KR20080024638A - Apparatus and method for syntactic analysis using detection of parallel structure - Google Patents

Abstract

A device and a method for analyzing syntax by recognizing a parallel structure are provided to divide excessively long sentences found in a patent document with high correctness, and raise syntax analysis efficiency/correctness by dividing the sentence into more syntaxes. A chunking part(100) tags and partially parses an inputted raw English document. A node recognizer(200) recognizes a parallel node starting point of the raw document. A similarity calculator(300) calculates similarity weight among parallel nodes based on the similarity of vocabulary/part-of-speed of a starting word, a head, and the word next to the head. A parallel structure recognizer(400) searches all available parallel structures based on the calculated similarity weight, calculates the weight of the searched parallel structures, and recognizes the parallel structure of the raw document based on the calculated weight of the parallel structure. A parallel structure parser(500) parses the recognized parallel structure. A whole sentence parser(600) parses the whole raw sentence again which is inputted from a parsing result.

Description

Parsing apparatus and method through parallel structure recognition {APPARATUS AND METHOD FOR SYNTACTIC ANALYSIS USING DETECTION OF PARALLEL STRUCTURE}

1 is a block diagram showing an apparatus for parsing through parallel structure recognition according to an embodiment of the present invention.

2 is a diagram illustrating a similarity table between nodes after node recognition according to an embodiment of the present invention.

3 is a diagram illustrating a similarity table between nodes after node recognition according to another embodiment of the present invention.

* Explanation of symbols for the main parts of the drawings *

100: chunking unit 200: node recognition unit

300: similarity calculation unit 400: parallel structure recognition unit

500: parallel structure parser 600: full sentence parser

The present invention relates to a syntax analysis apparatus for an automatic translation system, and more particularly, to a syntax analysis apparatus and method through frequent parallel structure recognition in patent sentences.

In the abstract or claims of the patent document, excessively long sentences frequently appear in the general domain. In order to analyze such a long sentence, the conventional syntax analysis method mainly uses a method of splitting a sentence around a boundary point of a clause such as a conjunction or a comma.

However, in the case of long sentences appearing in the patent document, the length of the sentence is considered to be longer, rather than the combination of the conjunctions or the arrangement of sentences is mainly due to the parallel structure in which the same structures such as nouns and verb phrases are arranged. Likewise, nouns and verb phrases that form a parallel structure are also complex structures, often containing clauses again.

Therefore, in the above-mentioned patent document, if the entire sentence is simply divided based on a clue such as a conjunction or a comma, there is a problem that it is likely to cause an error and may not be divided into enough units to parse. there was.

Accordingly, an object of the present invention is to provide an apparatus and method for parsing through parallel structure recognition, which performs syntax analysis by increasing the degree of recognition of parallel structures typically shown in patent documents.

To this end, the apparatus for parsing through parallel structure recognition according to an embodiment of the present invention, the chunking unit for performing tagging and partial parsing on the input English text; A node recognition unit for recognizing a parallel node start point with respect to the original text; A similarity calculator for calculating similarity weights between the parallel nodes; A parallel structure recognition unit searching all possible parallel structures based on the calculated similarity weights, calculating a weight of the found parallel structures, and recognizing a parallel structure of the original text; A parallel structure parsing unit for parsing the recognized parallel structure; And an entire sentence parsing unit that parses the entire original sentence again from the parsing result.

In addition, the syntax analysis method through a parallel structure recognition according to another embodiment of the present invention, (a) performing tagging and partial parsing on the input English text; (b) recognizing a parallel node start point for the text; (c) calculating similarity weights between the parallel nodes; (d) searching for all possible parallel structures based on the calculated similarity weights, and calculating weights of the searched parallel structures to recognize parallel structures for the original text; (e) performing parsing on the recognized parallel structure; And (f) parsing the entire original sentence again from the parsing result.

Hereinafter, with reference to the accompanying drawings and embodiments will be described the present invention in detail. Like reference numerals designate like elements throughout the drawings.

1 is a block diagram illustrating a syntax analysis apparatus through parallel structure recognition according to an embodiment of the present invention.

As shown in FIG. 1, a chunking unit 100 that performs tagging and parsing on an input original text, a node recognition unit 200 recognizing a parallel node starting point, and a similarity calculation unit calculating similarity weights between parallel nodes ( 300), a parallel structure recognition unit 400 for recognizing all possible parallel structures and obtaining parallel structure weights, a parallel structure parser 500 for parsing the recognized parallel structures, and the entire structure again from the parsed results. It includes a full sentence parser 600 for parsing a sentence.

The chunking unit 100 performs tagging and partial parsing on the input original text, and extracts only base noun phrases (base NP) from the result to perform basic noun phrase chunking. Here, the basic noun phrase means a noun phrase that does not include the noun phrase again inside.

The node recognition unit 200 recognizes the starting point of the parallel phrase that is an element forming the parallel structure. At this time, the types of parallel phrases are noun phrases, verb phrases, prepositional phrases, noun clauses, and the like, and are classified into three types according to their positions in the parallel structure: start node, intermediate node, and end node. The node recognizer 200 uses a syntax pattern to recognize starting points of these nodes. Table 1 below is an example pattern for recognizing the starting point of the parallel phrase for the English patent document. However, this pattern is merely exemplary and is not limited thereto.

Start node Intermediate node End node include / comprise / comprise / have + NP / VBG step of VBG such as NP VERB that, , NP / VBG / PREP / CONJ (,) and NP / VBG / PREP / CONJ

Where NP is a noun phrase, VBG is an ing verb, PREP is a preposition, and CONJ is a conjunction / subject.

The similarity calculator 300 calculates similarity between the recognized parallel sphere nodes. Similarity S _ij between node i and node j is calculated as shown in Equation 1 below.

S _ij = s0 × (w1 × s1 + w2 × s2).

Where s0: 0 if the nodes are different or the article is different, 1 otherwise

s1: indicates the vocabulary / part-of-speech similarity of the head (e.g., 3 if the vocabulary is the same, 1 if the vocabulary of the vocabulary is the same, 0 otherwise)

s2: The lexical / part-of-speech similarity of the word following the head (eg, 2 if the vocabulary of the head is the same, 1 if the part-of-speech is the same, 0 otherwise). Here, the head refers to a word that is a reference among the words constituting the phrase.

The parallel structure recognizer 400 generates all possible parallel structures, and calculates the final parallel structure weights for each parallel structure based on the similarity calculated by the similarity calculator 300. The parallel structure consists of a start node, zero or more intermediate nodes, and an end node. Thus, for all possible combinations of start and end nodes of nodes whose node similarities are nonzero, it is possible to create a possible parallel structure by adding intermediate nodes whose end and node similarities are not zero. In addition, the weight for the parallel structure is basically calculated as the sum of the similarities between the nodes, and the total length of other parallel structures, the existence of another start or end node in the middle of the parallel structure, the similarity difference between the parallel nodes, and the presence of a specific vocabulary. The weight according to and the like is added or subtracted to calculate the final parallel structure weight.

The parallel structure parsing unit 500 recognizes the range of each parallel sphere node of the corresponding parallel structure in the order of the highest weight, and parses the range of each recognized parallel sphere node, thereby parsing all the parallel sphere nodes. Choose the final parallel structure for success.

The whole sentence parsing unit 600 parses the entire sentence again in a state in which the recognized parallel structure is replaced with one node, thereby outputting a final syntax parsing result. Therefore, the result of parallel parsing with improved accuracy of the input sentence can be obtained.

Hereinafter, a syntax analysis method through parallel structure recognition according to an embodiment of the present invention will be described through the following actual English patent example sentences.

Example 1

"To attain the above object, the laser irradiating system according to the present invention comprises a light source for emitting a laser beam in elliptical shape, a light source holder for holding the light source, a base for rotatably supporting the light source holder, a first driving unit for rotating the light source holder, a deflecting optical means for deflecting the laser beam from the light source in a direction perpendicular to an optical axis, a rotator holder for holding the deflecting optical means and being rotatable around the optical axis of the light source, a cylindrical lens being arranged so that a center of the cylindrical lens is aligned with the optical axis and the laser beam from the deflecting optical means enters perpendicularly to the optical axis and for diffusing the transmitting laser beam to a fan-shaped laser beam, a control unit for controlling the first driving unit and a receiver for receiving a signal for remote-control operation, wherein an irrad iating direction of the fan-shaped laser beam can be operated in remote control operation by the control unit based on a command signal received by the receiver. "

2 illustrates a similarity table between nodes after node recognition according to an embodiment of the present invention. In the similarity table, row i (node field) and column j represent node similarity S _ij .

Referring to FIG. 2, each field includes a node field, a text position field, a parallel structure index field, and a syntax head field. In addition, the similarity calculation unit 300 fills each field data (node similarity S _ij ) based on Equation (1). Based on this similarity table, parallel structure recognition is started. Here, "<", "," and ">" represent a start node, an intermediate node, and an end node, respectively. For example, on a similarity table based on Example 1, a parallel structure can be formed starting at node 0 and ending at node 12, with nodes 1, 2, 3, 4, 5, 7, 11 as intermediate points. It is a parallel structure.

Therefore, in the case of Example 1, since there is no other candidate, the parallel structure is selected as the final structure. Then, for each node, syntax parsing of NP (noun phrase) is performed for the range except comma or parallel conjunction, and the final parallel structure is "NP-> NP, NP, ..., NP, NP, NP CC NP (Noun phrases-> noun phrases, noun phrases, ..., noun phrases, noun phrases, noun phrases, noun phrases, noun phrases noun phrases) "is replaced by a single NP, and the entire sentence structure is" To attain the above object, the laser irradiating system according to The present invention comprises NP ", which makes it easy to parse and get final syntax parsing.

Looking at one embodiment of the present invention through another example sentence is as follows.

[Example 2]

"In accordance with an aspect of the present invention, for use with a doctor blade holder of a printing machine, a combined doctor blade and shield includes a doctor blade adapted to be secured to the doctor blade holder in a manner so as to extend outwardly therefrom at least substantially in engagement with an outer surface of a roll for metering ink on the roll, the doctor blade being formed as a doctoring plate having a rear portion adapted to be secured to the doctor blade holder and a front portion with a doctoring edge adapted to at least substantially contact the outer surface of the roll to meter ink adjusted, and a doctor blade shield having one end fixedly secured to the doctor blade in spaced relation away from the doctoring edge and a second opposite free end unsecured to the doctoring plate , so as to cover at least a portion of the outer surface of the doctor blade holder to prevent ink from depositing on the outer surface of the doctor blade holder when the combined doct or blade and doctor blade shield is clamped to the doctor blade holder. "

3 illustrates a similarity table between nodes after node recognition according to another embodiment of the present invention.

Referring to FIG. 3, in the similarity table created based on the example 2, the combinable parallel structure includes a plurality of parallels such as nodes 0 to 9, nodes 4 to 9, nodes 0 to 6, nodes 4 to 6, and the like. The structure is possible. In this case, the simple weighted sum of the node similarity to which Equation 1 is applied is as follows.

Node 0 to Node 9: 7

Node 4 to Node 9: 7

Node 0 to Node 6: 7

Node 4 to Node 6: 11

..., etc. Here, in the case of node 4 and node 6, since the starting point or end point, the similarity of the corresponding node is not added at nodes 0 to 9 and nodes 0 to 6.

In this case, the sum of simple weights is highest among nodes 4 to 6, but considering the weight of the total length of the parallel structure, the penalty for the start node and the end node included in the middle of the structure, the parallel structure of nodes 0 to 9 is The final parallel structure is chosen.

In this sentence, when the general sentence segmentation method is used, it is likely to be divided first in a comma of "on the roll, the doctor blade being". If so, a completely incorrect parsing result may be obtained. Depending on the method, correct parsing results can be produced by first recognizing the parallel structure.

While the present invention has been described in detail with reference to preferred embodiments, those skilled in the art will recognize that various modifications or changes can be made without departing from the spirit and scope of the invention.

As described in detail above, according to the present invention, since sentence division accuracy is increased for excessively long sentences appearing in a patent document, and sentence division is possible with more phrases, syntax efficiency and accuracy are improved.

Claims (8)

A chunking unit that performs tagging and partial parsing on the input English text; A node recognition unit for recognizing a parallel node start point with respect to the original text; A similarity calculator for calculating similarity weights between the parallel nodes; A parallel structure recognition unit searching all possible parallel structures based on the calculated similarity weights, calculating weights of the found parallel structures, and recognizing parallel structures of the original text based on the calculated parallel structure weights; A parallel structure parsing unit for parsing the recognized parallel structure; And And a full sentence parsing unit which parses the entire original sentence again from the parsing result. The method of claim 1, wherein the node recognition unit, And a starting point of the parallel node from a basic noun phrase chunking result of the original text using a specific syntax pattern. The method of claim 1, wherein the similarity calculation unit, And a similarity between the parallel nodes using the similarity between the vocabulary and the part of speech of the starting word, the vocabulary and the part of speech of the head, the vocabulary and the part of speech of the next word of the head. The parallel structure recognition unit according to any one of claims 1 to 3, wherein Based on the calculated similarity weights, for all possible parallel structure combinations consisting of a start node, zero or more intermediate nodes, and an end node, a parallel weight is obtained by adding or subtracting a specific weight to the sum of similarities between the nodes. Parsing apparatus through parallel structure recognition. The parallel structure parsing unit according to any one of claims 1 to 3, wherein Recognizing the range of each parallel sphere node of the parallel structure in order of the parallel structure weights being high, parsing all the parallel sphere nodes after parsing the range of the recognized parallel sphere node is successful. Parsing apparatus using the parallel structure recognition, characterized in that for selecting the parallel structure with respect to. The whole sentence parsing unit according to any one of claims 1 to 3, wherein And parsing the entire sentence again in a state where the recognized parallel structure is replaced with one node, and outputting a final syntax parsing result. (a) performing tagging and partial parsing on the inputted English text; (b) recognizing a parallel node start point for the text; (c) calculating similarity weights between the parallel nodes; (d) searching for all possible parallel structures based on the calculated similarity weights, and calculating weights of the searched parallel structures to recognize parallel structures for the original text; (e) performing parsing on the recognized parallel structure; And and (f) parsing the entire sentence of the original text from the parsing result again. A computer-readable recording medium recording a parsing program through parallel structure recognition, The program, (a) performing tagging and partial parsing on the inputted English text; (b) recognizing a parallel node starting point for the text; (c) calculating similarity weights between the parallel nodes; (d) searching for all possible parallel structures based on the calculated similarity weights, and calculating weights of the found parallel structures to recognize parallel structures for the original text; (e) performing parsing on the recognized parallel structure; And and (f) parsing the entire sentence of the original text from the parsing result again.

Images