No software in the context

[Samuel-Scalbert]

For the second time when i process the file hal-03882318 wiht softcite's docker i have this mention with no software in the context:

{
  "type": "software",
  "software-type": "software",
  "software-name": {
    "rawForm": "nnU-Net",
    "normalizedForm": "nnU-Net",
    "offsetStart": 49,
    "offsetEnd": 56
  },
  "context": "best network configuration. In the end, the model (or ensemble) which got the best pe",
  }
}

[jameshowison]

That is odd. Can you point to the pdf you are processing?

[Samuel-Scalbert]

yes of course here is the link : https://hal.science/hal-03882318/document

[willbeason]

The paragraph in question:

We randomly split all 23 patients into four folds using a cross-validation scheme. Table 1 presents the partition of the
dataset. Multiple images (i.e., planning and daily images) from individual patients were not distributed among datasets. The
test data was only used to evaluate the performance of the model in this fold and was not involved in training. nnU-Net further
divided the training data into training and validation sets and performed a five-fold cross-validation to automatically select the
best network configuration. In the end, the model (or ensemble) which got the best performance was chosen to perform the
inference on the test sets of this fold. The number of epochs during training was 1000 for every fold. The evaluation of the
segmented volumes is described in part 2.4.1. Table 2 reports the network configurations generated by nnU-Net for the
considered dataset.

So nnU-Net does appear in the paragraph, and the paragraph (at least in PDF form) includes linebreaks for each line of text. In this case, "best network configuration" is at the end of a sentence beginning "nnU-Net" but is at the beginning of a line of text. I wonder if possibly sentence segmentation logic got confused by the multiple line breaks in the middle of the sentence, or the fact that the sentence begins with a lowercase letter?

I'd expect the context to be:

nnU-Net further divided the training data into training and validation sets and performed a five-fold cross-validation to automatically select the best network configuration.

(line breaks removed)

Another possibility (or coincidence) is that the sentence beginning with "nnU-Net" is the fifth sentence in the paragraph, and it is the fifth line of text that was chosen as the context.

[lfoppiano]

With the version lfoppiano/software-mentions:0.8.2 deployed also here, the issue is not occurring.

[jameshowison]

Just ran this today, using latest source for software-mentions (and grobid). Indeed the error doesn't occur, although that's perhaps a side-effect of not recognizing nnU-Net as software ... I only get a mention for Elastix on page 7 ...

Looking at the text I think this is due to annotation guidelines around models vs software. While nnU-Net is instantiated in a software package (with that name) the text refers to the model. But there are contextual clues that something is executing (e.g., "automatic configuration runs").

Will keep this one open and use it in future annotation.

The segmentation model was trained using the deep learning-based method nnU-Net, which has been demonstrated to be efficient in multiple medical image segmentation tasks (Isensee et al 2021). nnU-Net's automatic configuration runs without
and post-processing. The 3D full resolution U-Net was used since it has been shown to be one of the best performing models in many medical image segmentation tasks (Malimban et al 2022, Isensee et al 2021).

Just exploring: If one changes method to tool then it is recognized.

[lfoppiano]

@jameshowison I think the main/master version uses wapiti by default (the huggingface's version uses delft) so it's possible that this is the reason why nnU-Net is not recognised.

[jameshowison]

Ah, that makes sense. I looked for how to enable delft on my local machine? Should I be following these instructions: https://grobid.readthedocs.io/en/latest/Deep-Learning-models/ or should I be building the docker image locally?

…

On Mon, Sep 15, 2025 at 10:39 AM Luca ***@***.***> wrote: *lfoppiano* left a comment (softcite/software-mentions#39) <#39 (comment)> @jameshowison <https://github.com/jameshowison> I think the main/master version uses wapiti by default (the huggingface's version uses delft) so it's possible that this is the reason why nnU-Net is not recognised. — Reply to this email directly, view it on GitHub <#39 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAWOUWBKPRGFCRGVFHHNUT3S3MSZAVCNFSM6AAAAACGQ5VU7OVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTEOJSHAYTGMBVHE> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[lfoppiano]

@jameshowison Let's discuss tomorrow whether it's worth it to build it locally.