Abstract: We present a model for automatic extraction of metadata and keywords to be used in the classification of scientific documents. The model mainly consists of metadata extraction, keywords extraction and documents classification. At the metadata extraction stage, various metadata items are extracted from research documents in the domain of commerce such title of the thesis/research article, author/s, advisor/s, year, publisher, type, and abstract. At the keywords extraction stage, Latent Semantic Indexing (LSI) is used to extract the underlying topics from these documents. At the classification stage which depends on the metadata and keywords extraction stages, three classification algorithms are used which are Stochastic Gradient Descent (SGD), Linear Support Vector (LSVC) and K-Nearest Neighbor (KNN). SGD has achieved the highest classification accuracy (80.5%) compared to LSVC and KNN when applied to Arabic document corpus. LSVC has achieved the highest classification accuracy (81.5%) compared to SGD and KNN when applied to the English document corpus.

The figure below shows an abstract view of the basic machine learning model that is used to extract metadata and keyword and also classifiying the given research document/s according to the predefined library of congress classification scheme. Based on the domain specific research dataset prepaired on the first stage of the model, the model is trained and tested to extract metadata and keywords as well as to calssify research documents according the given classificatoin scheme. Metadata include research title, author/s, year and month, publisher, etc. Keywords include research topic-related terms and words and some phrases as needed. These output data items are fedback to the model so it can use it in future extractions and classifications.

The figure below shows the architecture of the cloud-transformed realization of the basic machine learning model shown above. After the basic model is trained and tested, it is serialized together with the research dataset (train and test) to Google Cloud Storage as pucket storage and to Google App Engine as implemneted model in Python toghether with needed Machine learning functions and data processing function. The model running in Google App Engine requests and loads research dataset/s from the Google Cloud Storage as needed. The implemented model (extraction tool now) has a RESTful interface where a user can interact with to submit and get extraction and classification tasks.