The English MIMIC lemmatizer fails to initialize with a KeyError

The English MIMIC lemmatizer (package="mimic", processors="tokenize,pos,lemma") fails to initialize with a KeyError when loading the mimic_nocharlm lemma model.

The issue occurs because the model checkpoint does not contain the optional configuration keys:

charlm_forward_file charlm_backward_file

However, the loader assumes these keys are always present and accesses them with dictionary indexing (self.args[...]), which raises a KeyError for the mimic_nocharlm checkpoint.

To Reproduce Install Stanza (tested with versions 1.13.0 and 1.14.0). Download the English MIMIC models. Run:

import stanza

nlp = stanza.Pipeline(
    lang="en",
    package="mimic",
    processors="tokenize,pos,lemma",
    use_gpu=False,
)

The pipeline fails during initialization with: KeyError: 'charlm_forward_file'

The traceback points to:

stanza/models/lemma/trainer.py

where the loader executes code equivalent to:

self.args["charlm_forward_file"] = args.get(
    "charlm_forward_file",
    self.args["charlm_forward_file"],
)

Expected behavior

The mimic_nocharlm checkpoint should load successfully.

Since this checkpoint is explicitly a no-CharLM model, missing charlm_forward_file and charlm_backward_file entries should be treated as optional and default to None.

Replacing dictionary indexing with .get() resolves the issue without changing the model's behavior.

For example:

self.args["charlm_forward_file"] = args.get(
    "charlm_forward_file",
    self.args.get("charlm_forward_file"),
)

self.args["charlm_backward_file"] = args.get(
    "charlm_backward_file",
    self.args.get("charlm_backward_file"),
)

before these values are accessed.

After applying this change, the MIMIC lemmatizer loads successfully and produces normal predictions.

Environment: OS: macOS Python version: 3.10.10 Stanza version: 1.13.0 and 1.14.0