span-marker-bert-base-multilingual-uncased-multinerd

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

span-marker-bert-base-multilingual-uncased-multinerd

This model is a fine-tuned version of bert-base-multilingual-uncased on an Babelscape/multinerd dataset.

Is your data always capitalized correctly? Then consider using the cased variant of this model instead for better performance: lxyuan/span-marker-bert-base-multilingual-cased-multinerd.

This model achieves the following results on the evaluation set:

• Loss: 0.0054
• Overall Precision: 0.9275
• Overall Recall: 0.9147
• Overall F1: 0.9210
• Overall Accuracy: 0.9842

Test set results:

• test_loss: 0.0058621917851269245,
• test_overall_accuracy: 0.9831472809849865,
• test_overall_f1: 0.9187844693592546,
• test_overall_precision: 0.9202802342397876,
• test_overall_recall: 0.9172935588307115,
• test_runtime: 2716.7472,
• test_samples_per_second: 149.141,
• test_steps_per_second: 4.661,

Note: This is a replication of Tom's work. In this work, we used slightly different hyperparameters: epochs=3 and gradient_accumulation_steps=2. We also switched to the uncased bert model to see if an uncased encoder model would perform better for commonly lowercased entities like, such as food. Please check the discussion here. Refer to the official model page to review their results and training script.

Results:

Below is a combined table that compares the results of the cased and uncased models for each language:

Short discussion: Upon examining the results, one might conclude that the cased version of the model is better than the uncased version, as it outperforms the latter across all languages. However, I recommend that users test both models on their specific datasets (or domains) to determine which one actually delivers better performance. My reasoning for this suggestion stems from a brief comparison I conducted on the FOOD (food) entities. I found that both cased and uncased models are sensitive to the full stop punctuation mark. We direct readers to the section: Quick Comparison on FOOD Entities.

Label set

Inference Example

# install span_marker
(env)$ pip install span_marker


from span_marker import SpanMarkerModel

model = SpanMarkerModel.from_pretrained("lxyuan/span-marker-bert-base-multilingual-uncased-multinerd")

description = "Singapore is renowned for its hawker centers offering dishes \
like Hainanese chicken rice and laksa, while Malaysia boasts dishes such as \
nasi lemak and rendang, reflecting its rich culinary heritage."

entities = model.predict(description)

entities
>>>
[
  {'span': 'Singapore', 'label': 'LOC', 'score': 0.9999247789382935, 'char_start_index': 0, 'char_end_index': 9},
  {'span': 'laksa', 'label': 'FOOD', 'score': 0.794235348701477, 'char_start_index': 93, 'char_end_index': 98},
  {'span': 'Malaysia', 'label': 'LOC', 'score': 0.9999157190322876, 'char_start_index': 106, 'char_end_index': 114}
]

# missed: Hainanese chicken rice as FOOD
# missed: nasi lemak as FOOD
# missed: rendang as FOOD

# note: Unfortunately, this uncased version still fails to pick up those commonly lowercased food entities and even misses out on the capitalized `Hainanese chicken rice` entity.

Quick test on Chinese

from span_marker import SpanMarkerModel

model = SpanMarkerModel.from_pretrained("lxyuan/span-marker-bert-base-multilingual-uncased-multinerd")

# translate to chinese
description = "Singapore is renowned for its hawker centers offering dishes \
like Hainanese chicken rice and laksa, while Malaysia boasts dishes such as \
nasi lemak and rendang, reflecting its rich culinary heritage."

zh_description = "新加坡因其小贩中心提供海南鸡饭和叻沙等菜肴而闻名, 而马来西亚则拥有椰浆饭和仁当等菜肴，反映了其丰富的烹饪传统."

entities = model.predict(zh_description)

entities
>>>
[
  {'span': '新加坡', 'label': 'LOC', 'score': 0.8477746248245239, 'char_start_index': 0, 'char_end_index': 3},
  {'span': '马来西亚', 'label': 'LOC', 'score': 0.7525337934494019, 'char_start_index': 27, 'char_end_index': 31}
]

# It only managed to capture two countries: Singapore and Malaysia.
# All other entities were missed out.
# Same prediction as the [uncased model](https://huggingface.co/lxyuan/span-marker-bert-base-multilingual-cased-multinerd)

Quick Comparison on FOOD Entities

In this quick comparison, we found that a full stop punctuation mark seems to help the uncased model identify food entities, regardless of whether they are capitalized or in uppercase. In contrast, the cased model doesn't respond well to full stops, and adding them would lower the prediction score.

from span_marker import SpanMarkerModel

cased_model = SpanMarkerModel.from_pretrained("lxyuan/span-marker-bert-base-multilingual-cased-multinerd")
uncased_model = SpanMarkerModel.from_pretrained("lxyuan/span-marker-bert-base-multilingual-uncased-multinerd")

# no full stop mark
uncased_model.predict("i love fried chicken and korea bbq")
>>> []

uncased_model.predict("i love fried chicken and korea BBQ") # Uppercase BBQ only
>>> []

uncased_model.predict("i love fried chicken and Korea BBQ") # Capitalize korea and uppercase BBQ
>>> []

# add full stop to get better result
uncased_model.predict("i love fried chicken and korea bbq.")
>>> [
  {'span': 'fried chicken', 'label': 'FOOD', 'score': 0.6531468629837036, 'char_start_index': 7, 'char_end_index': 20},
  {'span': 'korea bbq', 'label': 'FOOD', 'score': 0.9738698601722717, 'char_start_index': 25,'char_end_index': 34}
]

uncased_model.predict("i love fried chicken and korea BBQ.")
>>> [
  {'span': 'fried chicken', 'label': 'FOOD', 'score': 0.6531468629837036, 'char_start_index': 7, 'char_end_index': 20},
  {'span': 'korea BBQ', 'label': 'FOOD', 'score': 0.9738698601722717, 'char_start_index': 25, 'char_end_index': 34}
]

uncased_model.predict("i love fried chicken and Korea BBQ.")
>>> [
  {'span': 'fried chicken', 'label': 'FOOD', 'score': 0.6531468629837036, 'char_start_index': 7, 'char_end_index': 20},
  {'span': 'Korea BBQ', 'label': 'FOOD', 'score': 0.9738698601722717, 'char_start_index': 25, 'char_end_index': 34}
]



# no full stop mark
cased_model.predict("i love fried chicken and korea bbq")
>>> [
  {'span': 'korea bbq', 'label': 'FOOD', 'score': 0.5054221749305725, 'char_start_index': 25, 'char_end_index': 34}
]

cased_model.predict("i love fried chicken and korea BBQ")
>>> [
  {'span': 'korea BBQ', 'label': 'FOOD', 'score': 0.6987857222557068, 'char_start_index': 25, 'char_end_index': 34}
]

cased_model.predict("i love fried chicken and Korea BBQ")
>>> [
  {'span': 'Korea BBQ', 'label': 'FOOD', 'score': 0.9755308032035828, 'char_start_index': 25, 'char_end_index': 34}
]

# add a fullstop mark hurt the cased model prediction score a little bit
cased_model.predict("i love fried chicken and korea bbq.")
>>> []

cased_model.predict("i love fried chicken and korea BBQ.")
>>> [
  {'span': 'korea BBQ', 'label': 'FOOD', 'score': 0.5078140497207642, 'char_start_index': 25, 'char_end_index': 34}
]

cased_model.predict("i love fried chicken and Korea BBQ.")
>>> [
  {'span': 'Korea BBQ', 'label': 'FOOD', 'score': 0.895089328289032, 'char_start_index': 25, 'char_end_index': 34}
]

Training procedure

One can reproduce the result running this script

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 3

Training results

Framework versions

• Transformers 4.30.2
• Pytorch 2.0.1+cu117
• Datasets 2.14.3
• Tokenizers 0.13.3