERT embedding

• doc2vec_linear.py: doc2vec embedding

• doc2vec_tensor.py: tensor embedding

• doc2vec_infersent.py: InferSent embedding

Task partition

• cluster_doc2vec.py: doc2vec

• cluster_random.py: random

• cluster_tensor.py: tensor embedding

• cluster_infersent.py: InferSent

Methods

• cnn_joint.py: CNN

• infersent_joint.py: InferSent or doc2vec

• bert.py: BERT

• maml.py: meta-training, eval_maml.py: meta-adaption

Experiment templates

• cv_portion.py: cross validation

• example.sh: example shell

Note

• download LDLPackage-v1.2 and tensor_toolbox, which have their own licenses

• download SemEval Task #14 dataset

• the results of label distribution learning methods, e.g., PT-X, AA-X, SA-X, can be calculated with LDLPackage-v1.2/edl/cv_portion_ldl.m

• grid search can be conducted by cv_grids.py and LDLPackage-v1.2/edl/cv_grid.m

• a quick demo can be experimented with bash example.sh and LDLPackage-v1.2/edl/example.m

• if you want to use a different word embedding, first use compress_wv.py to get a compressed word embedding file and put it in the corresponding path

• some codes should be tuned a bit (set the right paths) to work, feel free to contact me if you have any question

• running the BERT method needs bert-as-service

• running the InferSent method needs InferSent

---

GNU GENERAL PUBLIC LICENSE