What is GALIOT AI?
GALIOT AI is aviation safety intelligence framework based on Natural Language Processing (NLP) and Machine Learning (ML) computer techniques.
Okay, but what exactly GALIOT AI can do for me?
1. Identity priority safety reports and find the similar historical occurrences
GALIOT AI finds similar historical occurrences based on contextual relations between sentences and words in reports narrative and documents metadata. Calculated similarity between safety reports is far more accurate than simple SQL query results.
2. Predict classifications for new received safety report
GALIOT AI can predict different report classifications like risk level, root cause(s), occurrence category, …
All predictions and estimated prediction accuracy are calculated from pretrained models fine tuned on historical report data.
3. Discover hidden patterns
Based on unsupervised machine learning technique, GALIOT AI can discover probabilistic topic model(s) in hidden semantic structures that occur in collection of occurrences descriptions
and audit findings.
4. Answer questions
Question-answering (QA) is computer science discipline within the field of information retrieval. GALIOT AI is capable of pulling answers (and related context) from unstructured
collections of safety documents and manuals based on questions posed by users in natural language format.
How GALIOT AI performs safety intelligence?
GALIOT AI combines different Data Science, Machine Learning and Natural Language Processing methods and advanced programming techniques to analyse complex safety and quality data correlation
in order to infer knowledge about hidden patterns, emerging hazards, occurrence risk classification/categorisation and prioritisation of safety issues.
GALIOT AI processes all occurrence metadata and narratives using NLP methods and transforms words, phrases and entire sentences into multidimensional vectors.
Multiple transformation models: Bidirectional Encoder Representations (Google BERT), Word Embedding (Word2Vec), and Bag-of-Words (Tf-Idf) are trained and evaluated on a
custom history data set provided. After the transformation, different methods on vectors embedings are be performed
(cosine similarity, k-means clustering, Latent Dirichlet Allocation, Naive Bayes classifier, …) depending on task goals listed above.