You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
senpy/docs/conversion.rst

178 lines
5.5 KiB
ReStructuredText

Automatic Model Conversion
--------------------------
Senpy includes support for emotion and sentiment conversion.
When a user requests a specific model, senpy will choose a strategy to convert the model that the service usually outputs and the model requested by the user.
Out of the box, senpy can convert from the `emotionml:pad` (pleasure-arousal-dominance) dimensional model to `emoml:big6` (Ekman's big-6) categories, and vice versa.
This specific conversion uses a series of dimensional centroids (`emotionml:pad`) for each emotion category (`emotionml:big6`).
A dimensional value is converted to a category by looking for the nearest centroid.
The centroids are calculated according to this article:
.. code-block:: text
Kim, S. M., Valitutti, A., & Calvo, R. A. (2010, June).
Evaluation of unsupervised emotion models to textual affect recognition.
In Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text (pp. 62-70).
Association for Computational Linguistics.
It is possible to add new conversion strategies by `Developing a conversion plugin`_.
Use
===
Consider the following query to an emotion service: http://senpy.gsi.upm.es/api/emotion-anew?i=good
The requested plugin (emotion-random) returns emotions using the VAD space (FSRE dimensions in EmotionML):
.. code:: json
[
{
"@type": "EmotionSet",
"onyx:hasEmotion": [
{
"@type": "Emotion",
"emoml:pad-dimensions_arousal": 5.43,
"emoml:pad-dimensions_dominance": 6.41,
"emoml:pad-dimensions_pleasure": 7.47,
"prov:wasGeneratedBy": "prefix:Analysis_1562744784.8789825"
}
],
"prov:wasGeneratedBy": "prefix:Analysis_1562744784.8789825"
}
]
To get the equivalent of these emotions in Ekman's categories (i.e., Ekman's Big 6 in EmotionML), we'd do this:
http://senpy.gsi.upm.es/api/emotion-anew?i=good&emotion-model=emoml:big6
This call, provided there is a valid conversion plugin from Ekman's to VAD, would return something like this:
.. code:: json
[
{
"@type": "EmotionSet",
"onyx:hasEmotion": [
{
"@type": "Emotion",
"onyx:algorithmConfidence": 4.4979,
"onyx:hasEmotionCategory": "emoml:big6happiness"
}
],
"prov:wasDerivedFrom": {
"@id": "Emotions0",
"@type": "EmotionSet",
"onyx:hasEmotion": [
{
"@id": "Emotion0",
"@type": "Emotion",
"emoml:pad-dimensions_arousal": 5.43,
"emoml:pad-dimensions_dominance": 6.41,
"emoml:pad-dimensions_pleasure": 7.47,
"prov:wasGeneratedBy": "prefix:Analysis_1562745220.1553965"
}
],
"prov:wasGeneratedBy": "prefix:Analysis_1562745220.1553965"
},
"prov:wasGeneratedBy": "prefix:Analysis_1562745220.1570725"
}
]
That is called a *full* response, as it simply adds the converted emotion alongside.
It is also possible to get the original emotion nested within the new converted emotion, using the `conversion=nested` parameter:
http://senpy.gsi.upm.es/api/emotion-anew?i=good&emotion-model=emoml:big6&conversion=nested
.. code:: json
[
{
"@type": "EmotionSet",
"onyx:hasEmotion": [
{
"@type": "Emotion",
"onyx:algorithmConfidence": 4.4979,
"onyx:hasEmotionCategory": "emoml:big6happiness"
}
],
"prov:wasDerivedFrom": {
"@id": "Emotions0",
"@type": "EmotionSet",
"onyx:hasEmotion": [
{
"@id": "Emotion0",
"@type": "Emotion",
"emoml:pad-dimensions_arousal": 5.43,
"emoml:pad-dimensions_dominance": 6.41,
"emoml:pad-dimensions_pleasure": 7.47,
"prov:wasGeneratedBy": "prefix:Analysis_1562744962.896306"
}
],
"prov:wasGeneratedBy": "prefix:Analysis_1562744962.896306"
},
"prov:wasGeneratedBy": "prefix:Analysis_1562744962.8978968"
}
]
Lastly, `conversion=filtered` would only return the converted emotions.
.. code:: json
[
{
"@type": "EmotionSet",
"onyx:hasEmotion": [
{
"@type": "Emotion",
"onyx:algorithmConfidence": 4.4979,
"onyx:hasEmotionCategory": "emoml:big6happiness"
}
],
"prov:wasGeneratedBy": "prefix:Analysis_1562744925.7322266"
}
]
Developing a conversion plugin
==============================
Conversion plugins are discovered by the server just like any other plugin.
The difference is the slightly different API, and the need to specify the `source` and `target` of the conversion.
For instance, an emotion conversion plugin needs the following:
.. code:: yaml
---
onyx:doesConversion:
- onyx:conversionFrom: emoml:big6
onyx:conversionTo: emoml:fsre-dimensions
- onyx:conversionFrom: emoml:fsre-dimensions
onyx:conversionTo: emoml:big6
.. code:: python
class MyConversion(EmotionConversionPlugin):
def convert(self, emotionSet, fromModel, toModel, params):
pass
More implementation details are shown in the `centroids plugin <https://github.com/gsi-upm/senpy/blob/master/senpy/plugins/postprocessing/emotion/centroids.py>`_.