Text-oriented relative position information extraction method

Abstract

The position information not only comprises geographically named entities, but also is an important part of relative position information between the entities. However, the expression forms of the relative positions are diversified, are very close to daily languages, are difficult to identify through fixed grammar rules, and need to explore a new method for extraction. The invention provides a text-oriented relative position information extraction method. According to the method, a BiLSTM (bidirectional LSTM) + CRF (conditional random field) model and a spatial semantic feature template are utilized to train data in a text, place name entity information and relative position information in the text are extracted, a semantic structure in a Chinese text and an obtained external feature library are combined, a position indicating information dictionary is constructed, a recognition rule of the relative position information is formulated. And the precision of relative position information extraction is further optimized.

Description

technical field [0001] The invention relates to a text-oriented location information extraction method, in particular to a text-oriented relative location information extraction method. Background technique [0002] Location information not only includes geographic named entities, but relative location information between entities is also an important part. However, unlike the spatial position expressed by coordinates and other means, the spatial relationship in text is often ambiguous. Most of the descriptions are qualitative expressions, such as common words such as "surrounding" and "next to", from which precise location descriptions cannot be obtained. Geographically named entities are often fixed and represent a definite location. Therefore, the combination of geographic named entities and relative location information between entities is an ideal method for extracting location information in text. However, relative positions can be expressed in various forms, which ...

AI Technical Summary

Problems solved by technology

However, relative positions can be expressed in various forms, which are very close to everyday language, and it is difficult to identify them through fixed grammatical rules. We need to explore new methods to extract them.

Existing related inventions often only focus on the extraction method of geographically named entities, ignoring the identification and transformation of the relative positional relationship between entities, and lack of relevant full-location information corpus

At the same time, the existing recognition methods still have many deficiencies, such as low recall rate for complex place name recognition, inaccurate recognition range, etc.

Images