Guiping Zhang

Research on Katakana phrase translation based on bi-directional integration

In order to solve the problem of Katakana reduced to English in Japanese-English translation, we employ the phrase-based statistical machine translation model to perform Katakana phrase (or word) translation from Japanese to English. The katakana phrase is segmented into words by CRF, and then Japanese-English and English-Japanese bi-directional integration translation is carried out on those segmented results. The translated results of all the segmented words are comprehensively scored to obtain the best English phrase translation result. The experimental results indicate that the katakana phrase translation precision reaches 76%, effectively addresses the problem of the katakana reduced to English.

Using Bilingual Segments to Improve Interactive Machine Translation

Recent research on machine translation has achieved substantial progress. However, the machine translation results are still not error-free, and need to be post-edited by a human translator (user) to produce correct translations. Interactive machine translation enhanced the human-computer collaboration through having human validate the longest correct prefix in the suggested translation. In this paper, we refine the interactivity protocol to provide more natural collaboration. Users are allowed to validate bilingual segments, which give more direct guidance to the decoder and more hints to the users. Besides, validating bilingual segments is easier than identifying correct segments from the incorrect translations. Experimental results with real users show that the new protocol improved the translation efficiency and translation quality on three Chinese-English translation tasks.

Automatic Construction of Domain Terminology Knowledge Base for HowNet Based on the Headword

HowNet is a Chinese-English Bilingual common-sense knowledge base, playing an important role in machine translation tasks. However, when facing domain-specific machine translation tasks, HowNet must be supplemented with domain-specific terminologies. In other words, we need to construct domain terminology semantic knowledge base. In this paper, we propose a method to automatically construct domain terminology knowledge base, based on the headword of a terminology. Specifically, the semantic meaning (HowNet DEF) of an unseen terminology is defined as one of the semantic meanings of the headword of the terminology. Headword disambiguation is done by considering the context of headwords and adding domain-specific disambiguation rules to the general disambiguation rules. Experiments on aviation domain show that our proposed method on headword disambiguation achieves 9.4% improvement based on the default disambiguation tools in HowNet. We also find that with our automatically constructed domain terminology knowledge base, HowNet machine translation system achieves better translation quality.

Improving interactive machine translation via multiple positive constraints

Since machine translation systems are still unable to produce satisfactory outputs, recently various interactive machine translation (IMT) approaches are proposed. State-of-the-art IMT systems use the human validated prefix as the only constraint that guides decoding, in which the human guidance is quite insufficient. This paper extends the human-computer interactions by allowing translators to provide multiple correct fragments (CFs) besides prefix. These fragments act as positive constraints for decoding. Four improvements are proposed to adapt traditional IMT to this new interaction mode. Experimental results on Chinese-English corpora show that our method achieves lower KSMT scores under comparable prediction speed in comparison with the traditional IMT method.

Incorporating Word Clustering into Complex Noun Phrase Identification

Since the professional technical literature include amounts of complex noun phrases, identifying those phrases has an important practical value for such tasks as machine translation. Through analysis of those phrases in Chinese-English bilingual sentence pairs from the aircraft technical publications, we present an annotation specification based on the existing specification to label those phrases and a method for the complex noun phrase identification. In addition to the basic features including the word and the part-of-speech, we incorporate the word clustering features trained by Brown clustering model and Word Vector Class (WVC) model on a large unlabeled data into the machine learning model. Experimental results indicate that the combination of different word clustering features and basic features can leverage system performance, and improve the F-score by 1.83 % in contrast with the method only adding the basic features.

Study on assistant concept acquisition in domain ontology construction for Chinese texts

Concept acquisition is an important part of domain ontology construction, and how to accomplish assistant concept acquisition becomes a research focus. In this paper, a character-based CRF model is adopted to obtain the set of candidate terms, and we propose an active learning algorithm to select a concept from the set of candidate terms for the user and use the stochastic gradient descent algorithm for training the weight of concepts. The experiment results show that this algorithm can effectively assist user acquire domain concepts, when the set of correct terms identified by the CRF model is used as candidate concepts, the value of MAP reaches 0.9335.

A new cascade algorithm based on CRFs for recognizing Chinese verb-object collocation

This paper proposes a new cascade algorithm based on conditional random fields. The algorithm is applied to automatic recognition of Chinese verb-object collocation, and combined with a new sequence labeling of “ONIY”. Experiments compare identified results under two segmentations and part-of-speech tag sets. The comprehensive experimental results show that the best performance is 90.65 % in F-score over Tsinghua Treebank, and 82.00 % in F-score over the segmentation and part-of-speech tagging scheme of Peking University. Our experiments show that the proposed algorithm can greatly improve recognition accuracy of multi-nested collocation, and play a positive role on long distance collocation.

A method of mining bilingual resources from Web Based on Maximum Frequent Sequential Pattern

The bilingual resources are indispensable and vital resources in the NPL fields, such as machine translation, etc. A large amount of electronic information is embedded in the Internet, which can be used as a potential information source of large-scale multi-language corpus, so it is a potential and feasible way to mine a great capacity of true bilingual resources from the Web. This paper proposes a method of mining bilingual resources from the Web based on Maximum Frequent Sequential Pattern. The method uses the heuristic approach to search and filter the candidate bilingual web pages, then mines patterns using maximum frequent sequential, and uses a machine learning method for extending the pattern base and verifying bilingual resources in accordance with the Japanese to Chinese word proportion. The experimental results indicate that the method could extract bilingual resources efficiently, with the precision rate over 90%.

Translation evaluation without reference based on user behavior model

How to evaluate the translation of a machine translation system is a very important research topic. The traditional method of translation evaluation without reference is used to evaluate the translation from the linguistic characteristics mostly. In this paper, the user cost of post-editing operation is considered, and a new method of evaluation translation based on user behavior model is proposed. First of all, track and record the process from the post-editing of machine translation to the forming of the final translation, and extract the decision knowledge of a user behavior; then use the knowledge as an indicator of translation evaluation, and evaluate the machine translation by combining with a language model. Experimental results show that in the absence of reference, the present method is much better than the method of using linguistic characteristics only, and the present method is close to BLEU method with one reference in Spearmens rank order correlation coefficient with human evaluation.