This paper presents a research for extracting technical terms from documents in the nursing domain.An exploratory study showed that a well-known term extraction method, which has proven to be effective in extracting technical terms specified to the computing domain, can not effectively extract technical words representing symptoms or treatments of diseases.We propose a new technical term extraction method to improve extraction performance.Its main characteristics are enhancing permissible combinations of word-class;and excluding fundamental vocabulary.Experimental results showed that our extraction method attained 99% recall, which is 16% better than the recall attained by the base method.In addition, 13% of precision rate is improved by excluding fundamental vocabulary.

View full abstract

Identifying discourse relations is important for many applications, such as text/conversation understanding, single/multi-document summarization and question answering. This paper focuses on discourse relations between two successive Japanese sentences, and classified the relations into six classes in terms of their relation types. We propose an example-based method in which the system determines the discourse relation between sentences similar to two input sentences. The method utilizes phrasal pattern made from input sentences and core words in input sentences. As an evaluation result, the accuracy attains over 75% for human judgment. Moreover, it is expected for the example-based approach that the accuracy has been improved further according to the increase of the instances.

View full abstract

A novel task for evaluating question answering technologies is proposed.This task assumes interactive use of question answering systems and evaluates among other things, the abilities needed under such circumstances, i.e.proper interpretation of questions under a given dialogue context;in other words, context processing abilities such as anaphora resolution and ellipses handling.This paper shows the design of the task and its empirical background.The task proposed is not only novel as an evaluation of the handling of information access dialogues, but also includes several valuable ideas such as a measuring metric in order to obtain intuitive evaluation of the answers to list-type questions and reference test sets for obtaining information on context processing ability in isolation.

View full abstract

This paper represents a method of acquiring polar sentences from HTML documents. The basic idea is to exploit three lexico-syntactic patterns and two layout structures of HTML documents.The method requires only a small amount of hand-crafted rules and can be implemented in low cost.In our experiment, the method was applied to one billion documents and 650 thouthands polar sentences were aquired.

View full abstract

The words which are not defined in Thesaurus appear in the daily conversation, including new words and proper nouns.It is unable to carry conversation with no knowledge about these words.We can search about an unknown word by using Web.However, it is difficult to obtain necessary information efficiently because large amounts of information exists in Web.The meaning of the unknown word is acquired by presenting an appropriate node of Thesaurus.As for the research on understanding of the unknown word, there are a lot of researches which relies on language data including corpus.These researches have a problem that there are unknown words which can not be responded.This paper proposes the technique for finding the best node for the unknown word included proper nouns by conceptualizing the unknown word and evaluating relationship to each node.

View full abstract

Recently, there are a lot of automatic summarization systems.Almost all previous works figure an importance for each word or each sentence, and compress or extract a sentence by using the importance of each word or each sentence.However, when we generate a summary, we use much knowledge and experience in our mind.Therefore, it is difficult to compute the importance which correlates with human sense.This paper proposes a new summarization method which is based on example-based approach. The method has three steps.First, system retrieves a similar instance in a instance collection to an input.The instance collection indicates summaries which are generated by human.In the second step, the system links the similar phrases in the input to a phrase in the similar instance.As third step, the system combines the corresponding phrases, and outputs summary candidates.Experimental results have proven that the summarization system attains approximately 1.81 accuracy on a scale 1 to 4 by human judgments.And the system has obtained better accuracythan previous work.From the examinations, the system has confirmed that the summaries were generated by combining the phrases in many position of the input, while those summaries are not given just by common methods such as sentence extraction methods and sentence compression methods.

View full abstract