Towards Explainable Table Interpretation Using Multi-view Explanations

ExplainTI, a framework of explaining table interpretation using multi-view explanations. ExplainTI consists of two phases: tables are converted to sequences and lightweight column graphs firstly, then a pre-trained transformer encoder is fine-tuned to aggregate contextual information and provide multi-view explanations: (i) Local Explanations, are the most relevant phrases (w.r.t. pairwise phrases) in columns (w.r.t. column pairs) for the predictions; (ii) Global Explanations are the most influential samples from training data, which tend to have similar labels to the input sample; (iii) Structural Explanations, we design a relatively small-scale but highly effective method to construct column graphs for tables, which can not only aggregate contextual information but also provide explanations from structural view via graph attention mechanism.

Framework

Requirements

• Python 3.7.11
• PyTorch 1.9.0+cu111
• HuggingFace Transformers 4.11.3
• faiss 1.7.1
• scikit-learn 1.0

Datasets

We used two real world large-scale benchmark datasets from different types for evaluation.

WikiTable

The WikiTable dataset is Web tables collected from Wikipedia. We used the same train/valid/test splits as TURL.

GitTable

The GitTable dataset is the first large-scale relational table corpus. We used its subset organism and split it into train, valid, and test using the radio of 8:1:1.

The raw dataset is from GitTable. The code of preprocessing is in gittable.py and the preprocessed data is in data/GitTable.

Build the column graph

python build_graph.py [<args>] [-h | --help]

e.g.

python build_graph.py --data_name=GitTable --sample_size=16

Training with ExplainTI

To train the model and get the explanations with ExplainTI:

python train.py [<args>] [-h | --help]

e.g.

python train.py --le=0.1 --ge=0.1 --se --save_model --data_name=GitTable --path=./data/GitTable --batch_size=160 --column_length=64 --num_type=1141

The meaning of the flags:

• --data_name: the name of the dataset. e.g. WikiTable
• --path: the local path of the dataset. e.g. ./data/WikiTable
• --column_length: maximum length of a column after being converted to a sequence. e.g. 64
• --n_epoch: the number of epochs. e.g. 40
• --lr: the learning rate. e.g. 5e-5
• --model: the language model to be fine-tuned. e.g. bert-base-uncased
• --use_large: whether use large version of the model
• --le: the loss weight of the local view. e.g. 0.05
• --ge: the loss weight of the global view. e.g. 0.05
• --se: a flag indicating whether turn on the structural view
• --window_size: the number of tokens in a window. e.g. 8
• --top_k: the number of explanations generated for each view. e.g. 10
• --num_type: the number of types. e.g. 255
• --num_relation: the number of relations. e.g. 121
• --attention_method: the aggregate method for graph attention. e.g. dot
• --save_model: if this flag is on, then save the checkpoint to {save_path}/{name}.pt.
• --save_path: see --save_model. e.g. ./checkpoint
• --update_epoch: indicates how often the global embeddings store is updated. e.g. 10
• --cell_duplicated: performs cell-level deduplication or not.