In this paper, we propose a new Korean Open Information Extraction system so-called SRDF. The SRDF system has been designed to effectively extract reified triples from Korean natural language texts based on the use of singleton property and other natural language processing techniques such as part-ofspeech tagging and chunking. The SRDF system is the Open Information Extraction system that enables extracting a multiple number of triples from a single sentence via reification.
Traditional Information Extraction (IE) thus far has been relying heavily on human
intervention of hand-crafted rules and hand-tagged training data. In recent years, on the
other hand, Open IE based on self-supervised learning has become more strongly
suggested to overcome such a limitation, and it is now possible to process massive text
corpora without having to require much human effort. TextRunner [
Throughout the following sections of this paper, we introduce SRDF, the new Korean
Open IE system, in much greater details. Since the Korean language, in a variety of
respects, has uniquely different grammatical structures and the system of postposition
and word spacing compared to other languages like English and Chinese in particular,
our team has been devoted to develop a new Open IE system specially designed to meet
the characteristics of Korean. We, at the same time, have also strived to build a system
through which multiple relationships between argument(s) and relation(s) within a
sentence can be extracted by using singleton property – the new method of reification [
SRDF simply receives as input a Korean text corpus and returns an extracted set of triples expressed in the form of singleton property. The system of SRDF operates through three steps of procedure in total that are “preprocessing”, “argument and relation detection”, and “triple generation” as described below. 2.1
When a Korean sentence is given as input, the SRDF system performs part-of-speech (POS) tagging and chunking first, as preprocessing.
The POS-tagged and chunked Korean sentence is then passed on to the next stage of argument and relation detection. This stage literally is to detect argument(s) and relation(s) from the given Korean sentence, and is further divided into three smaller steps similar to other Open IE systems based on self-supervised learning. ─ Labeling: At this stage, the preprocessed Korean sentence gets automatically labeled based on three important factors that are “the POS tag patterns”, “the position of words in sentence”, and “the postposition(s) within the sentence”. ─ Learning: An argument detection model and a relation detection model are learned here using decision tree. The former model uses “lemma”, “POS tag”, “length of the sentence”, “start position of argument”, “end position of argument”, “next lemma” and “next POS tag” as features, and the features that the latter uses include “lemma”, “POS tag” and “postposition”. ─ Extracting: Once a Korean sentence is received as input, the relation detection model in the SRDF system classifies whether a certain word in the given sentence is a relation or not, while the argument detection model classifies whether the word is a subject or an object. After that, they return all the classification results that are necessary for the next step of triple generation, including “the postposition(s) of detected argument(s)” and “the position of detected argument(s) and relation(s) in sentence”. 2.2
Our team has studied not only “how to extract information from Korean sentences” but
“how to generate triples for representing multiple relationships between argument(s)
and relation(s) within a sentence” as well. For example, the sentence ‘‘Barack
Obama was awarded the Nobel Prize in 2009.’’ is including multiple
relationships shared between one relation and two arguments, thus it is ideal that two
triples should be generated like <Barack Obama, award, the Nobel
Prize>, <Barack Obama, award, in 2009>. However, when we perform
information extraction on the above-mentioned sentence using the ReVerb [
As shown in Fig. 1 (blue = subject, orange = object, purple = relation), the method of generating triples using SRDF is as follows: 1. Identify the association between argument(s) and relation(s) based on the position of words in sentence. – Korean sentences have a different structure from English. Whereas an English sentence typically has a Subject-Relation-Object word order, the word order of Subject-Object-Relation is more common in Korean. In this light, the SRDF system can infer that the object(s) in Korean sentences are associated with the relation(s) located on the right side of them, and vice versa. In effect, the objects “the Nobel Prize” and “2009-year” are associated with the relation “award” on their right-hand side, as shown by the blue curved arrows in Fig. 1. 2. Identify whether the postposition(s) attached to the object(s) of the sentence is accusative. – In Korean, it is also common that almost every object is attached with a postposition, and the postposition is considered a very important factor when understanding syntax of the sentence. Among various postpositions, accusative postposition “EUL” specifically indicate that the relation of the sentence is a transitive verb. When the postposition attached to the object of the sentence is accusative, the SRDF system generates a triple with the following form in general <Barack Obama, award#1, the Nobel Prize>, in which the relation “award” is attached with a provenance #1. In other cases where the relation of the sentence is not a transitive verb and the postposition attached to the object is not accusative, triples are made by the SRDF system in the anonymous form of <subject, relation#provenance, ANOMYOUMS>. This method has an advantage of enabling representation of sentences with no object in the form of triple. 3. Generate reified triples using remained objects. – The main triple <Barack Obama, award#1, the Nobel Prize> has been made in the previous step and, at this stage, “2009-year-E” should be reified. When generating a reified triple, the SRDF system situates the relation of the main triple as the subject of the reified triple, places the postposition as the relation, and lets the object be the object as <award#1, E#1, 2009-year> for instance.
The performance of SRDF system has been evaluated by application to 100 Korean sentences randomly sampled from the web as a testing data set. The evaluation results have been assessed by two human evaluators based on the two criteria of Detection – how precisely the SRDF system has detected the argument(s) and relation(s) from the given sentence – and Triple Generation – how accurately the reified-triple has been generated from the detected argument(s) and relation(s) –. The results and error statistics are presented in Table 1 below. As shown in Table 1, the SRDF system is of an excellent capability of both detecting argument(s) and relation(s) and generating triples, where the performance of triple generation is relatively 18% lower. Having thoroughly examined the failed sentences, we found out that most errors occur in the course of detection followed by POS-tagging, and the least errors are made during the process of reification.
4
In this paper, we have demonstrated the feasibility of extracting structured information from Korean natural language texts without any human intervention. We have also proposed a novel method of combining Open IE with the singleton property technique in representation of multiple relationships between argument(s) and relation(s) within a sentence. Our project is still ongoing in active progress, and it is with great expectation for our forthcoming researches to more technically expand the scope of our project. All the expected accomplishments of the next phases of our project work will be made publicly available through the website at http://143.248.135.216:8080/SRDFREST/index.htm. Acknowledgement. This work was supported by Institute for Information & communications Technology Promotion(IITP) grant funded by the Korea government(MSIP) (No. R0101-150054, WiseKB: Big data based self-evolving knowledge base and reasoning platform)