언어 자원이 부족한 언어 쌍에 대한 다국어 사전 추출

Abstract

In general, bilingual and multilingual lexicons are important resources in many natural language processing fields such as information retrieval and machine translation. Such lexicons are usually extracted from bilingual (e.g., parallel or comparable) corpora with external seed dictionaries. However, few such corpora and bilingual seed dictionaries are publicly available for many language pairs such as Korean–French. It is important that such resources for these language pairs be publicly available or easily accessible when a monolingual resource is considered.

This thesis presents efficient approaches for extracting bilingual single-/multi-word lexicons for resource-poor language pairs such as Korean–French and Korean–Spanish. The goal of this thesis is to present several efficient methods of extracting translated single-/multi-words from bilingual corpora based on a statistical method.

Three approaches for single words and one approach for multi-words are proposed. The first approach is the pivot context-based approach (PCA). The PCA uses a pivot language to connect source and target languages. It builds context vectors from two parallel corpora sharing one pivot language and calculates their similarity scores to choose the best translation equivalents. The approach can reduce the effort required when using a seed dictionary for translation by using parallel corpora rather than comparable corpora. The second approach is the extended pivot context-based approach (EPCA). This approach gathers similar context vectors for each source word to augment its context. The approach assumes that similar vectors can enrich contexts. For example, young and youth can augment the context of baby. In the investigation described here, such similar vectors were collected by similarity measures such as cosine similarity. The third approach for single words uses a competitive neural network algorithm (i.e., self-organizing maps

SOM). The SOM-based approach (SA) uses synonym vectors rather than context vectors to train two different SOMs (i.e., source and target SOMs) in different ways. A source SOM is trained in an unsupervised way, while a target SOM is trained in a supervised way.

The fourth approach is the constituent-based approach (CTA), which deals with multi-word expressions (MWEs). This approach reinforces the PCA for multi-words (PCAM). It extracts bilingual MWEs taking all constituents of the source MWEs into consideration. The PCAM

2

identifies MWE candidates by pointwise mutual information first and then adds them to input data as single units in order to use the PCA directly.

The experimental results show that the proposed approaches generally perform well for resource-poor language pairs, particularly Korean and French–Spanish. The PCA and SA have demonstrated good performance for such language pairs. The EPCA would not have shown a stronger performance than expected. The CTA performs well even when word contexts are insufficient. Overall, the experimental results show that the CTA significantly outperforms the PCAM.

In the future, homonyms (i.e., homographs such as lead or tear) should be considered. In particular, the domains of bilingual corpora should be identified. In addition, more parts of speech such as verbs, adjectives, or adverbs could be tested. In this thesis, only nouns are discussed for simplicity. Finally, thorough error analysis should also be conducted.