In order to improve the accuracy of ASA, new ASA has been using deep learning model to train dataset and return 3 posibilities (positive, negative and neutral).
We also chose to define cases in which we do not want the model to score a mention which will be unassigned.
Here are the unassigned cases:
1. mentions with only an urls inside
2. long mention with more than 500 words/punctuation/emojis
3. mentions only composed with hashtags or @
Below is a description of the sentiment types and a brief definition.
| Type | Note |
| Positive | The value of positive sentiment is significantly higher than others. |
| Negative | The value of negative sentiment is significantly higher than others. |
| Neutral | There are no sentiment words. Or, the negative value and the positive value are almost alike. |
| Unassigned | First, if the language is not supported by ASA, we can't assign sentiment. In this case, the mention will be classified as "unassigned". |
| Second, if a mention is crawled in a dashboard before being processed by NLP service, this mention can potentially be "unassigned". |
- Mentions with a single sentiment, or a combination of the same sentiment, will be assigned a positive or negative sentiment.
- Mention will be set to neutral sentiment when there is a neutral tone to the conversation. A combination of positive and negative sentiment in a mention will not lead to a neutral sentiment.
- Mentions with two or more different sentiments will be set to unassigned.
- Volume of mentions with all automatic sentiment and, more specifically, automatic neutral sentiment, might decrease, depending on mentions' content and type.
- A new complementary approached will be explored next to handle the complexity of providing accurate sentiments to mentions with mixed sentiments or with specific complexity, such as industries specific tones.
- Sarcasm and irony in mentions will be interpreted literally by the new model, leaving interpretation to the user. (to be confirmed by tests)