Project
ELSA: entity-level sentiment
A Java library that finds the people, organisations, places, and products in text and works out the sentiment expressed toward each — and who holds it — using a layered ONNX model pipeline.
Sentence-level sentiment is too coarse: "Chris hates Android phones but loves the iPhone camera" is neither positive nor negative. ELSA identifies each entity and the sentiment expressed toward it, including who holds the opinion.
Input: "Chris hates Android Phones but loves the iPhone camera."
Output: Android Phones [PRODUCT] NEGATIVE (0.92) holder: Chris
iPhone [PRODUCT] POSITIVE (0.88) holder: ChrisHow it is built
A four-layer pipeline in Java on ONNX Runtime, where each layer can stop early when there is nothing more to find: a subjectivity gate (TF-IDF and logistic regression, about 0.1 ms), a sentence segmenter (OpenNLP), a DistilBERT sentence-sentiment filter, and entity and sentiment-target models. The four ONNX models are trained with Python scripts in the repository (about an hour on Apple Silicon), fetched prebuilt, or replaced with synthetic stubs for CI.