Senta sentiment analysis system v1.0

Senta is an open source sentiment analysis system developed by Baidu.

Sentiment analysis aims to automatically identify and extract subjective information such as tendencies, positions, evaluations, and opinions in texts. It includes a variety of tasks, such as sentence-level emotion classification, evaluation object-level emotion classification, opinion extraction, emotion classification, etc. Sentiment analysis is an important research direction of artificial intelligence and has high academic value. At the same time, sentiment analysis has important applications in consumer decision-making, public opinion analysis, personalized recommendations and other fields, and has high commercial value.

Recently, Baidu officially released the emotional pre-training model SKEP (Sentiment Knowledge Enhanced Pre-training for Sentiment Analysis). SKEP uses emotional knowledge to enhance the pre-training model and surpasses SOTA in 14 typical Chinese and English sentiment analysis tasks. This work has been accepted by ACL 2020.

In order to facilitate R&D personnel and business partners to share leading-edge sentiment analysis technology, Baidu has open sourced the SKEP-based sentiment pre-training code and Chinese and English sentiment pre-training models in Senta. Moreover, in order to further lower the user threshold, Baidu has integrated a one-click sentiment analysis and prediction tool for industrialization into the SKEP open source project. Users only need a few lines of code to implement SKEP-based emotional pre-training and model prediction functions.

SKEP

SKEP is an emotional pre-training algorithm based on emotional knowledge enhancement proposed by the Baidu research team. This algorithm uses an unsupervised method to automatically mine emotional knowledge, and then uses the emotional knowledge to construct a pre-training target, so that the machine can learn to understand emotional semantics. SKEP provides a unified and powerful emotional semantic representation for various sentiment analysis tasks.

The Baidu research team performed three typical sentiment analysis tasks: Sentence-level Sentiment Classification, Aspect-level Sentiment Classification, and Opinion Role Labeling, with a total of 14 Chinese and English data. The above further verified the effect of the emotional pre-training model SKEP. Experiments show that using the general pre-training model ERNIE (internal) as initialization, SKEP improves by about 1.2% on average compared to ERNIE, and improves by about 2% on average compared to the original SOTA.