--embedding_word_aggregation

Switch

--embedding_word_aggregation AGG [AGG ...]

Description

How to aggregate sub-word token embeddings back up to the word level, when extracting features with --bert_model/--emb_model (e.g. mean or concatenate). Aliased as --word_aggregation, --bert_word_aggregation.

Argument and Default Value

One or more aggregation-method strings. Default: ['mean'] (dlac.DEF_TRANS_WORD_AGGREGATION).

Details

Verified: ran successfully with concatenate as part of a full embedding-extraction command (the run got past feature computation and failed later only due to the unrelated --embedding_keep_msg SQLite issue — see the Known issue there).

Other Switches

Required Switches:

• --bert_model or --emb_model

Example Commands

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