| Resumo: | In the context of Natural Language Processing, the Named Entity Recognition (NER) task focuses on extracting and classifying named entities from free text, such as news, which usually has a particular phrasing structure. Entity detection supports more complex tasks, such as Relation Extraction or Entity-Oriented Search, for instance the ANT search engine. There are some NER tools focused on the Portuguese language, such as Palavras or NERP-CRF, but their F-measure is still below the F-measure obtained by other available tools, for instance based on an annotated English corpus, trained with Stanford CoreNLP or with OpenNLP. ANT is an entity-oriented search engine for the University of Porto (UP). This search system indexes the information available in SIGARRA, the information system of the UP. Currently it uses handcrafted selectors to extract entities, based on XPath or CSS, which are dependent on the structure of the page. Furthermore, it does not work on free text, specially on SIGARRA's news. Using a machine learning method allows for the automation of the extraction task, making it scalable, structure independent and lowering the required work effort and consumed time. In this dissertation, I evaluate existing NER tools in order to select the best approach and configuration for the Portuguese language, particularly in the domain of SIGARRA's news. The evaluation was done based on two datasets, the HAREM collection, and a manually annotated subset of SIGARRA's news, which are used to assess the tools' performance using precision, recall and F-measure. Expanding the existing knowledge base will help index SIGARRA pages by providing a richer entity-oriented search experience with new information, as well as a better ranking scheme based on the additional context made available to the search engine. The scientific community also benefits from this work, with several detailed manuals resulting of the systematic analysis of available tools, in particular for the Portuguese language. First, I carried an out-of-the-box performance analysis of some selected tools (Stanford CoreNLP, OpenNLP, spaCy and NLTK) with the HAREM dataset, obtaining the best results for Stanford CoreNLP, followed by OpenNLP. Then, I performed a hyperparamenter study in order to select the best configuration for each tool, having achieved better-than-default results in each tool, particularly for NLTK's Maximum Entropy classifier, increasing the F-measure from 1.11% to 35.24%. Finally, using the best configuration, I repeated the training process with the SIGARRA News Corpus, having achieved F-measures as high as 86.64%, for Stanford CoreNLP. Furthermore, given this was also the out-of-the box winner, it leads me to conclude that Stanford CoreNLP is the best option for this particular context.
|
|---|