Richard Zowalla created OPENNLP-1978:
----------------------------------------
Summary: SentenceVectorsDL does not truncate input to the model's
maximum sequence length
Key: OPENNLP-1978
URL: https://issues.apache.org/jira/browse/OPENNLP-1978
Project: OpenNLP
Issue Type: Bug
Reporter: Richard Zowalla
Fix For: 3.0.0
{{SentenceVectorsDL}} encodes the whole text and runs it through the model
without truncating to the model's maximum sequence length \(typically 512
positions for BERT\-style models\). A longer input fails inside ONNX Runtime
and surfaces as a generic inference failure. With batched execution one
over\-long input fails the whole batch.
Suggested fix: truncate to the maximum length by default \(keeping {{\[SEP\]}}
at the end\), with an option to reject over\-long input with an
{{IllegalArgumentException}} instead, and document the behavior.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
