--embedding_no_context

Switch

--embedding_no_context

Description

Encodes each sentence/segment independently (without surrounding message context) when extracting features with --bert_model/--emb_model. Aliased as --emb_no_context, --bert_no_context.

Argument and Default Value

No argument. Boolean switch, default: False (context-aware encoding).

Details

Known issue — silently produces an empty feature table. In this mode, FeatureExtractor.addEmbTable() calls tokenizer.encode_plus(...) inside a bare except: that catches and swallows any error, printing Message pair/ message unreadable. Skipping this.... and continuing. Verified: running with --embedding_no_context completed "successfully" (exit 0, DLATK exits with success!) but the resulting feature table had 0 rows — every message ended up skipped by that catch-all, and the true underlying error is masked. The same command without this flag (context-aware mode) worked correctly and produced thousands of real feature rows. Also noticeably slower (~13s vs ~3s here) since it processes per-sentence rather than per-message.

Other Switches

Required Switches:

• --bert_model or --emb_model

Example Commands

Not usable as of this writing — see Known issue above. Omit this flag for working (context-aware) embeddings:

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