International Journal of Computer Processing of Languages

Recently, with the development of Chinese semantically annotated corpora, e.g. the Chinese Proposition Bank, the Chinese semantic role labeling (SRL) has been boosted. However, the Chinese SRL researchers now focus on the transplant of existing statistical machine learning methods which have been proven to be effective on English. In this paper, we have established a word based Chinese SRL system, which is quite different from the previous parsing based ones. We use "semantic chunking" to represent our new method. Semantic chunking is named because of its similarity with syntactic chunking. The difference is that semantic chunking is used to identify and classify the semantic chunks, i.e. the semantic roles, only with the word-level information. Based on semantic chunking, the process of SRL is changed from "parsing — semantic role identification — semantic role classification", to "semantic chunk identification — semantic chunk classification". With the elimination of the parsing stage, the SRL task can get rid of the dependency on parsing, which is the bottleneck both of speed and precision. The experiments have shown that the semantic chunking based method outperforms previously best-reported results on Chinese SRL, if the word segmentation and part-of-speech (POS) tagging are both correct. On the automatic word segmentation and POS tagging, our method decreases a little. The greatest advantage of semantic chunking method is that it saves a large amount of time. Besides these, we also carry out some experiments only for semantic role classification. These experiments have shown that only with the word-level features, the performance of semantic role classification can still be very high, which proves that the syntactic structural information is not indispensable.

Among the techniques to solve the knowledge bottleneck problem of supervised learning models, active learning is a promising method. One of the popular techniques of active learning is uncertainty sampling which, however, often presents problems when outliers are selected. To solve this problem, this paper presents a density-based re-ranking technique, in which a density measure is adopted to determine whether an unlabeled example is an outlier. The motivation of this method is to use not only the most informative example in terms of uncertainty measure, but also the most representative example in terms of density measure. The second effort we made is that a technique of sampling by clustering (SBC) is presented to build a representative initial training data set for active learning. Experimental results of active learning for word sense disambiguation and text classification tasks show that the proposed techniques can improve active learning with uncertainty sampling.

This paper reports on the development of a Named Entity Recognition (NER) system in Bengali by combining the outputs of the three classifiers, namely Maximum Entropy (ME), Conditional Random Field (CRF) and Support Vector Machine (SVM). A part of the Bengali news corpus developed from the web-archive of a leading Bengali newspaper has been manually annotated with the four major named entity (NE) tags, namely Person name, Location name, Organization name and Miscellaneous name. We have also used the annotated corpus of the IJCNLP-08 NER Shared Task for South and South East Asian Languages (NERSSEAL). An appropriate tag conversion routine has been developed in order to convert the fine-grained NE tagged NERSSEAL corpus to the form, tagged with the coarse-grained NE tagset of four tags. The system makes use of the different contextual information of the words along with the variety of orthographic word-level features that are helpful in predicting the four NE classes. In this work, we have considered language independent features as well as the language dependent features extracted from the various language specific resources. Lexical context patterns, which are generated from an unlabeled corpus of 10 million wordforms using an active learning technique, have been used for developing a baseline NER system as well as the features of the classifiers in order to improve their performance. A number of post-processing techniques have been used in order to improve the performance of the classifiers. Finally, the classifiers are combined together into a multiengine NER system using three weighted voting techniques. The system has been trained and tested with the datasets of 272K wordforms and 35K wordforms, respectively. Experimental results show the effectiveness of the proposed approach with the overall average Recall, Precision and F-Score values of 93.81%, 92.18% and 92.98%, respectively. The proposed system also outperforms the three other existing Bengali NER systems. The language independent versions of the ME, CRF and SVM based NER systems have been evaluated for the four other popular Indian languages, namely Hindi, Telugu, Oriya and Urdu, with the datasets obtained from the NERSSEAL shared task data. The SVM based system yielded the best performance with the F-Score values of 76.35%, 72.65%, 69.34% and 65.66% for Hindi, Telugu, Oriya and Urdu, respectively.

This paper proposes a method that automatically creates a sentiment lexicon in a new language using a sentiment lexicon in a resource–rich language with only a bilingual dictionary. We resolve some of the difficulties in selecting appropriate senses when translating lexicon, and present a framework that sequentially applies an iterative link analysis algorithm to enhance the quality of lexicons of both the source and target languages. The experimental results have empirically shown to improve the sentiment lexicon in the source language as well as create a good quality lexicon in the new language.

Automatic evaluation is known as a consistent and relatively inexpensive method of evaluating machine translation. Thus, it is useful in developing machine translation (MT) systems. Automatic MT evaluation methods basically require reference (human) translation. There are two types of automatic machine translation evaluation methods: one assesses machine translation by comparing it with reference translation, e.g., BLEU, NIST, METEOR, and the other method evaluates machine translation by classifying it as "good" (human translation-like) or "bad" (machine translation-like) translation based on properties of the translation. Due to the cost of the preparation of reference translation, the former method is more expensive. By contrast, the latter is inexpensive, because it requires reference translation only for training a classifier using machine learning algorithms. Previous studies of classification-based evaluation assessed the validity of their evaluation methods based on classification accuracy for overall machine translation. Since fluency (naturalness) of machine translation varies, classification results should reflect the fluency of translation. More fluent machine translation should involve less machine translation-like translation than poor machine translation. Therefore, this study assesses classification-based evaluation based on the percentage of MT-like translations in more fluent translation. This study further investigates the validity of classification-based evaluation by comparing classification-based evaluation with reference translation-based evaluation and manual evaluation results by human evaluators. The experimental results show that our classification-based method can accurately evaluate fluency of machine translation.

Parsing is an important process of Natural Language Processing (NLP) and Computational Linguistics which is used to understand the syntax and semantics of natural language sentences confined to the grammar. Parsing models need syntax and semantic coverage for better interpretation of natural language sentences. Though statistical parsing with trigram language models gives better performance through tri-gram probabilities and large vocabulary size, it has some disadvantages like lack of support in syntax, free ordering of words and long distance relationship which are the challenging features of the Tamil language. Grammar based structural parsing provides solutions to some extent. To overcome these disadvantages, structural component is to be involved in statistical approach which results in hybrid models like phrase and dependency models. To add the structural component, balance the vocabulary size and meet the challenging features, lexicalized and statistical parsing (LSP) is to be employed with the assistance of hybrid models. To incorporate all the features in complex and large sentences, phrase structure model may not be suitable to a larger extent. When dependency relations are applied among words, direct relationships can be established. Lexicalized and statistical parsing of natural language text in Tamil language using dependency model will give better performance than using phrase structure model. New part of speech (POS) and dependency tag sets for Tamil language have been Treebank has been developed with 326 sentences which comprises more than 5000 words with manual annotation. It has been extended to 1000 sentences using bootstrapping and manual correction and used to train the dependency model. This LSP with dependency model provides better results and covers all the features of Tamil language.