Update README.md

README.md

@@ -25,4 +25,49 @@ model-index:
           - name: Relative polarity precision
             type: Relative polarity precision
             value: 93.19%
----
+---
+
+
+
+This repository contains a pretrained model (and an easy-to-run wrapper for it) for structured sentiment analysis in Norwegian language, pre-trained on the [NoReC dataset](https://huggingface.co/datasets/norec).
+This is an implementation of the method described in "Direct parsing to sentiment graphs" (Samuel _et al._, ACL 2022). The main repository that also contains the scripts for training the model, can be found on the project [github](https://github.com/jerbarnes/direct_parsing_to_sent_graph).
+The model is also available in the form of a [HF space](https://huggingface.co/spaces/ltg/ssa-perin).
+
+
+  The sentiment graph model is based on an underlying masked language model – [NorBERT 2](https://huggingface.co/ltg/norbert2).
+The proposed method suggests three different ways to encode the sentiment graph: "node-centric", "labeled-edge", and "opinion-tuple".
+The current model 
+- uses "labeled-edge" graph encoding
+- does not use character-level embedding
+- all other hyperparameters are set to [default values](https://github.com/jerbarnes/direct_parsing_to_sent_graph/blob/main/perin/config/edge_norec.yaml)
+, and it achieves the following results on the held-out set of the NoReC dataset:
+
+| Unlabeled sentiment tuple F1 | Target F1  | Relative polarity precision |
+|:----------------------------:|:----------:|:---------------------------:|
+|     0.434                    |  0.541      |        0.926                |
+
+
+In "Word Substitution with Masked Language Models as Data Augmentation for Sentiment Analysis", we analyzed data augmentation strategies for improving performance of the model. Using masked-language modeling (MLM), we augmented the sentences with MLM-substituted words inside, outside, or inside+outside the actual sentiment tuples. The results below show that augmentation may be improve the model performance. This space, however, runs the original model trained without augmentation.
+
+|                | Augmentation rate | Unlabeled sentiment tuple F1 | Target F1 | Relative polarity precision |
+|----------------|-------------------|------------------------------|-----------|-----------------------------|
+| Baseline       | 0%               | 43.39                        | 54.13     | 92.59                       |
+| Outside        | 59%              | **45.08**                    | 56.18     | 92.95                       |
+| Inside         | 9%               | 43.38                        | 55.62     | 92.49                       |
+| Inside+Outside | 27%              | 44.12                        | **56.44** | **93.19**               |
+
+
+
+The model can be easily used for predicting sentiment tuples as follows: 
+
+```python
+>>> import model_wrapper
+>>> model = model_wrapper.PredictionModel()
+>>> model.predict(['vi liker svart kaffe'])
+[{'sent_id': '0',
+  'text': 'vi liker svart kaffe',
+  'opinions': [{'Source': [['vi'], ['0:2']],
+    'Target': [['svart', 'kaffe'], ['9:14', '15:20']],
+    'Polar_expression': [['liker'], ['3:8']],
+    'Polarity': 'Positive'}]}]
+```