International Journal of Computer Processing of Languages

Patent classification is a large scale hierarchical text classification (LSHTC) task. Though comprehensive comparisons, either learning algorithms or feature selection strategies, have been fully made in the text categorization field, few work was done for a LSHTC task due to high computational cost and complicated structural label characteristics. For the first time, this paper compares two popular learning frameworks, namely, hierarchical support vector machine (SVM) and k-nearest neighbor (k-NN) that are applied to a LSHTC task. Our experimental results show that the latter outperforms the former for the LSHTC task, which is quite different from the existing results for normal text categorization tasks. In addition, this paper compares different similarity measures and ranking strategies in k-NN framework for LSHTC task. From our empirical study, conclusions can be drawn that k-NN is more appropriate for the LSHTC task than hierarchical SVM. BM25 outperforms other similarity measures and ListWeak gains a better performance than other ranking strategies. Our empirical results also indicate that using all the labels of the retrieved neighbors can remarkably improve classification performance over only using the first label of the retrieved neighbors.

The performance of a statistical machine translation (SMT) system depends heavily on the quantity and quality of the bilingual language resource. However, previous work mainly focuses on the quantity and tries to collect more bilingual data. In this paper, to optimize the bilingual corpus to improve the performance of the translation system, we propose some approaches to processing the training corpus by filtering noise and selecting more informative sentences from the training corpus. Also, to coordinate the parameter turning using minimum error rate training (MERT) approach, we propose two methods to select sentences from the large development data which are based on the phrase and sentence structure respectively. Different from the existing methods, our methods do not need so many development data but still obtains effective and robust parameters, while expending little time in the MERT process. The experimental results show that our methods can get better translation performance both in translation quality and speed.

In recent years, there has been an increasing interest in opinion analysis which seeks to identify and classify opinions in the text automatically. Many existing opinion analysis techniques use opinion lexicon as the most important features but lack the learning capacity of new opinion words and especially, the contextual behaviors of opinion words. Additionally, many works employ intra-sentence features while the contextual intersentence level feature is not well studied. This paper presents an opinion analysis technique which incorporates multiple level intra-sentence features from punctuation-, word- to collocation-level and inter-sentence feature. Based on the systematic observation and analysis on the opinion corpus, some linguistic clues and corresponding features for opinion analysis are discovered. These features are incorporated in a support vector machine (SVM) based classifier to identify opinionated sentences from running text and determine their polarities. To overcome the barrier caused by insufficient annotated training data, a semi-supervised learning algorithm is designed to train the classifier and enrich the opinion-related knowledge by using high quality instances identified from large raw text as new training data interactively. Evaluations on the datasets of NTCIR-6 opinion analysis task (OAT-6) and NTCIR-7 multilingual opinion analysis task (MOAT-7) show that the proposed opinion analysis approach achieved encouraging performance.

Recently, online opinions have received much attention, such as product reviews, economic analysis, political polls, etc. These opinions can provide people with valuable information and help them to make decisions. Therefore, how to effectively retrieve documents including opinions or comments specified by a target is becoming a popular research issue. A typical method for opinion retrieval is document-based method that splits each document into bag-of-word, and computes documental relevance and opinion separately. However, document-based methods neglect the contextual information and lead to the loss of the association between the opinion and its corresponding target which in turn affects the performance. In an extreme case, a relevant document full of irrelevant opinions will also be assigned a high ranking. In this paper, we will present a unified approach for opinion retrieval in the granularity of sentence because opinions would likely be expressed within one sentence. Thus, the association between an opinion and its corresponding target can be maintained and an overall score can be computed for each sentence. Moreover, we also tried to capture the contextual information of a sentence by its position. Compared with document-based approaches, experimental results on COSIB and COAE datasets showed that our approach achieved significant improvement.

With the rapid progress of storage technologies and communication networks, there has been an explosive growth of digital music information. How to organize and search large scale music collection is one of the most fundamentally challenging research problems. According to the musical form structure and repetition rule of musical theme, repeating patterns can be regarded as the representation of musical themes. Statistical analysis indicates that repeating patterns in music collection also follow Zipf's law. Therefore we introduce a dictionary based indexing approach to organize and index music collection, where the musical dictionary comprises repeating patterns. Then all music melodies are partitioned into words using forward or backward word segmentation method with the musical dictionary. Experimental results reveal that the proposed indexing method can be applied to music information retrieval effectively, and possesses obvious advantages when the scale of music collection increases drastically.

Events represent a crucial source of information in any text; dates and proper names (people, locations, organizations, products, diseases, etc.) combine to describe a certain event. In this paper, our focus is on these events, on how we find them in the text, extract them, organize them, save them in a database, and use them to answer any question about them. In our system, each event has three main types of identifying factors: keywords that indicate that an interesting event has occurred, dates that specify the time of the event, and proper nouns that specify the people, organizations, locations, products, etc., that are involved in the event. In this paper, we describe a system that uses these keys to find events, classify them, and save them in a database along with the identifier of the document that mentions that event. The retrieval process uses this information to provide the user with menus to form queries about the events; it then executes those queries, and finds the related documents.

This article focuses on the development of Natural Language Processing (NLP) tools for Computer Assisted Language Learning (CALL). First, we have developed some NLP tools: a labelled dictionary of Arabic (as complete as possible), a generator for morphological derivatives, a Conjugator and a morphological analyzer for Arabic. Second, we used these tools to create a number of educational applications for learning the Arabic language by using the proposed system SALA (an NLP-based authoring system, organized into three distinct layers: functions, scripts and activities).