Document processing device, document processing method, and document processing program

#The present invention relates to a document processing device, a document processing method, and a document processing program, and particularly to an interactive document search device, an interactive document search method, and an interactive document search program for obtaining an answer sentence to a question sentence.

#An interactive search device is a device that receives a question related to a specific field in a natural sentence and returns an answer based on the natural sentence by a system, and is also called a question-answer system. As a device of this type, for example, Patent Literature 1 (JP-A-2013-143066) discloses a technique in which numerous comment sentences related to keywords extracted from question sentences are classified into potential topic groups, answer sentence candidates are selected by associating answer sentences similar to a sentence group included in the topic group, and a representative keyword, which is a difference between the selected answer sentence candidates, is clearly indicated and selected by a user.

#According to the technique described in Patent Literature 1, the answer sentence to the question sentence is obtained, and in particular, when a plurality of answer sentence candidates exist for the question sentence of the user, an answer sentence reflecting an intention of the user is clearly indicated (narrowed down).

#FIG. 1 is a diagram showing an example of an interactive search system according to an embodiment.

#FIG. 2 is a diagram showing a configuration example of the interactive search device according to the embodiment.

#FIG. 3 is a diagram showing a flowchart of a normalization DB registration processing according to the embodiment.

#An example of an interactive search device will be described as a preferred embodiment of a document processing device. That is, the interactive search device for estimating a word that may be omitted as a synonym based on a document group of a search target, unifying the synonym and expression variation in both an input question sentence input by a user and a hypothetical question sentence of the search target, and performing a search after performing a normalization processing for complementing an omitted word will be described.

#When any of word A and word B in the document are used in the same context, it is considered that the word A and the word B are used in the document with relatively similar meanings. Whether two sentences are in the same context depends on words configuring the sentences and dependency relationships thereof, but the sentences in the same context are often configured with the same word, so that a determination that whether the two sentences are in the same context can be made to some extent only by the matching number of words. Therefore, if words appearing around the word A and the word B match to some extent, the word A and the word B are used in the same context, that is, the word A and the word B can be considered to be words having similar meanings.

#A word embedding vector is a technique for numeralizating a similarity of meanings between words under such an assumption. The word embedding vector is a method of converting words into real number vectors of about 200 dimensions based on a statistic of a plurality of words that appear before and after each of all words in a learning corpus, and can measure the similarity of the meanings between the words by a cosine similarity between vectors. By generating the word embedding vector based on a target document group and extracting a word whose cosine similarity is equal to or larger than a certain value for any word, a word used in a meaning similar to the word can be extracted.