Semi-Supervised Noun Compound Analysis with Edge and Span Features

Proceedings of COLING 2012: Technical Papers, pages 1915–1932, COLING 2012, Mumbai, December 2012. Semi-supervised Noun Compound Analysis with Edge and Span Features Yugo MURAWAKI Sadao KUROHASHI Graduate School of Informatics, Kyoto University Yoshida-honmachi, Sakyo-ku, Kyoto, 606-8501, Japan {murawaki, kuro}@i.kyoto-u.ac.jp ABSTRACT In this paper, we propose the use of spans in addition to edges in noun compound analysis. A span is a sequence of words that can represent a noun compound. Compared with edges, spans have good properties in terms of semi-supervised parsing. They can be reliably extracted from a huge amount of unannotated text. In addition, while the combinations of edges such as sibling and grandparent interactions are, in general, difficult to handle in parsing, it is quite easy to utilize spans with arbitrary width. We show that spans can be incorporated straightforwardly into the standard chart-based parsing algorithm. We create a semi-supervised discriminative parser that combines edge and span features. Experiments show that span features improve accuracy and that further gain is obtained when they are combined with edge features. TITLE AND ABSTRACT IN JAPANESE スパンとエッジ特徴量を用いた 半教師あり名詞句解析 名詞句解析において，エッジだけでなくスパンを手がかりとして使うことを提案する． スパンは名詞句を表しうる単語列であり，エッジと比べて半教師あり学習に適した性質 を持っている．すなわち，大量の生テキストから高い信頼性をもって抽出可能である． さらに，エッジは解析時に組み合わせ (兄弟や孫の関係など) を考えることが一般に難 しいのに対して，スパンは任意の長さの組み合わせを自明に利用できる．この論文で は，スパンが動的計画法による標準的な構文解析手法に簡単に組み込めることを示し， エッジとスパン特徴量を組み合わせた半教師ありの識別型構文解析器を提案する．実験 により，スパン特徴量が解析精度を改善し，エッジと組み合わせることでさらに精度が 向上することが示された． KEYWORDS: noun compound analysis, semi-supervised learning, parsing, span. KEYWORDS IN JAPANESE: 名詞句解析, 半教師あり学習, 構文解析, スパン. 1915 1 Introduction Words are used as a basic unit in a broad range of applications in natural language processing. However, it often happens that what we need to recognize turn out to be longer than single words. They are phrases, noun compounds in particular, that consist of more than one word. A noun compound is not just a sequence of words but has a latent structure. Consider the following example in Japanese. [jidou automation [onsei speech ninshiki]] recognition automatic speech recognition The brackets indicate the internal structure of the noun compound. In addition to the right- branching structure, it has another possible interpretation. [[jidou automation onsei] speech ninshiki] recognition recognition of automatic speech Our goal is to recognize that the former is semantically coherent while the latter is not. In order to analyze the internal structures of noun compounds, we need some automatic method because they are too large in number and too productive to be covered by a hand-crafted lexicon. This task is called noun compound analysis. Noun compound analysis can be seen as a task of dependency parsing (Lauer, 1995). However, it is different from usual full-sentence parsing in that part-of-speech tags help little, if at all. A noun compound is just a sequence of nouns and lacks grammatical markers. For this reason we take fully lexicalized approaches in noun compound analysis. Fully lexicalized approaches often suffer from the data sparseness problem. Apart from the observation that nouns have much higher domain specificity than other words, a model needs to learn the lexical association of a pair of words. However, we can never create an annotated corpus that covers the combinations of tens of thousands of words. To overcome data sparseness, it seems promising to take semi-supervised approaches that exploit a huge amount …

Content truncated to 500 words