Classes or also called labels, between which to discriminate (classify) the text, e.g.
Patterns, also known as label descriptors, allow the semantic matching between the analysed text and the different labels, e.g. This text is {}
Symanto Brain is based on zero- and few-shot technologies in accordance with the latest neural language models. Specifically, it uses semantic matching methods from a neural network with more than 300 million parameters and dozens of neural layers (deep neural network), trained with texts in more than 50 languages and fine-tuned for natural language inference, paraphrasing and other series of classification tasks (e.g., feeling, emotions, topic extraction, personality traits, etc.) in all types of data sources (e.g., social networks, reviews, news, etc.).
In order to speed up the text classification, which is one of the main drawbacks of current zero- and few-shot-based classification methods, Symanto has developed a proprietary technology based on Siamese networks and therefore, enabling researchers to build AI models quickly and procedurally.
To create any model using Symanto Brain, there are two key elements required:
Classes or also called labels, between which to discriminate (classify) the text, e.g.
Patterns, also known as label descriptors, allow the semantic matching between the analysed text and the different labels, e.g. This text is {}
Input: The product has no issues but the packaging causes so much extra to squirt out and you can't stop it. For how expensive it is it's such a waste.
Result:
With this, the zero-shot model allows rapid classification of the data set without the need for prior training and with results similar to those of a deep learning model trained with large amounts of data, whereas the few-shot model allows one to adjust the quality to a high extent by providing only a few annotated texts instead taking hours to annotate big datasets.
Specifically, Symanto Brain works as follows:
With zero-shot, the researcher only has to configure the task to approximate (defining patterns and labels) and expose the API for consumption. In this case, there is a single step, the use of the model based on requests to the exposed API.
With few-shot, the researcher retrains the model by providing it with some annotated data and exposes the API for consumption. In this case, there are two steps, 1) in model training; 2) the use of the exposed model.
The key to Symanto Brain, and where the research is focused, is the adequate definition of the task to be addressed, that is, the adequate definition of semantically appropriate classes and the patterns that allow the matching. |