--bert_model

Switch

--add_bert

Description

Adds a BERT feature table.

Argument and Default Value

No argument. Boolean switch, default: False.

Details

Aliases: --emb_model (--add_embedding, --add_emb_feat). Extracts contextual embedding features for every message using a Huggingface Transformer model (BERT by default) and aggregates them into a feature table, one row per group.

By default this uses bert-base-uncased (--bert_model), the mean of the second-to-last layer (--bert_layers), and mean word/message aggregation (--bert_msg_aggregation). Use --batch_size (default 32) to change how many messages are embedded per GPU batch.

Other Switches

Required Switches:

• -d, -t , -g

Optional Switches:

• --bert_model

• --bert_layers

• --bert_msg_aggregation

• --bert_layer_aggregation

Example Commands

Creates a default BERT feature table.

dlatkInterface.py -d dla_tutorial -t msgs -c user_id --add_bert