datasets:
- tner/tweetner7
metrics:
- f1
- precision
- recall
model-index:
- name: tner/twitter-roberta-base-2019-90m-tweetner7-continuous
results:
- task:
name: Token Classification
type: token-classification
dataset:
name: tner/tweetner7
type: tner/tweetner7
args: tner/tweetner7
metrics:
- name: F1 (test_2021)
type: f1
value: 0.6587179789871326
- name: Precision (test_2021)
type: precision
value: 0.6727755003617073
- name: Recall (test_2021)
type: recall
value: 0.6452358926919519
- name: Macro F1 (test_2021)
type: f1_macro
value: 0.6107285696131857
- name: Macro Precision (test_2021)
type: precision_macro
value: 0.6215631908472189
- name: Macro Recall (test_2021)
type: recall_macro
value: 0.6039860329938679
- name: Entity Span F1 (test_2021)
type: f1_entity_span
value: 0.7843692816244613
- name: Entity Span Precision (test_2020)
type: precision_entity_span
value: 0.8010610079575596
- name: Entity Span Recall (test_2021)
type: recall_entity_span
value: 0.7683589684283566
- name: F1 (test_2020)
type: f1
value: 0.6475869809203142
- name: Precision (test_2020)
type: precision
value: 0.7049480757483201
- name: Recall (test_2020)
type: recall
value: 0.598858329008822
- name: Macro F1 (test_2020)
type: f1_macro
value: 0.6057800656625983
- name: Macro Precision (test_2020)
type: precision_macro
value: 0.6627892226359489
- name: Macro Recall (test_2020)
type: recall_macro
value: 0.5669673771050993
- name: Entity Span F1 (test_2020)
type: f1_entity_span
value: 0.755331088664422
- name: Entity Span Precision (test_2020)
type: precision_entity_span
value: 0.8222357971899816
- name: Entity Span Recall (test_2020)
type: recall_entity_span
value: 0.6984950700570836
pipeline_tag: token-classification
widget:
- text: >-
Get the all-analog Classic Vinyl Edition of `Takin' Off` Album from
{@herbiehancock@} via {@bluenoterecords@} link below: {{URL}}
example_title: NER Example 1
tner/twitter-roberta-base-2019-90m-tweetner7-continuous
This model is a fine-tuned version of tner/twitter-roberta-base-2019-90m-tweetner-2020 on the
tner/tweetner7 dataset (train_2021
split). The model is first fine-tuned on train_2020
, and then continuously fine-tuned on train_2021
.
Model fine-tuning is done via T-NER's hyper-parameter search (see the repository
for more detail). It achieves the following results on the test set of 2021:
- F1 (micro): 0.6587179789871326
- Precision (micro): 0.6727755003617073
- Recall (micro): 0.6452358926919519
- F1 (macro): 0.6107285696131857
- Precision (macro): 0.6215631908472189
- Recall (macro): 0.6039860329938679
The per-entity breakdown of the F1 score on the test set are below:
- corporation: 0.5165775401069518
- creative_work: 0.480106100795756
- event: 0.4846715328467153
- group: 0.6041666666666665
- location: 0.6836268754076973
- person: 0.8458527493010252
- product: 0.6600985221674878
For F1 scores, the confidence interval is obtained by bootstrap as below:
- F1 (micro):
- 90%: [0.6500084574752211, 0.6675327789934176]
- 95%: [0.6480876172354417, 0.6695072839398589]
- F1 (macro):
- 90%: [0.6500084574752211, 0.6675327789934176]
- 95%: [0.6480876172354417, 0.6695072839398589]
Full evaluation can be found at metric file of NER and metric file of entity span.
Usage
This model can be used through the tner library. Install the library via pip.
pip install tner
TweetNER7 pre-processed tweets where the account name and URLs are converted into special formats (see the dataset page for more detail), so we process tweets accordingly and then run the model prediction as below.
import re
from urlextract import URLExtract
from tner import TransformersNER
extractor = URLExtract()
def format_tweet(tweet):
# mask web urls
urls = extractor.find_urls(tweet)
for url in urls:
tweet = tweet.replace(url, "{{URL}}")
# format twitter account
tweet = re.sub(r"\b(\s*)(@[\S]+)\b", r'\1{\2@}', tweet)
return tweet
text = "Get the all-analog Classic Vinyl Edition of `Takin' Off` Album from @herbiehancock via @bluenoterecords link below: http://bluenote.lnk.to/AlbumOfTheWeek"
text_format = format_tweet(text)
model = TransformersNER("tner/twitter-roberta-base-2019-90m-tweetner7-continuous")
model.predict([text_format])
It can be used via transformers library but it is not recommended as CRF layer is not supported at the moment.
Training hyperparameters
The following hyperparameters were used during training:
- dataset: ['tner/tweetner7']
- dataset_split: train_2021
- dataset_name: None
- local_dataset: None
- model: tner/twitter-roberta-base-2019-90m-tweetner-2020
- crf: True
- max_length: 128
- epoch: 30
- batch_size: 32
- lr: 1e-05
- random_seed: 0
- gradient_accumulation_steps: 1
- weight_decay: 1e-07
- lr_warmup_step_ratio: 0.3
- max_grad_norm: 1
The full configuration can be found at fine-tuning parameter file.
Reference
If you use any resource from T-NER, please consider to cite our paper.
@inproceedings{ushio-camacho-collados-2021-ner,
title = "{T}-{NER}: An All-Round Python Library for Transformer-based Named Entity Recognition",
author = "Ushio, Asahi and
Camacho-Collados, Jose",
booktitle = "Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations",
month = apr,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.eacl-demos.7",
doi = "10.18653/v1/2021.eacl-demos.7",
pages = "53--62",
abstract = "Language model (LM) pretraining has led to consistent improvements in many NLP downstream tasks, including named entity recognition (NER). In this paper, we present T-NER (Transformer-based Named Entity Recognition), a Python library for NER LM finetuning. In addition to its practical utility, T-NER facilitates the study and investigation of the cross-domain and cross-lingual generalization ability of LMs finetuned on NER. Our library also provides a web app where users can get model predictions interactively for arbitrary text, which facilitates qualitative model evaluation for non-expert programmers. We show the potential of the library by compiling nine public NER datasets into a unified format and evaluating the cross-domain and cross- lingual performance across the datasets. The results from our initial experiments show that in-domain performance is generally competitive across datasets. However, cross-domain generalization is challenging even with a large pretrained LM, which has nevertheless capacity to learn domain-specific features if fine- tuned on a combined dataset. To facilitate future research, we also release all our LM checkpoints via the Hugging Face model hub.",
}