JPS63313275A - Sentence analyzing system - Google Patents

Abstract

PURPOSE:To attain a rapid sentence analysis by providing plural sentence structure analyzers and analyzing the sentence structures of respective sections in a divided sentence in parallel. CONSTITUTION:An input sentence 1 is divided into two parts 8, 9 by a controller 4 and the sentence structures of respective parts 8, 9 are analyzed in parallel by sentence structure analyzers 10, 11 based on a CYK method using a grammatical rule 2. The analyzed result of the analyzer 10 is stored in a triangular part ADB of an analysis tree table 3 and the result of the analyzer 11 is stored in a triangular part BCE. Whether a rule for combining the two parts exists or not is checked, and when the rule is detected, the analysis is ended. When the rule is not detected, a dividing point is changed and said operation is retried. Since the processing time is almost proportional to the cube of the number of words in the analyzed sentence, the sentence analysis is extremely speeded up by division parallel processing.

Description

[Detailed Description of the Invention] [Field of Industrial Application] This invention relates to a text analysis method using a parsing algorithm based on context-free grammar rules, which is a basic algorithm for text analysis processing using a computer. ] Figure 3 is based on, for example, the following document (Suzuki et al.: A study of CYK method syntactic analysis: About quick parsing, Information Processing Society of Japan Natural Language Processing Study Group, Material 59-4゜198)
7) is a configuration diagram showing a conventional text analysis method well known by 7), in which 1 is an input text, 2 is a grammar rule, and 3 is a record of syntactic analysis results including intermediate results of the parsed text. 4 is a parsing device that performs syntactic analysis on one input sentence l using the above grammar rules 2 and parse tree table 3 and records the results in parse tree table 3; 5 is a parsing device for input sentence 1; , 6 is an example of grammar rule 2, and 7 is an example of parse tree table 3. Note that the syntax analysis device 4 is realized by a computer, and the human sentences 12, grammar rules 2, parse tree table 3, etc. are stored in the memory of the computer.

Next, the operation will be explained. When input sentence 1 consists of n words, each word is written as W+, Wl..., W in order from the beginning.
Call it n. Also, the element in the i-th row and j column of the analysis tree table 3 is ti
Call it j.

Grammar rule 2 consists of a set of rules having the form AB, C or Aa. Here, A, B, and C are S.

In grammatical categories such as NP and VP, a represents a word.

The syntactic analysis device 4 sequentially reads each word of the input sentence 1 from the beginning. First, when reading W+, we search for grammar rule 2, find the rule that has Wl on the right side of the arrow (N→W+), and parse the grammar category (A+) on the left side at t+1 in Tree Table 3.
to be recorded.

Next, when Wl is read, the rule (A knee Wl) is similarly found, and A2 is recorded in tzz. Furthermore, 1+
Find the rule (B-4A + + Az) that has the category (and A2) recorded in + and t22 on the right side,
Record B at t12. However, sometimes no rule is found, in which case it does nothing. Also, multiple rules may be found, in which case multiple records will be made.

Generally, the operation of the parser 4 when reading Wi (2≦i≦n) is as follows. First, a rule (Ai-Wi) having Wi on the right side of the arrow is found and Ai is recorded in tit. Furthermore, for tji (1≦j≦1-1), t i4t+ t ! Recording is performed in the order of -211..., ±11. Calculation of the value to be recorded in each tji is based on j≦
For all k such that ≦i-1, have the category (B) recorded in tjk as the first category on the right side,
Find a rule (A-B, C) that has the category (C) recorded in tk+1i as the second category on the right side,
This is to record A in tji. Zero or more categories are recorded in each tji.

As mentioned above, by processing Wi in order from i=1 to n, the triangular matrix (ti
The syntactic analysis result is recorded in the part of j where i≦j). If the grammatical category corresponding to the entire input sentence 1 is recorded in tin, the parsing is successful; otherwise, it is a failure. The result of parsing input sentence example 5 using grammar rule example 6 is parse tree example 7.

[Problem that the invention seeks to solve]

Since the conventional text analysis method is configured as described above, in order to determine the success or failure of parsing, the triangular matrix in the upper left half of parse tree table 3 must be filled in completely.
There was a problem with processing time.

This invention was made to solve the above problems, and it is possible to judge the success or failure of parsing without necessarily filling in the entire triangular matrix in the upper left half of parse tree table 3, and it is better than the conventional method. The purpose is to obtain a text analysis method that can analyze texts in a short amount of time.

[Means for solving problems]

The text analysis method according to the present invention includes a plurality of parsing devices in parallel, divides an input sentence and passes it to each parsing device, and judges the contents of an analysis tree table in which the parsing results are recorded. It is equipped with a control device that controls re-segmentation of sentences and re-execution of syntactic analysis.

[Effect]

Each of the plurality of syntactic analysis devices in this invention is in charge of each fragment of a sentence divided by the control device, executes syntactic analysis in parallel, and independently writes each analysis result to a corresponding portion of the parse tree table. Record.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings. 1st
In the figure, 1 is an input sentence, 2 is a grammar rule, 3 is an analysis tree table for recording syntactic analysis results including intermediate results of the sentence analysis, and 4 is a syntactic analysis device that divides input sentence 1. 8 is an input that has been divided by the control device 4; The left half of sentence 1, 9 is the right half of input sentence 1 divided by the control device 4, and 10 is the result of parsing the left half 8 using grammar rule 2 and parse tree table 3. The parsing device 11 records in the corresponding part of the parse tree table 3, and the parsing device 11 also performs parsing on the right half 9. Note that the control device 4, syntax analysis device 10.11, etc. are realized by a computer, and the input sentence 1 and the left half 8.1 of the divided input sentence 1 are realized by a computer. The right half 9, grammar rules 2, parse tree table 3, etc. are stored in the memory of the computer.

FIG. 2 is a flowchart showing the operation of the control device 4.

Next, the operation will be explained with reference to FIG. Assume that input sentence 1 is composed of n words W+, W2, . . . , Wn.

At step 12, the control device 4 converts the input sentence 1 into Wj
It is divided into a left half 8 from Wj to Wj and a right half 9 from Wj+1 to Wn. Here, j is n/ when n is an even number.
2. When the number is odd, it is set as n+1/2. That is, the input sentence 1 is divided into a left half 8 and a right half 9 at an approximately midpoint position.

The divided left half 8 and right half 9 are then processed in step 13.
and 14, the syntax is parsed in parallel by the parsers 10 and 11.

The processing in steps 13 and 14 is the same as in the conventional method, and the respective analysis results are calculated using the two triangles ΔADB in the analysis tree table 3.
and ΔBEC are respectively recorded. At this time, the analysis result of the entire left half 8 of the input sentence 1 is recorded in tlj, and the analysis result of the entire right half 9 of the input sentence 1 is recorded in tj+In. Next, in step 15,
It is checked whether the analysis result for the entire input sentence 1 can be obtained from the values of tlj and tj+In. That is,
Categories recorded in tlj and tj+ln (At
and A2) on the right-hand side and the grammatical category S corresponding to the entire input sentence 1 on the left-hand side (S-A r , A
If 2) is found, it means that the solution has been found; otherwise, it can be seen that the solution has not been found.

If a solution is found, the process ends in step 16. At this time, compared to the conventional method, it is possible to omit the processing for recording the part of the rectangle DBEF in the parse tree table 3, and the processing time is also lower than that of the conventional method for words included in input sentence 1. Since it is approximately proportional to the cube of the number n, a parallel analysis of the divided left half 8 and right half 9 results in about 178.

If no solution has been found, a check is made in step 17 to see if all n-1 ways of dividing input sentence 1 have been tried. If so, in step 18, the analysis is deemed to have failed and the process ends.

Otherwise, in step 19, another division method is selected for input sentence 1. Specifically, similar to the process in step 12, the control device 4 converts input sentence 1 from Wj to w.
It is divided into a left half 8 up to Ma 3 and a right half 9 from Wj'+1 to Wn. The value of jl at this time is selected to be different from the method of division already performed in the previous step 12 or step 19. In this example, the parse tree table 3
Divide at the position indicated by point I.

Subsequently, the subdivided left half 8 and right half 9 are parsed in parallel by parsers 10 and 11 in steps 13 and 14. At this time, since the values already recorded in the analysis tree table 3 can be used, the part of the analysis tree table 3 that needs to be newly recorded becomes the rectangle DBHG. That is,
In reality, only the parsing device 10 in charge of the left portion 8 operates, and the parsing device 11 in charge of the right portion 9 only checks that no calculation is required and then stops. Steps 13 and 1
As a result of step 4, the analysis results of the entire left half 8 and the entire right half 9 of the input sentence 1 are recorded in t lj' and tj''+In, respectively.

Thereafter, the process described above is repeated from step 15 onwards.

If a solution exists for input sentence 1, it means that the parsing was successful and the process ends in step 16, and if no solution exists even after repeating the above process, it means that the parsing has failed and the process ends in step 18. The process ends at .

When the process ends in step 16, unlike the conventional method, it is possible to determine whether the parsing is successful without necessarily filling up the entire triangular matrix ΔAFC in the upper left half of the parsing tree table 3. The area where recording may be omitted is rectangle D.
The closer the part is inside BEF and to the vertex F, the more likely it is to be omitted, and the more calculations will be omitted. Therefore, sentences can often be analyzed in a fraction of the processing time compared to conventional methods. Note that when the process ends in step 18, that is, when syntactic analysis fails, the amount of time and complexity is the same as in the conventional method.

In the above embodiment, two parsing devices 10 and 11 were used, but even if three or more are used, it is possible to configure the device according to the same analysis method, and the same effect as in the above embodiment can be achieved. .

〔Effect of the invention〕

As described above, according to the present invention, a plurality of parsing devices are provided in parallel, an input sentence is divided and passed to each parsing device, and the contents of an analysis tree table in which the parsing results are recorded. It is equipped with a control device that judges the re-segmentation of the text and re-executes the syntactic analysis, and executes syntactic analysis in parallel using multiple syntactic analysis devices for each fragment of the divided text. Since the results are recorded independently in the corresponding parts of the analysis tree table, this method has the effect of allowing sentences to be analyzed in less processing time than the conventional method.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing a text analysis method according to an embodiment of the present invention, FIG. 2 is a flowchart showing a processing procedure according to an embodiment of the present invention, and FIG. 3 is a block diagram showing a conventional text analysis method. be. ■ is the input sentence, 2 is the grammar rule, 3 is the parse tree table, 4 is the control device, 8 and 9 are the left and right halves of the divided input sentence 1, 1
0 and 11 are syntactic analysis devices. In addition, in the figures, the same reference numerals indicate the same or corresponding parts. Agent Masuo Oiwa (and 2 others) Figure 1 Figure 2

Claims (1)

[Claims] Performs syntactic analysis on the input text and analyzes the results based on predetermined grammar rules stored in the computer's memory and a parsing table for recording syntactic analysis results including intermediate results of syntactic analysis. In a text analysis method that is equipped with a syntax analysis device that records data on a tree table, and performs syntax analysis of the entire sentence by sequentially parsing the words of the input text, it is equipped with a plurality of the above syntax analysis devices in parallel, and The system is equipped with a control device that divides the divided text and passes it to each parsing device, judges the contents of the parse tree table in which the parsing results are recorded, and controls the re-segmentation of the text and the re-execution of the parsing. A sentence characterized in that syntax analysis is executed in parallel on each fragment of a sentence using a plurality of syntax analysis devices, and each analysis result is independently recorded in a corresponding part of an analysis tree table. Analysis method.