This paper presents the results of the two shared tasks associated with W-NUT 2015: (1) a text normalization task with 10 participants; and (2) a named entity tagging task with 8 participants. We outline the task, annotation process and dataset statistics, and provide a high-level overview of the participating systems for each shared task.

Models for statistical spoken language understanding (SLU) systems are conventionally trained using supervised discriminative training methods. In many cases, however, labeled data necessary for these supervised techniques is not readily available necessitating a laborious data collection and annotation effort. This often results into data sets that are not expansive enough to cover adequately all patterns of natural language phrases that occur in the target applications. Word embedding features alleviate data and feature sparsity issues by learning mathematical representation of words and word associations in the continuous space. In this work, we present techniques to obtain task and domain specific word embeddings and show their usefulness over those obtained from generic unsupervised data. We also show how we transfer these embeddings from one language to another enabling training of a multilingual spoken language understanding system. Index Terms-spoken language understanding; natural language processing; word embedding; named entity recognition; vector space models

Abstract. In this paper, we consider the problem of sentence difficulty analysis from various angles. Past works have endeavored to design deterministic scoring algorithms depending only on semantic and syntactic information. We propose instead not only to hire local feature space representing individual sentence with its syntactic and semantic structure, but also to consider global distributional difference among corpora. For the local feature space, we select 28 linguistic features and transform them into conjuncted and discretized form. By applying global score classification, we can show its much improved results. We test our proposed model to 1,000 sentences and get much higher accuracy than traditional learning models such as SVM and AdaBoost.

Abstract. Supersense tagging is a problem of finding a corresponding semantic super tag (eg. Phenomenon, Act) based on syntactic information and annotated corpora. However, we employ semantic information rather than syntactic one and annotated corpora, because Korean language has relatively flexible syntactic structure and is lack of annotated corpora. To construct the automatic sense tagging system for Korean language, we use semi-supersenses of first and second level in Sejong's Noun Semantic Class System. We employ a hybrid approach consisting of three phases: one morphological matching phase and two semantic matching phases. The morphological phase is based on suffix pattern matching which assigns compound word to the class including the suffix word. One of the two semantic matching phases is based on concept similarity on WordNet, and the other is based on the term similarity in term matrix reduced by singular value decomposition (SVD). Above semantic phases are using weighted k-Nearest Neighbor classifier commonly but are also using different similarity metrics. In experiments, 79,103 unknown words are extracted from 225,779 noun words from syntactic tagged corpus, and 98% of the unknown words are addressed by our hybrid method.