Ensemble Technique Utilization for Indonesian Dependency Parser

31st Pacific Asia Conference on Language, Information and Computation (PACLIC 31), pages 64–71 Cebu City, Philippines, November 16-18, 2017 Copyright c⃝2017 Arief Rahman, Kuncoro Adhiguna and Ayu Purwarianti Ensemble Technique Utilization for Indonesian Dependency Parser Arief Rahman Ayu Purwarianti Institut Teknologi Bandung Indonesia Institut Teknologi Bandung Indonesia 23516008@std.stei.itb.ac.id ayu@stei.itb.ac.id Abstract Two of the main problems in creating an Indonesian parser with high accuracy are the lack of sentence diversity in treebank used for training and suboptimal uses of parsing techniques. To resolve these problems, we build an Indonesian dependency treebank of 2098 sentences (simple and complex sentences) and use ensemble techniques to maximize the usage of available dependency parsers. We compare the combination of seven parsing algorithms provided by MaltParser and MSTParser, which provides both transition-based and graph-based models. From our experiments, we found that the graph-based model performs better than the transition-based model for Indonesian sentences. We also found no significant accuracy difference in models between several simple ensemble models and reparsing algorithms. 1 Introduction Text parsing is one of the major tasks in natural language text processing (NLP). Text parsing is the process of determining the syntactic structure of a sentence. The result of text parsing is a syntactical tree, which is mostly used for higher-level NLP tasks, like sentiment analysis (Di Caro and Grella, 2013) and semantic role labeling (Johansson and Nugues. 2008). There are two kinds of text parsing to date: constituent parsing and dependency parsing. Constituent parsing parses a sentence by determining the constituent phrases of the sentence hierarchically, usually by using a grammar (Aho, 2003). Dependency parsing, on the other hand, parses a sentence by determining a dependency relation for each word in a sentence. In this research, we use dependency parsing, because it is suited for analyzing languages with free word order, such as Indonesian (Nivre, 2007). Figure 1 shows an example of a parsed Indonesian sentence using dependency structure. Figure 1. Example of a parsed Indonesian sentence (TL: That allegation does not miss) with dependency structure Up until now, there have been only a few studies regarding Indonesian dependency parsing (Sulaeman, 2012; Green et.al, 2012). Most of the previous researches focused on rule-based parsing (Purwarianti et.al, 2013), which yielded quite a low accuracy, compared to other languages. Based on these researches, we use ensemble parsing techniques (Surdeanu and Manning, 2010) in our works. We also built a dependency Treebank corpus used for the model training with 2098 sentences.� In the following sections, we describe the relevant studies and some basic concepts about 64 dependency parsing and its models. We then describe the corpus used in this research, our experiment settings, and finally the results and analysis.� 2 Related Works There are two studies that are related to ensemble dependency parsing, which is Surdeanu & Manning's work for English (Attardi and Dell'Orletta, 2009), and Green et al.'s work for Indonesian (Green et.al, 2012). Surdeanu & Manning created an ensemble dependency parser using parsing algorithms from both MaltParser and MSTParser for English. This research …

Content truncated to 500 words