A Hybrid Machine Learning Approach for Information Extraction

Abstract

Information Extraction (IE)

Automatic Information Extraction in Semi-structured Official Journals

“…A promissing ML approach applied by different authors is to use learning algorithms as text classifiers for IE [11,12,13,14,15]. In this approach, the input document is initially broken into fragments which are candidates to fill in the output fields.…”

“…The extraction is then accomplished by considering the classifications provided by the ML algorithm. The text classification approach has been applied, for instance, to extract information on business cards [11], bibliographic references [12,13,15], author affiliations [13], job advertisements [14], among other applications.…”

“…It has advantages compared to other ML approaches since it may use conventional and easily available algorithms, which is not the case with most IE systems based on automata induction and pattern matching techniques, that use complex and specific learning algorithms [15]. Among the algorithms used as text classifiers, we mention, for instance, bayesian classifiers [13], decision trees [15], Hidden Markov Models [11] and Conditional Random Fields [12].…”

“…In some IE tasks being tackled in previous work, it is necessary to deploy different delimiters in order to generate text fragments which are good candidates to be associated to the required data fields (e.g., [4,11,15]). In our domain of application, however, the data fields to be extracted always correspond to one or more complete lines of text, and hence, the delimiter used to fragment the input documents was solely the end-of-line character.…”

Hidden Markov Models and Text Classifiers for Information Extraction on Semi-Structured Texts

“…In [23], the authors provided the initial experiments which evaluated the viability of the proposal. In the current work, we present more discussion and new experiments with the proposed approach, including the use of Support Vector Machines in the first phase of classification and a comparison to a benchmarking IE approach successfully used in previous work [4].…”