---
language:
- 'no'
- nb
- nn
license: cc-by-4.0
pipeline_tag: token-classification
---
# Targeted Sentiment Analysis model for Norwegian text
This model is a fine-tuned version of [ltg/norbert3-large](https://huggingface.co/ltg/norbert3-large) For Targeted Sentiment Analysis (TSA) on Norwegian text. The fine-tuning script is avaiable [on github](https://github.com/egilron/seq-label.git).  
In TSA, we identify sentiment targets, "That what is spoken positively or negatively about" in each sentence. Our models performs the task through sequence labeling, AKA "token classification". 

The dataset used for fine-tuning is [ltg/norec_tsa](https://huggingface.co/datasets/ltg/norec_tsa), at its defaul settings, were sentiment targets are labeled as either "targ-Positive" or "targ-Negative". The norec_tsa dataset is derived from the [NoReC_fine dataset](https://github.com/ltgoslo/norec_fine).


## Quick start
You can use this model in your scripts as follows:
```>>> from transformers import pipeline
>>> origin = "ltg/norbert3-large_TSA"
>>> trust_remote = "norbert3" in origin.lower()
>>> text = "Hans hese , litt såre stemme kler bluesen , men denne platen kommer neppe til å bli blant hans største kommersielle suksesser ."
>>> if trust_remote: # Downloads configurations for norbert3
...     pipe = transformers.pipeline( "token-classification",
...                                 aggregation_strategy='first',
...                                 model = origin,
...                                 trust_remote_code=trust_remote,
...                                 tokenizer = AutoTokenizer.from_pretrained(origin)
...                 )
...     preds = pipe(text)
...     for p in preds:
...         print(p)

{'entity_group': 'targ-Positive', 'score': 0.6990814, 'word': ' Hans hese , litt såre stemme', 'start': 0, 'end': 28}
{'entity_group': 'targ-Negative', 'score': 0.5721016, 'word': ' platen', 'start': 53, 'end': 60}
```  



## Training hyperparameters
- per_device_train_batch_size: 64
- per_device_eval_batch_size: 8
- learning_rate: 1e-05
- gradient_accumulation_steps: 1
- num_train_epochs: 24 (best epoch 18)
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08

## Evaluation
```               precision    recall  f1-score   support

targ-Negative     0.4648    0.3143    0.3750       210
targ-Positive     0.5097    0.6019    0.5520       525

    micro avg     0.5013    0.5197    0.5104       735
    macro avg     0.4872    0.4581    0.4635       735
 weighted avg     0.4969    0.5197    0.5014       735
 ```