JPH0414175A - Analyzing device for natural language sentence - Google Patents

Abstract

PURPOSE:To obtain an analyzing device for natural language sentences which has a wide range of application by calculating the inadequacy of a function structure by means of the limitation of a limiting part and a dictionary part at production of the function structure. CONSTITUTION:An input sentence given from an input device 2 is divided into morphemes by a dictionary part 6, and plural component element structures are produced for those morphemes based on the contents of a syntax rule part 4. Then a function structure is produced from each component element structure. At production of the function structure, a controller 1 applies the limitation by means of the limitation of a limiting part 5 and the dictionary 6. Then the inadequacy of the function structure is calculated and the analyzing result is outputted from an output device 3. In such a constitution, the ambiguity is satisfactorily deleted and at the same time even an exceptional metaphorical expression can be analyzed. Thus it is possible to obtain an analyzing device for natural language sentences which has a wide range of application with no occurrence of explosion in terms of space and time.

Description

[Detailed Description of the Invention] The present invention relates to a natural language sentence analysis device, and more specifically,
The present invention relates to a natural language sentence parsing device used in realizing a system that takes natural language sentences as input, such as a machine translation device.

When parsing a natural language sentence, precise analysis is difficult only by constructing a structural data structure such as a parse tree. Therefore, regular functional grammar is a grammatical theory that constitutes a semantic data structure in addition to a structural data structure, and provides a framework for describing grammar complementary to these two data structures.

Regarding this, “Ding he Mental Repr.
estation of GramIIlatica
l Relations, J (Bresnan, J,
MITPress, 1982).

In this overhead functional grammar, the latter semantic data structure is called a functional structure, and the functional structure is the final output.

Although the functional structure is ``semantic,'' it is a structure similar to a superficial case frame, and cannot be said to be an appropriate structure as a semantic structure for natural language sentences. Therefore, applications that use an analysis system, such as machine translation, must obtain necessary semantic information from an external information source such as a dictionary based on information about a single functional structure.

The information extracted from external information sources is the semantic information of miscellaneous expenses, and we will try to handle this semantic information within the framework of business function grammar, and furthermore, use this semantic information to resolve ambiguity in analysis. For example, rLF
"Towards the fusion of G and semantic analysis" (Yoshito Nitta Information Processing Society Natural Language Processing Research Group Proceedings, 68-2, 1988)
There is. Here, "ambiguity" means that multiple solutions are output as an analysis result. However, this conventional technique has the following problems.

(1) Grammatical properties cannot be leveled by syntactic strength.

(A) There cannot be more than one subject in a simple sentence.

(B) The subject of "drink" is an animal.

Regarding the above two grammatical phenomena, (A) was given as an example of a very strong grammatical property, and (B) was given as an example of a weaker grammatical property. These two grammatical properties differ greatly in their ``strength'', or in other words, the number of exceptions. Therefore, in syntactic analysis, property (A) must always hold true, but property (B) should be treated as not holding true in some cases. In other words, a syntactic analysis that treats an analysis result in which the property (B) does not hold as a failure cannot analyze an exceptional and comparative sentence such as "Cars drink gasoline." However, parsing that does not use properties like (B) cannot eliminate ambiguity in the parsing.

In addition, since all grammatical properties are checked when constructing a functional structure, the choice of whether to use or not use grammatical properties is based on ``removal of ambiguity'' and ``resistance to exceptions.'' It becomes a trade-off. In other words, there is a problem in that if you try to remove more ambiguity, you will not be able to analyze exceptional expressions, and if you make it possible to analyze even exceptional expressions, you will not be able to remove ambiguity. This is because grammatical properties cannot be leveled by syntactic strength.

In addition, semantic analysis is performed as a phase after syntactic analysis,
Conventional techniques that attempt to resolve ambiguity in analysis include, for example, "Systematic Utilization of Linguistic Knowledge in Semantic Analysis" (Yuji Ikeda and two others, Information Processing Society of Japan's 39th National Conference Materials, 5F-8゜1989) There is. This technology describes linguistic knowledge as production rules, and for each production rule, scores are given according to the degree of conformity of the solution to the accompanying conditions. This eliminates ambiguity in analysis by selecting the one with the best score. However, there are the following problems.

(2) - Spatiotemporal and temporal explosions are likely to occur.

If we try to make it robust against exceptions, there are only a few analysis results that can be removed at the time of syntactic analysis, and as a result, resolving ambiguity in analysis will depend heavily on semantic analysis processing.

Therefore, at the end of the syntax analysis process, the number of analysis results temporarily increases significantly, which can make processing difficult both spatially and temporally.

This phenomenon always occurs when the length of a sentence becomes long, and the problem is that the sentences that can be analyzed are short. The cause of this problem is the two-phase processing of syntactic analysis and semantic analysis. In other words, unless semantic analysis is performed in parallel from the point of syntactic analysis, ambiguity in the analysis cannot be removed and spatial and temporal explosions are likely to occur.

Regarding the above problem (1), grammatical properties are expressed as constraints, information about the grammatical strength of the constraints is added to each constraint, and ambiguity in the analysis is removed using this information. The information regarding the grammatical strength is expressed, for example, as a numerical value, and the magnitude of the value serves as a standard for leveling.

Regarding problem (2) above, by not modularizing processing such as syntactic analysis and semantic analysis, but by describing both as the same constraint, it is possible to solve the problem by
is uniformly treated as a grammatically strong constraint, and ``semantic analysis'' is treated as a grammatically weak constraint. In addition, even during the analysis, if some constraints are not satisfied and it is considered not to be a solution, it is temporarily removed from the analysis process to prevent spatial and temporal explosions.

■-The present invention was made in view of the above-mentioned circumstances.
To provide a natural language sentence analysis device which can sufficiently remove ambiguities, can also analyze exceptional figurative expressions, does not cause spatial or temporal explosions, and therefore has a wide range of applications. It was made for the purpose of

In order to achieve the above-mentioned object, the present invention includes an input device for inputting a natural language sentence, a dictionary section that divides the input sentence from the input device into morphemes, and divides the morpheme-divided sentences into Then, a plurality of constituent structures are created using the contents of the syntax rule section, a functional structure is created from each of the constituent structures, and when creating the functional structure, the constraints from the constraint section and the dictionary section are used. The present invention is comprised of a control device that applies constraints using the control device, and an output device that outputs analysis results by the control device, and is characterized in that the incompatibility of the functional structure is calculated when the functional structure is created. .

The natural language sentence analysis device of the present invention divides an input sentence into sentence components, recursively combines some of the components to form a larger component, and finally obtains a valid sentence. Output the analysis results that have become constituent elements as valid analysis results.

When constructing a larger component by combining several components, several constraints are checked to realize the configuration. Since each constraint has information regarding the grammatical strength of the constraint, the information regarding the grammatical strength is comprehensively evaluated for constraints that are not satisfied, and the constraints are considered to be incompatible at the time of composition. When the incompatibility becomes large to a certain extent, the analyzer considers the analysis state to be unsuitable and temporarily excludes it from future analysis. Therefore, only analysis states with smaller incompatibility, ie, greater compatibility, are subject to analysis processing. Hereinafter, the explanation will be based on an embodiment of the present invention.6 Fig. 1 is a block diagram for explaining an embodiment of a natural language sentence analysis device according to the present invention.In the figure, 1 is a control device, 2 is an input 3 is an output device, 4 is a syntax rule section, 5 is a constraint section, and 6 is a dictionary section.

The input device N2 includes input from an input device such as a keyboard or a storage device. The output device 3 includes a case where the analysis result is visually confirmed on a display or the like, or a case where it is an application using a syntax analysis device such as in the case of machine translation.

Figures 2 (a) to (c) show a dictionary D, a syntax rule R1, a constraint C
are shown respectively.

Each element of the dictionary shown in Figure 2 (a) is a headword, a word range! I (part of speech) is expressed as a feature. A feature is a list of the form (feature name, feature value). Here, the characteristics are extracted characteristics of miscellaneous expenses.

The syntactic rule R shown in FIG. 2(b) is written in a phrase structure grammar with labels. Each element on the right-hand side is a labeled or unlabeled nonterminal. (R1) in (
N P ; topic), the nonterminal symbol is NP and the label is topic. Also, among the non-terminal symbols on the right side, the lower-case alphabetic characters are pre-terminal symbols (opening ranges), and the upper-case alphabetic characters are non-terminal symbols other than the pre-terminal symbols. Also, the labels in the syntax rules represent function names. Here, the term "function name" is used in the same manner as in the known practice function grammar. That is, the description of (R1) is the same as the following RIO in the various quantity functional grammar.

(RIO) VP->NP VP↓=↑
topic ↓:↑ Constraint C is a production rule with a function name and a penalty. The notation is in the form of a (Function name: Penalty) constraint rule. Each constraint rule refers to information in the functional structure, particularly feature values extracted from a dictionary. Here, the term "functional structure" is used in the same manner as in the known word search functional grammar, and is a recursive matrix in which the functional name is the attribute name and the functional structure is the attribute value. FIG. 11 shows an example of the functional structure.

FIG. 3 is a flowchart of a control device for a natural language sentence analysis device according to the present invention. Below, each step will be explained in order.

"Siri"2; Input the Japanese sentence "solid writing" from the input device.

Qian Chu I: The control device to which the Japanese sentence is input divides the input sentence into a list of morphemes and substitutes them into KL. for example.

If (al) He swims is the input sentence, then KL becomes (a2) (He: n) (Ha: P) (Swim: v). Here, each morpheme is expressed in the format (Headword 2 Opening Range II). "r-opening range" is the so-called part of speech,
In syntax rules, it is represented by lowercase alphanumeric characters.

KL is a list of morphemes, and from now on, we process each morpheme from the beginning of KL, and when there are no more morphemes,
If a valid analysis is obtained as the final state, the analysis is successful. This algorithm uses the Left-to-Right
It is a syntactic analysis method, and is suitable for an interactive natural language sentence analysis device as a real-time syntax analysis algorithm that synchronizes with the user's input of sentences.

Check whether the list (column) of morphemes has ended.

When the analysis is completed, the analysis results are listed in the NSS, and from there, those that are valid as the final state, that is, the analysis results that are composed of syntactic elements corresponding to a sentence, are extracted.

W: Output the analysis result to the output device 3. In the case of syntactic rule R, it is determined that a syntactic element corresponding to a sentence is constructed when a non-terminal symbol S is constructed.

Danko Sat; 5tep If the string of morphemes has not ended in step 3, substitute the first morpheme to W. Since the first morpheme was popped, the length of KL is one less.

Now, the analysis state at the time when the previous morpheme has been processed is listed in SS. Therefore, the analysis is performed using a new morpheme W for each element of SS (an analysis state in which the previous morpheme has been processed).

Check if S is empty.

If it is empty, the new processing for the previous analysis state has been completed, so post-processing is performed.

If it is not empty, let S be the analysis state in which the previous morpheme has been processed.

Hiroyu Sha: Perform shift-reduce processing for S and W. The shift-reduce process in S7 will be described later.

5 step If ss is empty in step 5, post-processing is performed, but at this time, the analysis state that has been processed up to the current morpheme is listed in NSS. Check if NSS is empty. 5tep if not empty
At step 14, NSS is substituted for Ss, and the processing regarding this morpheme is completed.

2: If NSS is empty, it means that there are no more possible parsing states, and it means that parsing has temporarily failed.

Processing in this case will be described later.

FIG. 4 is a flowchart of the shift-reduce process in step 5 in FIG. 3. Below, each step will be explained in order.

This part of the processing is a partial extension of a known shift-reduce parser algorithm. When the shift-reduce process is called, the analysis state at the time when processing up to the previous morpheme is completed is assigned to S, and the current morpheme is assigned to W.

The analysis state S is similar to a general shift reduce parser,
It is represented by a stack of syntactic elements (nonterminal symbols). FIG. 9(,) shows an example of a stack in an analysis state.

5tep21; First, empty is assigned to TSS in order to temporarily list the analysis state that has been processed up to the current morpheme.

Master: Shift with S and W and assign the result to NS. Shifting means pushing a preterminal symbol W onto a stack S of syntax elements.

7; Regarding the newly created analysis state NS, all analysis states created by performing reduce operations zero or more times are listed in the TSS.

Here, the reduce processing in step 23 will be explained in more detail using FIG. 5.

Extract the syntax rules applicable to the prayer state NS from the syntax rules. This process may be performed using a known glue use table, or there may be a method of extracting a rule in which the top of the stack (in this case, the new morpheme W) is the final term on the right side of the syntax rule. The syntax rules thus extracted are listed in the RL.

Trigger: Checks if there are more syntactic rules that can be applied (RL is not empty). If it is empty, this means that the applicable syntax rules have ended, so exit.

d; If it is not empty, pop RL and assign it to R.

When applying the next syntax rule (next element of RL), NS will be needed again, so we copy it to TS so as not to break it.

d; Apply R to the copied TS. More specifically, it involves replacing some syntax elements (non-terminal symbols) at the top of the stack with the left side of the syntax rule that is the right side.

FIG. 9 shows an example of reduction. The stack in the parsed state in FIG. 9(a) has patterns N and P at the leading end, which match the pattern on the right side of (R6) of the syntax rule R.

Therefore, this part is replaced with NP, which is the left side of (R6). The result of reduce in step 5 is assigned to TS.

Step 36: Check whether the newly created analysis state TS can reach the final state. This check can be performed using a known speed table. Here, the speed table is a table configured in advance from syntax rules, and is used to quickly refer to whether a stack can reach its final state based on a stack pattern.

If it is found in step 36 that the final state is about to be reached, TS is pushed to TSS. If you find that it cannot be reached, follow these 5 steps 37
Skip. Now, since TS is in the state of analyzing the result of one reduction, in order to perform further reduction, it is necessary to assign TS to NS and call the reduce process recursively. At this time, if NS is destroyed, it will cause problems in the next processing of the loop, so it must be substituted locally.

When the reduce process is completed in step 23 of FIG. 4, the TSS lists analysis states in which reuse has been completed zero or more times. Penalty calculation and penalty classification will be performed for these analysis states.

Transfer Lλ4: If the TSS is empty, the shift-reduce process ends.

麩肚斗田; if it is not empty in the above 5tep 24, T
Pop SS and assign it to PS. Therefore, PS is the analysis state resulting from shifting and reducing zero or more times for the current morpheme.

畦肚I且: Calculate the PS penalty and substitute it for P. This process will be described later as a penalty calculation process.

4; Next, check whether P is greater than or equal to KILL. KILL is set to 200 here. This check means that an analysis state with a sufficiently large penalty is completely treated as an analysis failure.

Step 28: Next, for analysis states where the penalty is below KILL, it is checked whether P is greater than or equal to TH.

7; If P is greater than or equal to TH, bush Ps to SL.

If P is less than TH, push ps to NSS. TH means the upper limit value of the penalty of the analysis state to be processed at the current time. If P is greater than or equal to TH, it cannot be processed, so it is bushed to SL so that it will not be processed later. Here, SL is a list of analysis states that are excluded from subsequent processing and are, so to speak, "sleeped". On the other hand, if P is smaller than TH, there is no problem with the analysis state, so it is sent to NSS. NSS
is a list of analysis states to be processed later.

Here, the calculation of the penalty in step 526 will be explained in more detail using FIG.

First, the input analysis state (step 26
``C is PS) is set as U. This is because the penalty calculation process is also called from 5tep43, which will be described later.

扛虻danlord: Constructs a functional structure from the analysis state U.

This process can be realized by an algorithm that constructs a functional structure from constituent structures known in word search functional grammars, as long as the analysis state leaves not only a stack of nonterminal symbols but also a history of shift-reduces that have been applied so far.

ste 53 ste 54; Once the functional structure is created, calculate the penalty for the functional structure.

Regarding this process, since the functional structure is a recursive matrix as shown in Figure 11, it is necessary to define a method for calculating partial penalties for the substructures of the functional structure and a method for calculating the penalty by integrating the partial penalties. good.

In the embodiment, the method of integrating the partial penalties is simply adding the partial penalties. Further, calculation of the partial penalty will be explained in detail below using FIG. 7.

Dear Rainbow Caller: The substructure of the functional structure for which a penalty is to be calculated finds the relationship name (function name) in the parent functional structure and substitutes it into REL. Here, the function name is a label in the syntax rules, and can be found by checking the history of reduce or from the function structure if the function structure has already been configured.

4; Next, check all constraints indexed in REL by function name and list them in CL.

Constraints are indexed by function name like constraint C, so CL is easy to configure.

nth day'-; Initialize the variable PEN to which the penalty is assigned. Since CL lists the constraints to be applied, PEN is calculated by applying all of them and adding the penalties of the constraints that are not satisfied.

This process is performed in a loop of 5 steps 64 to 5 steps 68.

The calculated PEN is the penalty for function name REL. Since a function name is a name that indicates a substructure of a functional structure, the above process is performed recursively for all function names in the functional structure, and the penalty for the entire functional structure is calculated by adding the obtained penalty. 6 Regarding the creation of the functional structure and the calculation of penalties as described above, the analysis of the example sentence of (al) is shown in Figures 10, 11, and 1.
This will be explained using Figure 2. FIG. 10 is a parse tree (C3) showing the result of analyzing the morpheme sequence of (C2) by the process shown in FIG. Shift-reduce history equivalent to this parse tree (C3) is shown in FIGS. 12(a) to 12(g).

In FIG. 12, S means a shift operation, that is, the work of incorporating one new morpheme, and R means a reduce operation, that is, the application of a syntactic rule. The number after R is the number in syntax rule R of the applied syntax rule. Figure 12 (g
) reached the final state. The labeled parsing tree shown in FIG. 10 can be uniquely created from the shift-reduce history shown in FIG. 12. The parse tree shown in FIG. 10 is called a constituent structure in the terminology of the practitioner functional grammar described above, and the process of constructing the functional structure (C4) shown in FIG. 11 from this constituent structure is well known. Functional structure (C4) is er+try
There is a list of main words in the section, and the characteristics of the main words are also described. A dictionary was used for background information. Next, a penalty is calculated for the functional structure shown in FIG.

Ignoring features, the functional structure (C4) is a matrix with function names as attribute names. The depth of the matrix is one level, and the function of the outermost structure is only the topic function. Therefore, it is sufficient to calculate a partial penalty for this topic function. REL in FIG. 7, step 61 is t
It becomes opic. Constraints indexed by topic in constraint C are (C4) to (C7), and these are listed in CL. These four constraints are applied to the functional structure (C4), and the sum of the penalties for the unsatisfied constraints becomes the penalty for the topic function.

The functional structure (C4) satisfies the constraints (C4) (C5) (C6). In order to satisfy the 6th order constraint (C7), the functional structure (C4) is transformed into the functional structure (C5) shown in Figure 13. be done.

The difference between (a4) and (a5) is 5ub in (a5)
j function exists and its value matches the topic function. Therefore, in the functional structure (a5), 5ubj
Functional constraints (CO) to (C3) must also be satisfied. Especially when examining (C3),
The semantic attribute feature of the 5ubj function is "human," which matches the semantic attribute described after 5ubj in the lower categorization features in the parent functional structure. That is, (
Since C3) is satisfied, the penalty is O. The same holds true for other constraints, resulting in a functional structure (a
The penalty for 5) is O. On the other hand, in the functional structure (a4), the penalty is 20 because the constraint (C3) is not satisfied.

Finally, the remaining 5 steps in Figure 3 11 to 5 steps 1
4 will be explained.

A new analysis state is listed in NSS by the process explained above, but if the value is not empty, take 5 steps
At step 14, the value of NSS is substituted into the analysis state list SS to process the next morpheme.

If it is empty, the penalty calculated for the functional structure of each analysis state is 5te thicker than TH.
This is because it was listed in SL in p3Q, so the value of TH is recalculated, and more specifically, it is changed to a slightly larger value. This process can be realized by predetermining the initial value of TH and the value to be changed each time 5 tep 12 is passed.

5telL: When a new TH is determined, an emergency process is performed for the SL and the new TH.

扛扝上田; The result is in NSS, so it is NSS in SS.
Substitute , and process the next morpheme.

The last-minute processing will be explained in detail using FIG. 8.

Since the analysis state in the SL is processed, the process ends when the SL becomes empty.

Σmλ,; Pop SL and assign it to L. Processing from step 543 to step 546 is performed regarding the analysis state during SL.

Based on the penalty calculation process already mentioned, L
Calculate the penalty for P and assign the result to P.

1 week main 1-: If P is larger than TH, nothing is done, but if it is smaller than TH, the analysis state that was put to sleep is awakened. This process is to process the morphemes and analysis states from the time when the user fell asleep to the current morpheme, so it is similar to running the loop from step 3 to step 5 in FIG. 3. However, SS is a list consisting only of L as an initial value.

If the result of the process in Step 45 is included in PS, NSS and PS are concatenated and assigned to NSS. As a result, in the NSS, among the analysis states listed in SL, the analysis states that have been "awoke" by changing the penalty threshold to TH and have been processed up to the current morpheme are listed.

It is guaranteed that an analysis state that is once excluded from subsequent analysis due to the penalty exceeding the threshold TH through the last-life processing will be subject to analysis again if there is no other possible solution. Therefore, it is possible to suppress the number of analysis states and prevent spatial and temporal explosions, while at the same time being able to handle exceptional and comparative expressions.

Efficacy - As is clear from the above explanation, according to the present invention, grammatical properties are expressed as constraints, and information regarding the grammatical strength of the constraint is added as a penalty to each constraint,
This information is used to remove ambiguity in the analysis by calculating the penalty for the analysis state. The magnitude of this penalty value is the level of grammatical properties, and enables precise grammatical description and analysis that could not be achieved with conventional techniques.

In addition, in the present invention, syntactic analysis, which was distinguished in the prior art,
Instead of modularizing processing such as semantic analysis, all constraints are expressed as the same constraint, which is called "syntax analysis" in conventional technology.
is uniformly treated as a constraint with a large penalty, and ``semantic analysis'' is treated as a constraint with a small penalty.Also, for analysis states that have a large penalty due to not satisfying some constraints during the analysis process and can be considered not to be a solution, By temporarily removing it from further analysis processing,
Prevent spatial and temporal explosions.

In addition, it is possible to sufficiently remove ambiguity in the analysis, and it is possible to analyze even exceptionally comparative expressions, and it is also possible to analyze spatially,
It is possible to configure a natural language sentence analysis device that does not cause time explosion and has a wide range of applications.

[Brief explanation of the drawing]

FIG. 1 is a block diagram for explaining an embodiment of a natural language sentence analysis device according to the present invention, and FIG. 2 shows a dictionary, syntax rules,
3 is a flowchart of the control device of the natural language sentence analysis device according to the present invention; FIG. 4 is a flowchart of the shift-reduce processing in FIG. 3; and FIG. 5 is a flowchart of the reduce processing. 6 is a flowchart of penalty calculation, FIG. 7 is a flowchart of partial penalty calculation, FIG. 8 is a flowchart of last-minute processing, and FIG. 9 is a diagram showing an example of reduce.
FIG. 10 is a diagram showing a parse tree, FIG. 11 is a diagram showing a functional structure, FIG. 12 is a diagram showing a shift-reduce history, and FIG. 13 is a diagram showing other functional structures. . DESCRIPTION OF SYMBOLS 1... Control device, 2... Input device, 3... Output device, 4... Syntax rule part, 5... Constraint part, 6... Dictionary part. Patent applicant Rico Co., Ltd. Figure 1 Figure 2 (a) Dictionary Figure 6 (b) #1. tltllllR (c) Constraint C

Claims (1)

[Claims] 1. An input device that inputs a natural language sentence, and a dictionary unit that divides the input sentence from the input device into morphemes, and divides the morpheme division into multiple configurations using the contents of the syntax rule unit. a control device that creates an elementary structure, creates a functional structure from each of the component structures, and applies the constraints using the constraints from the constraint section and the dictionary section; 1. A natural language sentence analysis device comprising: an output device for outputting an analysis result by a control device, and calculating incompatibility of the functional structure when the functional structure is created.