NL8802271A - Grammar processor for natural language sentences - Google Patents

Abstract

The grammatical processing is performed by measures applied separately to each word unit of the sentence. The measures applied are as follows:- A. For each word unit and each constituent, its functional word category within the constituent is determined by referring to data concerning the verbal category of the word unit and category of the constituent. B. Finding a step leading to closure of the constituent by reference to data concerning the dominant category of the constituent. C. Testing the current constituent against grammar based rules regarding the context of words/constituents and, if required, assessing the probability factor of the sentence representation.

Description

Method for parsing a phrase, expressed in natural language, into phrases to be described with functional indications, as well as a device for carrying out such a method

The invention relates to a method for parsing a phrase, expressed in natural language, into phrases to be described with functional indications on the basis of word units which have been indexed into verbal categories and application-oriented categories, and to a device for performing such a method.

Such a method, which is referred to in professional circles as "parser", is known from: Allen, James; Natural Language Understanding. The Benjamin / Cummings Publishing Company, Ine., Menlo Park, U.S.A. 1987.

The parsers described there have a so-called "rewriting mechanism", which means that the parsers perform a parsing according to a large number of rewriting rules.

These rewriting rules link a group of words and / or sentence constituents on the one hand, and a parent constituent, that is, a constituent that dominates this group on the other. The number of rewriting rules depends on the size of the description mechanism to be used, which underlies the parser; the descriptive mechanism is in turn determined by the syntax and morphology of the language, which imposes restrictions on the parser in terms of its ability to come up with a solution to the sentence in case of a non-grammatical input. Only by including a very wide range of rewriting rules in the parser is it possible to parse incorrectly formatted as well as less common sentence constructions. As a result, the memory space of the computer on which the parser runs is intended to be used for the rewrite rules, while the time taken to parse such a sentence may be long. In addition, it will be very difficult to detect non-grammatical expressions. The object of the invention is to solve these drawbacks to a great extent.

The invention is based on the insight to bind the parsing to rules ("right associate grammar") in order to fit a word within the existing or newly to be formed constituent in the already obtained phrase representation, while granting a certain functionality at word level. (functional word category) on the basis of at least one verbal category appropriate to that word, and rules ("constituent closure grammar rules") to investigate the possibility of closing a constituent by granting a functionality, whether provisional or not, on constituent level (functional constituent category), during which during the parsing procedure an expectation factor for the sentence is maintained on the basis of an investigation into the relationship between the various words and / or constituents within the same constituent level, on the basis of which the most common obvious sentence representation can be selected. According to the invention, the method of the type described in the preamble is characterized by the following steps to be performed per word unit from the sentence: a. Determining the functional word category per word unit and per constituent within the constituent and / or within a new context. constituent to be created, based on information about the verbal category of the word unit concerned and the category of the constituent; b. Describing for each constituent, on the basis of information about the category of the constituent and the category of the constituent dominating above this constituent, a closure relating to the measure relating to the measure, as well as the granting of a functional label, whether or not provisionally, constituent to be closed; c. Testing the present constituent against syntax-based rules regarding the coherence of words and / or constituents within this constituent in at least one of the latter two steps, and, if necessary, revalue an expectation assigned to the sentence representation! ngs factor, as well as selecting any sentence representation whose expected ngs factor is above a certain threshold.

The parser to be described in a first embodiment is a parser with an external grammar (syntax directed parser) in which the grammatical rules are kept outside the actual program part, and are stored in memory parts or memory modules. This offers the possibility to change the grammatical content outside the program. Furthermore, by only changing the contents of the memory parts or memory modules, the parser can be made suitable for another natural language. The grammatical rules have such a form that it is possible to process a large number of possible structures with a limited number of rules.

Moreover, this parser will generally still come up with a solution for non-grammatical input and even indicate what kind of error has been made. This makes the parser suitable for use within an editor, for example.

In addition to the above-mentioned embodiment of a parser with an external grammar, a parser with an internal or integrated grammar (syntax embedded parser) is of course also possible by accommodating the grammar rules within the parser. This can result in a faster parser.

The invention will be explained in more detail with reference to the annexed figures, of which:

Fig. 1 shows by means of a flow diagram a general overview of the working method to be followed when parsing a text;

Fig. 2 shows a flow chart of the basic principle of the parser;

Fig. 3 shows in a flow diagram a first embodiment of the core portion of the structure shown in FIG. 2 flow diagram shown;

Fig. 4 shows in a flow diagram another embodiment of the system shown in FIG. 3 core section shown;

Fig. 5 shows in a flow diagram yet another embodiment of the system shown in FIG. 3 core section shown;

Fig. 6 shows a detailed flow diagram of a first common portion from the core portion shown in FIGS. 3 to 5;

Fig. 7 shows a detailed flow diagram of a first portion of the structure shown in FIG. 6 common part shown;

Fig. 8 shows a detailed flow diagram of a second portion of the structure shown in FIG. 6 common part shown;

Fig. 9 shows a detailed embodiment of a second common portion from the core portion shown in FIGS. 3 to 5;

Fig. 10 shows a flow diagram of an application-oriented part of the system shown in FIG. 2 flow diagram shown;

Fig. 11 in a flow chart shows a detailed view of part of the structure shown in FIG. 10 flow chart shown, and FIG. 12 in a flow chart shows a modification of a portion of the structure shown in FIG. 7 flow diagram shown.

A parser provides the means to convert a sentence, expressed in a natural language, into a syntactic representation using a formalism, which expresses the relationships between the different entities. That formalism is based on the idea that it is possible to represent the self-contained phrases (constituents) of a sentence through a hierarchical structure, such as a branching network (tree structure).

Strictly speaking, the input of the parser does not comprise a sentence line, but a series of words obtained by lexicalisation of the sentence. As shown in Fig. 1, this means that every sentence presented at program phase 1 must first be lyxized at program phase 2 before the parser parses the sentence at program phase 3. When lexicalizing a sentence, each word is called lexical information related to the possible verbal categories and the application-oriented information of each word by means of a so-called lexical memory, which is part of the computer memory or which is present on a memory disk (disk). categories are searched for and / or generated, and this information is stored by means of one or more word structures in a word memory, which is also part of the computer memory. The results of the lexicalization process of a word unit with respect to the application-oriented categories will be collected below under the term "Features". In the course of the description it will be discussed why there may be more than one word structure, but first the term word structure will be explained in more detail.

A word structure is an assembly of data elements associated with a word, whereby those data elements are assigned a structure with a number of memory fields. The word structure includes the following fields: a Category field, a Command field, a Row field, and a Features field. The Category field contains the word category found in lexical memory, for which the following word indications are eligible, for example: Article ("article"), noun ("noun"), pronoun ("pronoun"), verb (" verb "), preposition (" P-word "), conjunction (" C-word "), adjective (" adjective ") or adverb (" adverb "). This is only an example and depends on the dictionary and grammar table1 and used.

The Command field is initially empty, but will be filled in later by the parser. The meaning of the Command field will be further explained in the description of another structure, namely the constituent structure, in which such a Command field is also present.

The Row field is filled with a "String" representation of the Word, ie a linear arrangement of the elements of the Word. The Attributes field contains the other Word attributes, ie those word attributes, which are used during the lexical isation process. obtained but not included in the Category field The word attributes included in the Characteristics field are of a functional nature.

The above will be explained on the basis of two examples: The word "house" in Category-field is filled in: "noun (noun)" (noun); in the Command sheet d: "NIL"; in the Row field: "house" and in the Characteristics field: "3rd person singular" (sing 3) and "neuter" (neuter).

As a result of such a representation of a word structure, a word such as "regent" has a Category field with the entry "verb" (verb) as well as a Category field with the entry "noun" (noun) with bij each Category field obtains its own Attributes field.

Since a Word can thus occur in several word categories, and furthermore as a result of ambiguous relations between Constituents, multiple sentence structures between these phrases are conceivable, this approach leads to multiple forms of tree structures, and thus to multiple syntactic representations of the parsed sentence. The parser will always try to include every word structure of the current word within every open constituent structure. This means that the parsing speed decreases when many representations are present.

Since in hindsight not all representations appear to be suitable, it is the object of the invention to investigate how, in the interim, that is to say, during the active phase of the parser, one or more less attractive forms of representation of phrase structures can be eliminated. turn into. The starting point is that after each phase in which a word is added to the existing representation form, the Constituent involved is subjected to certain syntactic rules, and on the basis of certain error detections the initial probability factor, which is assigned the Constituent, according to a system of respective correction factors is lowered, and then the corrected probability factor is tested against a certain threshold value. All forms of Constituent representation with a probability factor that is less than that threshold can then be eliminated as "less likely". This procedure is then repeated to include the next word within the representation structure.

Such an action can take place immediately after the inclusion of a word in an existing or new constituent, but can also take place after a subsequent investigation into the conclusion of that constituent. With the parser to be described here, such a filtering action will take place in both situations.

Although the lexicalization as described here takes place before the actual decomposition so that the parsing process starts only after all words have been lexicalized, it is also possible to lexicalize one word at a time and then parse that word, lexicalize the next word - if possible in the meantime - and then parsing etc. An advantage of such an approach could be that it is then possible to start parsing a sentence while a user is still typing the sentence (real time parsing).

However, the flow diagram to be described below assumes a lexicalization that precedes the actual parsing process. The parser is then called after the lexical isation. If it succeeds (Y) in finding an analysis in step 4, this analysis can be used in the rest of the program (step 5). This can be a tree drawing program, in which a graphical representation of the found decomposition is drawn, an editor, an indexing and retrieval system, etc., after which the next sentence can be continued in step 1. However, step 4 does not lead to a usable result (N), the parser failed to find a good analysis. This will mean that the sentence was very ungrammatical. An action can be taken for this, after which the next sentence can be continued.

Before starting with the actual description of the parser (see step 3), it is important to describe another structure, namely the Constituent structure. The Constituent Structure is a phrase-related assembly of data items, which includes the following memory fields: Category field, Category field, Command field, Members field, Features field d, Stack field, Error reporting field (Yiolations field) and a Probability field.

The Category field is reserved for information about the category of the constituent, which in our example grammar should include the following phrases: main, sub- or clause (Sentence with the abbreviation S), the nominal phrase or phrase (nounphrase with the abbreviation NP), the preposition group or phrase (Prepositional phrase with the abbreviation PP) and the adjective / adverbial phrase (adjective / adverbial phrase with the abbreviation AP). The Category field initially contains the designation NIL, but will be filled in with the information about the constituent during the parsing process. The Command field is the field that contains, after parsing, address information (pointer) about the Parent constituent, i.e., the Constituent which dominates the present constituent.

The Structural Elements field is reserved for information about one or more pairs of the functional label that occur in pairs with a Word or Constituent structure belonging to this constituent, as well as the address information about the relevant Word or Constituent structure. In general, a Constituent can be composed of a number of Word structures and / or Constituent structures.

The functional labels used here in the sample grammar include the following: - Subj: Subject (Subject) - Indobj: Contributing Object (Indirect Object) - Obj: Direct Object (Object) - Smod: Sentence Modification (Sentence Modification) also known as a "further adverbial provision" (Adverbia!

Modi fi cation) - Comp: Complement - Pred: Predicate (Predicate) - fNP: functional NP, to be used as a provisional label, which will later be replaced by a definitive functional label at sentence level - Nmod-a: a modifying provision for a independent

(used) noun in the form of an AP

- Nmod-s: a modifying provision for a self-employed person

(used) noun in the form of an S

- Nmod-p: a modifying provision for a self-employed person

(used) noun in the form of a PP

- Oet: a determining pronoun or article (determiner)

- Head: central element within an NP, PP or AP

- Pobj: a section, connecting to a preposition

sated by a POI

- Amod: a modification of the main word of an AP

- Seq: a connecting word with a juxtaposition between two phrases or sentences (sequencer) - Conj: one of the subordinate phrases.

The Characteristics field is reserved for the information about the characteristics belonging to the Constituent. As soon as a Constituent is created with a certain Category, the Attribute Sheet d will be filled with the attributes belonging to the Category. In the case of a Word category, the Attributes field is provided with the attributes associated with that Word, which have been found during the lexialization process. During the parsing process, for each element (word or Constituent) that must be included in the relevant Constituent, the Characteristic field of the relevant Constituent will be adapted to the information associated with that element. The Stapelveld contains a list of the functional labels, which are also mentioned in the Structural Elements field of the Constituent. The Error Reporting field contains information about all error codes related to incorrect grammar made within the Constituent. This field is initially empty.

The Probability field indicates the probability factor attributed to the occurrence of this Constituent. Initially this is "1", but for each detected incorrect gram materiality, this value will decrease by a value which depends on the detected type of such incorrect granularity.

With the help of the Constituents it is possible to compose a parser tree. After all, every Constituent can in turn contain Constituents and / or Words. In addition to the functional category, the Structural Element field also indicates the memory address information associated with the Constituent or Word structures. On the basis of this information it is possible to descend into the parsing tree. Conversely, the Command field of a Constituent provides address information about its Parent constituent, i.e. the other Constituent dominating Constituent, which provides the opportunity to get higher in the parsing tree. It is noted that Words cannot contain structural elements, while the Command field of the highest Constituent (root) in the parsing tree contains no information (NIL).

The parser, as can already be seen to some extent from the Probability field, works with probability factors and has a threshold value that can be modified if necessary, against which all available structures are tested for their probability factor, and where only structures with a probability factor greater than or equal to the threshold value, to be processed by the parser. This makes it possible to perform the decomposition of grammatical and post-grammatical sentences more quickly.

In the following section, the operation of the parser will be explained in more detail, it is important to note here that the parser is based on the understanding that a Word of a certain Category within a Constituent of a certain Category often occurs will have a clearly defined functional meaning within the Constituent. The same goes for Constituents within Constituents. The parsing method is therefore based on the fact that it will usually be sufficient to know the Categories of the Parent Constituent on the one hand, and those of the Word or Constituent on the other in order to know the function (functional category) of that Word or Constituent within determine the Parent Constituent.

A program associated with the parser is shown in the following figures by means of a flow diagram, according to which the method of parsing a text in a computer is carried out.

The entire parser will first be described in a fairly global manner. Two program parts, the so-called "fitting a word in an existing or new Constituent" ("word right associate") and the "closing a Constituent" ("closure") will only be indicated briefly, after which In a subsequent part of the description, these program parts are explained in more detail with reference to detailed figures.

The method to be described can be divided into two phases, in which in the first phase, which in fact comprises the actual parsing process, the possible sentence representations are formed by means of a syntactic analysis, and in the second (additive) phase definitive functional labels on constituents with still have an unclear functional character. Appropriate filter procedures will be discussed in both phases.

This method is carried out word by word (with all its functions of use) and will be repeated each time when a next word is to be included within the representation structure.

In the flow diagram of Fig. 1 commences the reading phase of a sentence in order to be able to carry out the parser's method with reference numeral 1. In this phase, the sentence to be parsed is presented to the computer in a known manner (for example by means of a keyboard), and the sentence word for word written in a memory part to be designated as word memory in the computer, wherein the number of words (kmax) of which the sentence is composed is also calculated. The guideline here is that the words are separated by spaces, tabulations and punctuation marks, and that each punctuation mark is included in a word category. The stated number of kmax is also kept in the word memory.

Then the next phase (2), the lexical description of all words from the offered sentence, takes place. The word structures present per word are searched for by the words of the sentence offered, which is done on the basis of a memory part, which can be designated as lexical memory, which is present in the comouteraeheuaen, and in which a large collection with one or more word structures possibly occurring is stored.

A detailed explanation of phase 2 will be described with reference to FIG. 2 occur. The lexical description takes place in the next program step 6. A counting unit (k-telier), which indicates the rank number of the word to be treated (k), is reset to 0.

Then, in step 7, the count of the k-counter is increased by 1 (k = k + 1), after which in step 8 it is examined whether the count of the k-counter has already exceeded the maximum k-value (kmax) (k > kmax). If this is the case (Y), the program proceeds to step 11, if not (N), the program proceeds to step 9, whereby the next word from the word memory is assigned an address determined by the count of the k-counter. is retrieved and written to a memory section to be designated as word memory in the computer. Furthermore, in step 10, the word written in the word memory is searched for in the lexical memory and the specific information occurring there is added to the word in the word memory.

Then the program returns to step 7. Since in parsing each word must be examined for all listed word functions (word categories), each word with a deviating word category and / or other deviating characteristics will have its own rank number (1) in the corresponding Word structure while the number of assigned rank numbers (lmaXj k) Per w ° ord (k) will also be kept. Instead of (1 rnax, k ^ * ian ° ° k suffice to assign a specific label to the last array to indicate in the syntactic analysis below that the last word function of that word has been reached.

During the syntactic analysis to be carried out in phase 3, multiple representations with respective constituent structures may simultaneously appear as possible solutions, which have different probability factors due to conflict with certain grammar rules. Since one is only interested in the solution or series of solutions with the highest probability factor, solutions with a lower probability factor should be classified as an uninteresting solution and should therefore be eliminated. This is done by testing the probability factors associated with these solutions on the basis of a certain threshold value. This threshold value is set to the value "1" at the start of the syntactic analysis, but can be adjusted to a lower value during that analysis, if it appears that none of the probability factors obtained meet the threshold value. Accordingly, the program includes a step 11, wherein the initial threshold value is set to "1".

Then, at step 12, an initial constituent for the sentence is made. This is a Constituent, the initial stages of which include the following information:

Category field: S

Command field: NIL

Structural element field: MIL

Characteristic field: MIL

Stacking field: MIL

Error message field: MIL

Probability field: 1

In order to fill in the Characteristics field, a memory section to be designated as the first tabular memory (see table A) looks at the Characteristics indicated under category S. Since no Characteristics are indicated in this preprogramme, the corresponding field remains empty (NIL). The memory also comprises a so-called Representation memory for recording current Constituents (for example in list form as is usual with the programming language LISP), with which the parser will always work. In this Representation Memory, which has hitherto been empty, the initial constituent S is written.

Then, at step 13, the k-counter is reset to the initial value "0", and at step 14, the count of the k-counter is increased by "1". After step 14, the program arrives at step 15, which examines whether all words (including the punctuation marks) in the sentence have already been involved in the syntactic analysis and whether the count of the k-counter therefore applies: k> kmax. If this is the case (Y) then apparently every Word is included in the Constituents and the syntactic analysis should be considered as completed.

If in step 15 the statement k> kmax is not true (N), the program proceeds to step 15. The fitting of the word indicated with the aid of the k-counter in the Constituent structures available at that time will start, in the not yet completed Constituent structure, which is located in the aforementioned representation memory.

All Word structures (1) of this word are now available as a result of the exicalization process. When fitting a Word with a certain word structure within a Constituent with a certain Category, the Word will in many cases fulfill a certain function, and thus a certain functional Word Category can be assigned. The same will also apply to Constituents within their Parent Constituent at a later stage. On the basis of the Constituent Category on the one hand and the Word Category on the other hand, the function (functional Category) of the Word within the Constituent will be determined. If the described incorporation of the Word into the aforementioned Constituent is not directly possible, it is sometimes possible that a new Constituent can be linked to it, within which the rapture of the Word is possible. It should also be noted that the inclusion of a Word structure in an open Constituent structure on the basis of a number of grammar rules and because of the number of Word structures (1 max, k) in that Word (k) can result in a certain multiplication of the number of Constituent structures. . Since the results of the parsing process will always be included in the Representation memory, and the current Constituent structures occurring therein must be taken as input values for the parsing process, in step 17 a memory part of the computer, to be designated as "Temporary Memory", is first empty, the Representation Memory with all its Constituent Structures is then transferred to the Temporary Memory in step 18, and then the Representation Memory is empty in step 19. Several variants of the following program section 20 are possible. These will be explained successively with reference to Figures 3, 4 and 5.

Assuming that with a Word (k) with several Word structures (1) the connection possibility must be checked for each structure, the computer comprises a counting unit (so-called 1-counter), the counting position of which corresponds each time to the relevant word structure.

For a next word (k + 1), which must be included in the existing phrase structure (Sm), the 1-to-1 Irish must also be reset to 0, which is also the case with the counting unit (so-called m-counter ), which represents the rank numbers (m) of the constituents.

Accordingly, at step 21 in FIG. 3 the count position of the 1-counter is reset to 0 (1 = 0), after which in step 22 the count position of the 1-counter is increased by "1" (1 = 1 + 1).

In step 23 the question is asked whether the count of the 1-te er has already exceeded the maximum value (lmax ^), which represents the number of Word structures of the word with rank number k (1> ^ max)

If the question is answered positively (Y), the program proceeds to step 24.

However, if the question in step 23 is answered in a negative sense (N), then in the next step 25 the 1st Word structure will be brought from the Word memory to the Working memory to transfer the previous Word structure, in order to transfer to the newly registered Word structure perform a number of operations. Then, in step 26, the m-counter will be reset to 0 (m = Q), and then in step 27, the count of the m-counter is increased by "1" (m = m + 1). Then, at step 28, the question is asked whether the count of the m-counter has exceeded the number ι% 3Χ, which represents the number of constituents occurring at that time (m> mmax). If this is the case (Y), then it has been established that the word (k) to be fitted with the conscious structure (1) has been investigated in connection with all Constituent structures present in the Temporary Memory, and that therefore this part of the program must be repeated for the same word (k) but then for a subsequent Word structure (1 + 1). Accordingly, the program jumps back, and returns to step 22. If the question is answered negatively (N) in step 28, step 29 follows, where the mth Constituent is retrieved from Temporary Memory and sent to Working Memory. to transfer the previous Constituent. Thus, both the 1-th Word structure of the k-th Word and the m-th Constituent are in the Working Memory. With these two structures (Word structure and Constituent structure) a number of operations are performed in the program section 30.

In fact, an attempt is then made to place the Word structure within the Constituent structure, possibly by creating other Constituent structures. The program section is further shown in FIG. 6 elaborated. After execution of the program section 30, the program returns to step 27.

If in step 28 the question is answered in the affirmative (Y), the program returns to step 22. This means that the process is to fit a word right (word right associate-process) for one word structure (1) at one word (k ) performed on all actual constituents (% 3χ).

If at some point after the question is answered in the affirmative at step 23 (Y), the program proceeds to step 24. At that moment the process is to "insert a word to the right" for all word categories (lmax) at a word ( k) performed on all actual constituents (mmax), and the constituents thus formed are stored in the Representation memory. All these constituents now need to be subjected to a "closing constituents" program section. To this end, the Temporary Memory will first be emptied at step 24 and all the constituents stored in the Representation Memory will be transferred to the temporary Memory at the next step. Then, at Step 32, the Representation Memory is cleared. In the subsequent step 33, an m * counter, the count of which corresponds to a particular constituent in the Temporary Memory, is reset to 0 (m * = 0). In the next step 34, the count position of the m * counter is increased by "1" (m * = m * + l). Then, in step 35, the question is asked whether the m * counter has already exceeded the maximum value (m * max) corresponding to the number of constituents (m *> m * max). If this question is answered in the affirmative (Y), then all representations (m * max) obtained in all word categories (lmax) of the kth word have been subject to the program section for "closing constituents". The program then returns to step 14 to make representations using the next word (k + 1).

If, however, the question is answered in the negative in step 35 (N), then in the next step 36 the m * -th representation is taken from the Contemporary Memory and then in step 37 to the previously mentioned program section for "closing constituents". "subject. The result obtained during this step is written in the Representation Pleasure, after which the program returns to step 34.

At step 15 in FIG. 2 the question is answered in the affirmative (Y), then all words (kmax) with all associated Word structures Omax k) are accommodated in the constituents, and in the next step 38 the question is asked whether at least one Constituent exists in the Representation memory, of which the Command field is empty. After all, an empty Command field indicates that we are dealing with the upper Constituent or Top constituent (Root). This question is real, because in steps 30 and 37 a filtering procedure has been built in, which filters out (eliminates) faulty representations, and this by means of a threshold circuit in which the value of the Probability field is compared with the current threshold value. This comparative study will be further explained in Figures 6 to 9.

If there is a negative answer (N) to the question in step 38, it must be ascertained whether the program section thus described about the parser will yield a useful result, i.e. a meaningful representation, at a lower threshold value. However, the threshold value should not be brought below a certain value, for example 0.2, since it is then considered that the sentence presented is structurally and grammatically wrong in such a way that no useful representation from the parsing process should be expected. Accordingly, if the question is answered negatively (N) at step 38 during the next step 39, the question will be asked whether the threshold value is still above a predetermined minimum value, for example 0.2. In the affirmative answer (Y), in the next step 40, the threshold value will be set to a next lower number value according to a fixed series of number values. The program will then return to step 12 to rerun the program with regard to parsing the sentence presented, but at a lower threshold value. This also allows grammatically incorrect sentences. If the question in question 39 is answered in the negative (N), it must therefore be assumed 'F

If no useful analysis can be found, a corresponding action will have to be taken. The sentence can possibly be classified as suspicious, with presumably many and / or very serious grammatical errors. If necessary, a notification can also be made. It is allowed to let step 40 precede step 39, wherein step 39 then takes another value as a criterion. In the affirmative answer (Y) to the question of 'step. 38, and thus in the presence of at least one representation of Top Constituent, in the next step 42 the structural elements of the Top constituents present are adjusted. This is based on the situation that the sentence associated with the sentence has been created. However, some Command Fields have yet to be completed or edited. The tents (Parent constituents) in the Representation Memory will contain in their structural element many ÖjëenSö number Constituents (Member constituents) and / or Words '-vy · .j.; Yv (Memberwords). A member constituent is a constituent' which. "is dominated by its parent constituent. The Command Sheet ^" These Member Constituents and Member Words are now filled with · -an designation or reference to the appropriate Parent Constituents, thus becoming the Command Fields of all Member Constituents. yen: t | ^ Member words, which are mentioned in the Structural element of a particular Parent constituent, also provided with an indication or reference concerning this Parent constituent. The same also happens for 1 Command field. of the Constituents and the Words within a Constituent, whose Command Field has been completed during the parsing process.

Since temporary functional labels may have been assigned during the parsing process, these are replaced in step 43 by definitive ones. In the example grammar yet to be described, the preliminary fNP is used for each Constituent with Category NP thunder; i another Constituent with Category S. During this step, such labels are replaced by the final labels "Subject" "(Subject)," Suffering Object "(Object)," Contributing Object "; (Indobj.) And "Predicate" (Pred). This step 44 will be at a later stage; further explained. Finally, at Step 44, from the "Reproduction" memory, that Representation is selected, which has the highest probability factor. This can also result in several

Representations. For example, in the sentence "I saw a girl with the binoculars." two solutions are possible, since the phrase "with the binoculars" can be a more detailed definition of the "verb" as well as of the "Suffering Object".

A second embodiment of the program section 20 will now also be described with reference to FIG. 3 are explained. In addition, the index 1 now refers to a certain constituent structure (and not to a certain Word structure), while thereby the index m refers to a certain word structure. As a running index at steps 33 to 37, 1 * resp. l * max i-p-v · m * and m * max are used. Accordingly, at step 21, the count of the 1-counter with respect to the rank number of the constituent to be used is reset to "0" (1 = 0), at step 22, the count of the 1-counter with "1" (1 = 1 + 1), and in step 23 the question will be asked whether the count of the 1-boiler has already exceeded the maximum value Umax), corresponding to the number of constituents present (1> 1max). In step 25, the constituent indicated by the count of the 1-counter is retrieved from the temporal memory, which constituent will overwrite the previous constituent in the working memory. Furthermore, at step 26, the m-counter, the count of which corresponds to the rank number of a certain Word category, is reset to 0 (m = 0). In step 27 the count position of the m counter is increased by "1" (m = m + 1) and in step 28 the question is asked whether the count position of the m counter has the maximum value% 3Χ, corresponding to the number of word structures, has already exceeded (m> ί% άΧ). At step 29, the mth Word is retrieved from the Word memory and the mth Word will overwrite the previous word in the working memory. The m-th Word and the 1-th Constituent are then subjected to the adjustment process of step 30.

An affirmative answer (Y) to the question at step 28 implies that all word structures (ι% 3Χ) at one word (k) after connection to one Constituent (1) in the program section 20 have been examined. The other steps proceed as described above.

In the embodiments already described, all possible representations of constituents are first created or expanded, which one word (k) may be for all word structures on the basis of available constituents, before the program section 37 for "closing these constituents" is executed.

In the following embodiment, all possible representations (m * max) of constituents are first created or expanded, which are possible with one Word (k) for one Word structure (1) on the basis of all constituents (mmax) before the program part is completed. for closing these constituents (m *) is performed. Then the same part of the program is repeated, but for the next Word structure (1 +1).

The third embodiment based on this train of thought is shown in FIG. 4 is shown.

At step 21, the 1-count concerning the rank number of a Word structure at Word (k) is reset to "0" (1 = 0), and at step 22, the count of the 1-count is raised by "1" ( 1 = 1 +1). Then, in step 23, the question is asked whether the count of the 1-counter has already overwritten the maximum value dmax) (1> 1max). In step 25, the word with the Word structure defined by the 1-counter is taken from the Word memory and placed in the Working memory. Then, in step 26, the m-counter whose count position represents the rank number of a constituent in the Temporary Memory is reset to "0" (m = 0), and in step 27 the count position of the m-counter is reset by " 1 "raised (m = m + 1). In step 28 the question is asked whether all constituents (fnmax) have already been examined for connection by the program part of step 30 (m> mmax). If the question is answered in the negative (NI), in step 29 the constituent defined by the count of the m-counter is taken out of the Temporary Memory and placed in the Working Memory, after which in step 30 the connection of the 1-th Word structure on the mth Constituent structure is examined, and the result thereof is written in Representation memory.

In the affirmative answer to the question in step 28, in the next step 45 the contents of the Representation memory and that of the Temporary Switch memory are exchanged with each other, after which in step 32 the Representation memory is cleared. Next, at step 33, the m * counter, whose count represents the rank number of a particular Constituent in the Temporary Switch memory, is reset to "0" (m * = 0). In step 34 the count position of the m * counter is increased by "1" (m * = m * + 1), and in step 35 the question is asked whether the count position of the m * counter has the maximum value ( m * max) has already exceeded (m *> m * max). If this is not the case (NI), the constituent indicated by the m * -teer is removed from the Temporary Switch memory at step 36 and subjected to the program part for closing Constituent at step 37. The result is then stored in the Represents memory, and the program goes back to step 34. If the question is answered affirmatively (Y) in step 35, in the next step 47 the contents of the Representation memory are copied to a buffer memory and step 48 the Representation memory is cleared.

The program then goes back to step 22. The program part thus executed is then repeated, but then for the next Word structure (1 + 1) at the kth word.

If the question in question 23 is answered in the affirmative (Y), the program goes to step 46, in which the buffer memory is copied to the Representation memory and then the buffer memory is emptied. The program then goes back to step 14 (see Fig. 2).

Instead of placing the result of the process steps 37 and 30 first in the Representative memory and then in the Buffer memory, it is also possible to place the result directly in the Buffer memory. However, this will have the necessary consequences when dealing with these program parts in Fig. 6 and 9.

A fourth embodiment, which similarly differs from the above embodiment as the second embodiment from the first embodiment, will now be described in more detail. The indices 1 and m refer to a rank number with respect to a Constituent and a Word, respectively. The indication 1 * will be used as the running index in steps 33 to 37.

Accordingly, at step 21, the count of the 1 counter is reset to "0" with reference to the Constituent to be used, (1 = 0), at step 22, the count of the 1-counter is reset to "1" increased (1 = 1 + 1), and in step 23 the question will be asked whether the count of the 1-grower has already exceeded the maximum value (lmax), corresponding to the number of constituents present (1> lmax) In step 25, the Constituent indicated by the count position of the 1-count Irish is retrieved from the Temporary Memory and placed in the Working Memory. Furthermore, at step 26, the m-counter, the count of which corresponds to the rank number of a certain Word structure, is reset to "0" (m = 0). In step 27 the counter position of the m counter is increased by "1" (m = m + 1) and in step 28 the question is answered whether the count position of the m counter has the maximum value mmax> corresponding to the number of word structures , has already exceeded (m> ιτ ^ χ).

At step 29, the mth word structure is retrieved from the Word memory and placed in the Working memory. This m-th Word structure and the 1-th Constituent are then subjected to the incorporation process of step 30, the result being stored in Representation memory, and the program returning to step 27. Thus, an affirmative answer (Y) to the question keeps track step 28, that all Word structures (mmax) at one word (k) after connection to one constituent (1) in the program section 30 have been examined; then, at step 45, the contents of the Representation Memory are transferred to the Temporary Switch memory, and the Representation Memory is cleared. Subsequently, a phase begins at step 33, which leads to the closing process of the Constituents present at step 37. If the question is answered in the affirmative at step 35 (Y), the program proceeds to step 47, after which the program part thus described is repeated, but then for the next Constituent (1 + 1) in the Temporary Memory. If the question in question 23 is answered in the affirmative (Y), the program returns via step 46 to step 14 (see Fig. 2).

In FIG. 5 shows a fifth embodiment of the program part 20, which is based on the idea that each time after one Word and one Constituent have been transferred to the Working Memory, both the program part 30 regarding the fitting of that Word to that Constituent, as well as immediately after the program part 37 to take place regarding the closing of the obtained Constituents. Accordingly, at step 21, the 1-count 1 er, whose count position represents the rank number of the Word structure, is reset to "0" (1 = 0), at step 22, the count position of the T-count 1 is reset with "1" "increased (1 = 1 + 1), and in step 23 the question is asked whether the count with the 1-count 7er already has the maximum value (lmax), which corresponds to the number of Word structures with the kth word exceeded. If this is not the case (Ni), then in step 25 the 1 st Word structure indicated by the 1-counter is taken from the Word memory and written into the Working memory. Then, in step 26, the m-counter, whose count position corresponds to the rank number of the constituent in the Temporary Memory, is reset to "O" (m = O), and in the next step 27 the count position of m- counter increased by "1" (m = m + 1). At step 28, the question is asked whether the count of the m-counter has already exceeded the maximum value (mmax), corresponding to the number of constituents in the Temporary Memory (m> ι% 3Χ). If this is not the case (N), in the next step 29 the mth Constituent is written from the Temporary Memory to the Working Memory. Then, the program portion 30 regarding the connection of the 1 st Word structure to the m th Constituent structure is executed and included in the Representation Memory. Immediately thereafter, the replenished (or new) constituent (s) is retrieved and removed from the Representation Memory and then subjected to the program portion 37 regarding its termination. During step 37, the result is written into Representation memory, after which the program returns to step 27.

If the question of step 28 is answered in the affirmative (Y), the program returns to step 22.

If the question is answered in the affirmative at step 23 (Y}, the program returns to step 14 (see Fig. 2).

In a sixth embodiment, indices 1 and m are swapped compared to those in the fifth embodiment.

Therefore, steps 21, 22, 23 and 25 relate to the Constituent representation with index 1 and steps 26, 27, 28 and 29 to the Word structure with index m.

Using the flow diagram of Fig. 6, the program section will now be explained in too much detail.

By means of this program part an attempt is made to place the offered Word structure in the offered Constituent Structure. Although a punctuation such as the "comma" and the conjunctions "and", "as well as" and "or" are considered as Words, and as such are also stored in lexicological memory, they still occupy a separate position with respect to other Words in the parsing process; they can be treated as a Member element within a new Parent Constituent to be created of the respective Constituent, within which the process of fitting those conjunctions or punctuation should actually take place. For this reason, such conjunctions and said punctuation will follow a different path within the word fitting process. Hence, at the beginning of this program section at step 50, the question is asked whether the Word category of the offered Word is indicated with "Seq". If this is the case (Y), the program goes to step 51. If this is not the case (N), in the next step 52 the Word structure as well as the Constituent structure is written to the Working memory. Then, at step 53, on the basis of the information about the Category field of the Word and that of the offered Constituent, a search is carried out in a memory section to be designated as second tabular memory on grammar rules. For example, the word category "article" and the offered Constituent "NP" have the grammatical rule "article (NP (det))" in the indicated preproduction program, which means that the Word is incorporated into the functional label "det" in the offered Constituent Several grammar rules of this type can be found in attached table B.

The next step 54 asks whether the requested grammar rule exists and therefore has a data field. If the data field is empty (Y), and the grammar line therefore does not provide a functional Word label, then the program section 30 regarding the connection of a Word within a Constituent is considered to be finished. This is the case, for example, when the connection of a verb (verb) in an NP needs to be investigated: this does not offer a solution (label length = 0). In the negative answer (N) at step 54, the next step 55 asks whether the grammar line has a data field with only one functional label. If this is the case (Y), as with the line "article (NP (det))", then in the next program part 56 the offered word in this case "article" is included within the current constituent structure.

This program section will be explained in more detail with reference to FIG. 7 are explained. However, if a negative answer is given in step 55 (N) and thus multiple elements occur in the data field, as is the case with the grammar rule "article (PP (det NP))", then in the next step 57 last element in this case NP created a new Constituent, which is inserted between the Parent constituent PP and the Word "article".

This new Constituent thus has the following structure: Category field: the last element of the grammar rule in this case NP. Mission field: Parent constituent in this case PP.

Structural Element Field: NIL, that is, the field contains no information. ·

Characteristic field: All the characteristics belonging to this Constituent Category (NP), which can be derived from the first tabular memory. In this case: "neuter, inneuter, sing 1, sing 2, sing 3, plu I, plu 2, plu 3, definite, indefinite". Several examples of this type are included in Table A.

Stacking field: NIL, ie the field contains no information.

Error message field: NIL, that is, the field contains no information.

Probability field: 1.

This new constituent structure is now considered to be the current constituent structure.

In the next step 58, the last element i.c. NP is removed from the grammar line, and the program returns to step 55.

Since in the present case only one element remains in the grammar line, the program will now proceed to step 56. Step 56 is performed with the current word structure and the current constituent structure. It should be noted that what is considered to be the current Constituent structure may have been changed at step 57.

At step 51, the offered Constituent (within which steps 25 to 30 should fit a Word) is copied. In the next step 59 it is checked whether the copied Constituent already has an element "Conj" within the Stack field of the Constituent. If this is the case (Y), it is not advisable to enter a new Constituent again with the prefix "Conj" as Parent Constituent for the copied Constituent and the Word or Punctuation to be inserted. The process then proceeds to step 60, where the copied Constituent is placed in Representation Memory pending subsequent steps 61 to 66, which will involve this Constituent. If the question is answered in the negative (N) at step 59, then step 67 follows, where it is checked whether this Constituent currently already fulfills the conditions for closing a Constituent. A first condition is that the last element added to the Stacking Field is also a suitable element on which the Constituent is allowed. end. A second condition is that certain elements must at least be present in the Stack field of the Constituent.

The first condition can also be replaced by another condition, which indicates on which elements a Constituent may not end.

The first, replacement condition and the second condition are examined by means of grammar rules, which are stored in a memory section to be designated as third tabular memory. For example, the information (NP (det Nmod-a) head) indicates that in a NP the last element may not be a "determiner" or "Nmod-a", and that there must in any case be a "head". Several examples are included in Table C. If such conditions are not met by a POI, the value of the Probability field is decreased, and a corresponding error code is entered in the Error Reporting field; included. If these conditions are met, no further action takes place at this step. By lowering the probability factor, it is likely that this constituent will be eliminated in a subsequent filtering procedure (in program section 69).

In the next step 63, for the inclusion of a "comma" or the juxtaposed conjunctions next to the copied Constituent, a new Parent Constituent is generated, "which gets the same Category designation as the copied Constituent, but with the prefix" Conj "included. The Command field of the new Constituent becomes identical to that of the copied Constituent. The remaining fields are completed in the same manner as was the case in step 57.

In the next step 69, a proposal is made to fit the copied constituent within the newly formed constituent with the functional label "Conj". This is done using the functional label "Conj", the copied Constituent structure itself and the newly created Parent constituent. The associated program portion will be explained in more detail with steps 81 to 89, and may lead to the addition of a new Constituent Structure to the Representation Memory.

In the next step 70 it is checked whether a Constituent structure has indeed been added to the Representation memory. If not (N), and therefore the copy constituent is eliminated, the sequencer insertion process is considered terminated. If, on the other hand, the question is answered affirmatively at step 70, then step 61 follows, in which the last added constituent is read from the Representation memory and moreover is removed from it. The next step 62 follows an identical program as that which took place in step 56, however, in the stack field the functional label "seq" is represented. Since an element may have been added to the Representation memory at step 62, the question is asked at step 63 whether this is indeed the case. In the case of a negative answer (N), this program part is considered to have ended. In the affirmative answer (Y), the last added element is extracted from the Representation memory in step 64 and a new Constituent structure is then created in step 65 in a manner already described in step 57. The assignment field is filled in with a reference to the Parent constituent. As the Parent Constituent, the Constituent retrieved from Representation Memory in step 64 is now considered to be the current Constituent.

The category field gets the same value as the category field of the member constituent that has the functional label "conj" within the current constituent.

The other fields are filled in as was the case in step 57.

In the next step 66, the newly created constituent structure is added to the Representation Memory, after which the word fit program within a constituent is considered terminated.

Fig. 7 depicts a flow chart of a detailed program referenced during program sections 56 and 62.

At step 71, a copy of the current Constituent Structure, that is, the structure of the constituent, contains the relevant

Word must be lodged, written in working memory.

At the next step 72, the Functional Label, which at steps 53 and 56, resp. 62 is obtained, added. Then, in step 73, a combination of structural elements, namely the functional label and the Word structure to be fitted, is added as a label pair to the Structural element field. In practice, the address of the Word Structure will be added instead of the Word Structure itself.

Then, in step 74, in a memory section to be designated as the fourth tabular memory, the corresponding set of features is selected for the category of the current Constituent and the functional label of the Word. Any element from this set of Attributes, which does not appear on the Attributes field of the Word Structure, and which does occur on the Attributes field of Constituent Structure, is removed from the Attribute Sheet d of the Constituent Structure.

To illustrate what has been mentioned above, the following example is given: for "NP" and "determiner" according to the fourth tabular memory (see table D) the series of characteristics belongs: "neuter, inneuter, sing 1, sing 2, sing 3 , plu 1, plu 2, plu 3, definite, indefinite ". Suppose that the Constituent structure with category "NP" has the characteristics sing 1, sing 2, sing 3, plu 1, plu 2, plu 3, neuter, inneuter. Since now the Word structure in the Attribute sheet d has the characteristics "neuter, definite , sing 3 "(as is the case for the article" het "), the features sing 1, sing 2, plu 1, plu 2, plu 3 and neuter must be removed from the Feature sheet d of the Constituent structure.

In the next step 75, if desired, some control operations are carried out, which is done on the basis of a memory part to be designated as fifth table memory (see table E). In addition, for each constituent type (constituent category), incorrect combinations, incorrect order and the number of functional labels in the Stapelveld are examined. For example, the occurrence in a constituent NP of two functional labels "head" can be interpreted consecutively as an incorrect combination, the functional labels "Nmod-S" and then a "head" as a combination with an incorrect order, and more than one functional label "determiner" in an NP as an incorrect number of functional labels for such a constituent. Furthermore, an incorrect order of two functional labels is present when a "head" is followed by an "Nmod-A". It must be taken into account here that in the Constituent Structures the last added elements are placed at the front.

Table E indicates these checks with: "NP ((head head)) ((head Nmod-S) (Nmod-a head)) ((det 1))".

This table can be extended for other constituents and also in the number of test lines per constituent. A second check is made for each constituent category by detecting incorrect combinations in the characteristic field using table F, which is stored as such in a memory part to be designated as sixth tabular memory. In a NP, for example, the combination "inneuter" and "neuter" associated with the word combination "de huis" can be regarded as incorrect. This is reflected in the absence of both neuter and neuter in the Characteristics field.

It is also possible to look at combinations deemed necessary in the Characteristics field.

Each constituent must display at least one sanctioned combination of Feature labels in the feature field.

For example, the Characteristic field of the NP constituent "a big house" includes the combination "indefinite, sing 3, adj-not-infl., Adj. Enumerative, neuter" while a proven standard combination is for example (indefinite, adj.-not-inflected , neuter), which is found in the said Feature sheet d. The NP constituent "the big house" has the Characteristics field (definite, si ng 3, adj. Inflected, adj. Enumerative, neuter), in which the standard combination (definite, adj. Inflected) is found. Thus, for a NP constituent, a series of standard combinations of Feature labels can be drawn up. Such a set of standard combinations for an NP could be: NP ((definite, adj.-inflected) (plu 3, adj.-inflected) (inneuter, adj.-inflected) (neuter, indefinite, adj.-not inflected )). Such standard combinations can also be drawn up for other types of constituents, which are stored together in the memory part to be regarded as the sixth tabular memory (see Table F).

A constituent must then satisfy at least one of the associated standard combinations. If not, the value of the probability field for that constituent must be decreased. For example, the erroneous NP constituent has "a big house" as Attribute sheet d: (indefinite, sing 3, adj. Inflected, adj.-emumerative, neuter), in which none of the tried-and-tested standard combinations are found.

An error code can also be entered here in the Error message field, which is also made visible on the screen when the parsed sentence is presented.

In the next step 76, the question is asked whether the value of the Probability calibration field of the copy constituent is greater than or equal to the threshold value. In the negative answer to question (N), the copy constituent is eliminated, and the step 56 or 62 to be considered in FIG. 6 considered terminated.

In the affirmative answer (Y) to the question of step 76, the next step 77 asks whether the Constituent can still be closed in view of the current threshold value. The possible functional labels for the current Constituent within the Constituent are tested on the Command field. If the present Constituent in the Stapelveld has more than one functional label (Y), this test has already been passed with a positive result in a previous stage, and therefore does not need to be repeated. Thus, as in the situation where it appears that the Constituent can still be terminated at the current threshold, proceed to step 78. Otherwise, in FIG. 7 displayed program section will then be considered terminated (N). This step can be considered optional, since it is not necessary for the parsing process, but rather accelerates the obtaining of the end result. In the next step 78, this constituent is included in Representation memory.

Using the flow diagram of Fig. 8 now follows a detailed explanation of the program section 69.

Since this program portion has many similarities with program steps 71 to 78, this will be referred to in some program steps.

At step 79 a copy is made of the Parent constituent for the purpose of the working memory, and as such is also used in the following steps.

In the next step 80, the probability factor of the copied constituent is adjusted by that

Probability to be corrected by the Probability factor of the constituent dominated by the Parent constituent (member constituent), for example by multiplying both probability factors. Another suitable function f (x, y) from both probability factors x and y can also be used to determine the new value of the probability factor.

At step 81, the functional label, as mentioned at step 69 and at steps 99 and 104 to be described later, is included in the Stack field of the copy constituent.

Then, in step 82, a combination pair of the functional label and the dominated constituent is added successively to the structural element field of the copy constituent.

At step 83, in the fourth tabular memory, a Characteristic list is passed through at the category designation of the Parent constituent and the functional label of the Member constituent dominated by this constituent (see table 0). All Attributes, which appear in this as well as in the Attributes list of the constituent copied in step 79, but not in the Attributes !! jst of the Member constituent, are removed from the Attributes field of the copied constituent. In addition, all attributes contained in the second Attribute List from the table, if not already present in the Attribute Sheet d of the copied constituent, are added to the Attribute Sheet d of the copied constituent. This step is similar to step 74.

At step 84, a program section is executed as also described at step 75.

The subsequent steps 85, 86 and 87, which apply to the copied constituent, are similar to steps 76, 77 and 78, provided that this is a Constituent and not a Word.

Using the flow diagram of Fig. 9 now follows a detailed explanation of the program section 37 regarding the termination of the current Constituent structure.

At step 88, the current Constituent structure, which was presented in the previous step 36, is from the Temporary memory (see Fig. 3), from the Temporary Switch memory (see Fig. 4), from the program section 3Ü (see Fig. 5) or from the program sections 103 and 108 in FIG. 9 added to the Representation Memory. Since at step 65 a Constituent structure may have been created in which the Stacking field does not contain an element, it is not allowed to close such a constituent. As a result, at step 89 the question is asked whether the Stacking Field is empty. In the affirmative answer (Y), this program part is considered to have ended. In the case of a negative answer (NI), step 90 asks whether the Command field of the current Constituent contains a value. If this is not the case (N), the Constituent is a Top constituent, and is not eligible for closure. At a Top Constituent, the program portion of Fig. 9 then considered terminated. If the answer at step 90 is affirmative (Y), then step 91 follows, where a Copy Constituent is made of the current Constituent. In the next step 92, the Category of the Copy Constituent, hereinafter referred to as Member Constituent, is determined. Then, at step 93, the category of the Parent Constituent is determined, based on the Member Constituent's Command Field. Subsequently, using the data obtained in steps 92 and 93 on the category of the Member constituent and its parent constituent and on the basis of a memory part to be designated as seventh tabular memory, the corresponding grammar rule is selected in step 94 (see table G).

If the category of the Member Constituent is an NP and the category of the Parent Constituent is an "S", it follows from the grammar rule NP (S, FNP) that the Member Constituent can be assigned a functional label fNP. In principle, there can also be more than one element in that place in the table.

Then, at step 95, a filtering procedure is performed, in which conditions regarding the elements in the Stacking Field are tested. The assessment criteria are stored in the third table memory, the assessment itself being entirely in accordance with that of step 67 (see table C).

In the next step 96 it is examined whether the value in the Probability field is greater than or equal to the threshold value. After all, it is possible that in step 95 the value of the Probability field has decreased. If not (N), the program section 37 is considered to be finished. Otherwise (Y), step 97 follows, asking whether the grammar line selected in step 94 contains functional labels. If this is not the case (N), closing of the current Constituent is not possible, and the process of closing the current Constituent should be considered as ended. However, if an affirmative answer (Y) is given to this question, then at least a meaningful closure procedure is possible. It must however be ascertained whether one or more functional labels have been proposed at step 94. To this end, the next step 98 asks whether exactly one functional label was proposed at step 94. In the affirmative answer (Y), step 99 follows, and in the negative answer (N), step 100. In the program section 99, for the current Constituent it is proposed that as an element with the functional label as indicated in the grammar line of step 94, to function within the Constituent Structure of the Parent Constituent. The Parent constituent is indicated in the Command field of the current Constituent. The current Constituent with the proposed functional label is then included in the Memer Field of the Parent Constituent, which has been discussed in detail at steps 79 to 83. At step 101, it is determined whether during the program portion 99 (as explained in step 78) an element has been added to the Represents memory. In the case of a negative answer (N), this program part is considered to have ended. In the affirmative answer (Y) to the question at step 101, proceed to step 102, where the last added element is retrieved from the Representation memory, removed from the Representation memory, and then subjected to the procedure regarding the step 103 at terminating this element as indicated in the sequence of steps starting with number 88. At step 100, the current Constituent structure is copied, with the Copy further considered as the current Constituent. At program portion 104, it is proposed for the current Constituent to function as an element with the functional label indicated as the first element in the grammar line of step 94, within the Constituent Structure of the Parent Constituent. The current Constituent with the proposed functional label is then included in the Memer field of the Parent constituent in question, which is explained in detail at steps 79 to 87. At step 105, the first element (the functional element) is removed from the grammar line.

Next, it is checked at step 106 whether during the program part 104 the number of Constituents in the Representation memory has increased, which could have happened with the program part 104. In case of a negative answer (N) to this question, it is immediately returned to step 98 to perform the closing process of the current Constituent using the purified grammar rule. In the affirmative answer (Y) to the question of step 106, in the next step 107, the last added Constituent structure is retrieved from Representation memory and removed from it and considered as the current Constituent, whereupon in the next step 108 the termination program is indicated by the series of program steps starting with number 88 applies. After step 108, the program returns to step 98.

Since temporary labels may have been assigned during the previous program, these temporary labels will be replaced by final labels in connection with the actual parsing process in the following program section. For example, for the Sample Grammar used here in all Constituents stored in Representation Memory, the fNP labels will be replaced, where possible, by the final labels Subject (Subject), Direct Object (Object), Indirect Object Object) and Predicate (Predicate), which are respectively used under the following abbreviations: Subj., Obj., Indobj., Pred.

To this end, at step 109, the count position of the m * counter, with which a reference to the rank number of a constituent is obtained, is reset to 0, after which at step 110 the count position of this counter is increased by "1".

In step 111 it is checked whether the count of the m * -teil has already exceeded the maximum value (m * max) corresponding to the number of constituents. In the affirmative answer (Y), this program section is considered to have ended. In the negative answer (N) of the question at step 111, the next step 112 asks whether the Command field of the m * -th Constituent structure contains a data. If the question is answered in the affirmative (Y), then it is not a Top Constituent, and it is therefore not necessary for definitive labels to be assigned.

If the question is answered in the negative (N), program phase 113 follows with regard to replacing the provisional functional labels with final ones, after which the program returns to step 110.

With reference to Fig. 11 now follows a more detailed explanation of program phase 113.

At step 114, an n-counter, the count of which refers to the rank number of the member pair in the Structural Elements field of the current Constituent, is reset to 0. Then, at step 115, the count of this counter is increased by "1".

In the next step 116, the question is asked whether the count of the n-counter has already exceeded the maximum value (nmax), which corresponds to the number of member pairs in that Structural element field. If this is the case (Y), the program portion 113 is considered terminated, and the program goes back to step 110. If the question (N) is answered negatively at step 116, the second element of the next step 117 becomes the pair of the current Constituent to be determined by indicating the count of the n-counter, and designated as the current Constituent. In the next step 118, the question is asked whether the element designated as the current Constituent is indeed a Constituent, and not a Word. If not (N), the program goes back to step 115. If it is (Y), the current Constituent becomes on its structure and therefore on the contents of the Stack field of the current one, i.e. Constituent designated at step 117, examined. This implies that the same procedure as described at steps 114, 115, 116, 117 and 118 must now also be carried out again within the program part 119 with steps 120, 121, 122, 123 and 124, but now on the Member pairs of the element designated as the current Constituent. Should one of the Member pairs have a second element, which does not represent a Word but a Constituent (which may be the case with relative clauses, for example), then the next step 125 must also be replaced by a program section marked with indication 119. In this description, the situation is examined, in which the second element is designated as Word for all Member pairs at step 118. The consequence of this assumption is that step 119 can actually be omitted and the program ends up at step 126, where the question is asked whether one of the elements on the Stack field of the Constituent selected at step 118, namely the second element of the nth pair, is a temporary label. If this is not the case (N), the program returns to step 115. In case of an affirmative answer (Y) to the question in step 126, a program part follows in step 127, where the temporary functional label is replaced by a definitive functional label. The contents of the Stack field of the m * -th Constituent structure are thereby examined, and this content is compared with certain grammar rules, which are stored in a memory section to be designated as eighth tabular memory. Such a line could be: {(fNP, Vfin-main> Smod, fNP, Endmark) (Subj., Obj.), Which means that the first fNP is replaced by a label "Subject" and the second fNP by a label "Suffering Object" (see Table H).

If a number of solutions occur with a certain data content of the Stack field, this leads to a corresponding number of Top constituents, all of which are written down in the Representation memory. This is done by copying the current structure with the preliminary functional labels.

First of all, the provisional functional labels in the Structural Elements field of each Summit Constituent are now replaced in a corresponding manner.

Then, at step 128, a filtering process is performed using a ninth tabular memory to be designated (see Table I), verifying whether the conversion of preliminary functional labels generated by the grammar rule of this program section is permitted is based on the obligatory agreement of certain elements in the Characteristic Field for some related Constituents. In particular, there is such a relationship between the Constituent structures "Subject" and "Verb", where the inflection indication should be consistent. Table I lists the appropriate grammar rules, which are arranged by category indication of the constituent to be investigated. Each grammar line also contains a list of elements and a list of characteristics. Of any structure (Constituent or Word), which with one of the od de liist. matching elements, the Attribute field is retrieved. The intersection of all these Attribute fields with each other and with the found list of Attributes cannot be empty.

For example, at the sentence "He is walking" the cross section of the Characteristics of the "Subject" with the found list of Characteristics will yield "sing 3". after which the section becomes empty by inserting the "Verb" Attributes. There is thus an ungrammatical sentence construction. The analysis obtained is now made less likely by decreasing the value in the Probability field, while also providing the Error message field with an error code.

After the filtering process, the program returns to step 115. Steps 129, 130 and 131 are the same as steps 126, 127 and 128. After the "syntax directed parser", a "syntax embedded parser", which operates according to the described inventive idea, are explained. The latter parser differs from the former parser in that the steps to be performed with a tabular memory are replaced by actions to be carried out programmatically. As an example, program step 75 (Fig. 7) will be considered, with part of the contents of Table E in the flow chart of Fig. 12 is shown.

In FIG. 12, at step 132, the question is asked whether the constituent is of the "S" category. If this is the case (Y), program section 133 follows, after which the program of step 75 is considered to have ended.

If not (N), step 134 follows, where the question is asked whether the constituent is of type Srei. If this is the case (Y), then a research program 135 aimed at it follows. If this is not the case (N), step 136 follows, where the possible processing phase starts at the constituent Scomp using step 137.

The following program section 138 concerning an NP constituent with the associated research program will now be explained in more detail. At step 138, the question is asked whether the constituent is an NP. If the answer to this is affirmative (Y), program step 139 follows, where the question is asked whether the functional label "head" occurs twice in a row. If this is the case (Y), program step 152 follows, where the value of the Probability field is decreased by a certain correction factor, after which the program proceeds to step 140. If the question at step 139 is answered in the negative (N), one follows or several similar questions, which are not further explained here, but which ultimately end up in program step 140. The question is whether the functional label "head'1 is followed by the label" det "in the constituent to be investigated. case (Y), then this is an incorrect order, and the value of the Probability field is corrected by a certain factor in the next step 141. In case of a negative answer (N) of the question at step 140, step 142 follows asking whether the functional label "head" is followed by the label "nmod-a" If this is the case (Y), in step 143 the value of the Probability field of the test is oeken Constituent decreased by a certain factor. If there is a negative answer (N) to the question at step 142, a similar question is processed in a next step. Eventually, and therefore also after steps 141 and 143, the program then arrives at the first step 144 of the next phase of the research program, where the question is asked whether the label "det" occurs once in the constituent. In the case of a negative answer (N) to this question, the value of the Probability field is decreased by a certain factor in the next program step 145. In the affirmative answer (Y), step 146 follows, where the question is asked whether the constituent contains only one functional label "head". If not (N), a decrease of the Probability field value is made at step 147. However, in the program steps to be performed in this phase, the program run at step 75 is considered completed. If the question is answered negatively (N) in step 138, then step 148 follows where the question concerns a PP constituent. Upon confirmation (Y) of the question at step 138, a related research program is run. A similar program portion but for an AP constituent occurs at steps 150 and 151, after which the program equivalent to step 75 of FIG. 2 is considered terminated.

A similar programmatic treatment is also possible for the other tables of the "syntax directed parser".

Table A

(ap (adj-inflected adj-not-inflected)) (np (definite indefinite neuter inneuter singl sing2 sing3 plul plu2 plu3)) (conj-np (definite indefinite neuter inneuter))

Table B

(article (np (det)) (pp (det np)) (s (det np)) (s-rel (det np)) (s-comp (det n))) (adj (ap (head)) ( np (head ap)) (s (head ap np) (head ap-adj)) (s-rel (head ap np) (head ap-adj)) (s-comp (head ap np) (head ap-adj )) (ap-adj (head))) (adv (s (head ap-adv)) (s-rel (head ap-adv)) (s-comp (head ap-adv)))) (advmod (ap ( mod ap)) (np (mod ap ap)) (s (mod ap ap np) (mod ap-adj)) (s-rel (mod ap ap np) (mod ap-adj)) (s-comp (mod ap ap np) (mod ap-adj))) (p-word (np (head pp)) (pp (head)) (s (head ppHvfin-particle)) (s-rel (head ppHvfin-particle)) ( s-comp (head pp) (vfin-particle))) (c-word (s (comp s-comp))) (noun (np (head)) (pp (head np)) (s (head np)) (s-rel (head np)) (s-comp (head))) (pro-subst (np (head)) (pp (head np)) (s (head np)) (s-rel (head np) ) (s-comp (head np))) (pro-rel (np (head np s-rel))) (pro-adj (np (det)) (pp (det np)) (s (det np)) (s-rel (det np)) (s-comp (det np))) (verb (s (vfin-main)) (s-rel (vfin-main)) (s-comp (vfin-main))) (interpunction (s (endmark)))

Table C

(np (det nmod-a) head) (pp O pobj) (ap (mod)) (ap-adj (mod) head) (ap-adv (mod) head) (s () fnp) (s-rel ( fnp) vfin-main) (s-comp (fnp) vfin-main) (conj-s (seq)) (conj-pp (seq)) (conj-ap (seq)) (conj-s-rel (seq) )

Table D

(ap (head (adj-inflected adj-not-inflected)) (amod (adj-inflected adj-not-inflected))) (np (det (neuter inneuter sing3 plu3 definite indefinite))) (head (singl sing2 sing3 plul plu2 plu3 neuter inneuter)))

Table E

(s ((fnp fnp) (smod smod) (smod fnp) (fnp smod) (vfin-particle fnp) (vfin-particle) (fnp vfin-particle vfin-main fnp) (fnp vfin-particle fnp vfin-main fnp )) () ((vfin-main 1) (fnp 3))) (s-rel () () ((vfin-main 1) (fnp 3))) (s-comp () () ((vfin- main 1) (fnp 3))) (np ((head head)) ((det head) (nmod-a head) (head nmod-s) (det nmod-s)) ((det 1) (head 1) (nmod-a 1))) (pp ((head head) (head head pobj head)) () {(pobj 1)))

Table F

(np ((neuter inneuter) (singl sing2 sing3 piul pl u2 plu3)) {(definite adj-inflected) (piu3 adj-inflected} (inneuter adj-inflected) (neuter indefinite adj-not-inflected)))

Table G

(np (conj-np conj) (pp pobj) (s fnp) (s-rel fnp) (s-comp fnp)) (conj-np (pp pobj) (s fnp) (s-rel fnp) (s- comp fnp)) (ap (ap amod) (np nmod-a) (s smod) (s-rel smod) (conj-ap conj) (s-comp smod)) (conj-ap (ap amod) (np nmod -a) (s-smod) (s-rel smod) (s-comp smod)) (ap-adj (s fnp) (s-rel fnp) (s-comp fnp)) (ap-adv (s smod) ( s-rel smod) (s-comp smod)) (pp (np nmod-p) (s smod) (s-rel smod) (conj-pp conj) (s-comp conj)) (conj-pp (np nmod -p) (s-smod) (s-rel smod) (s-comp smod)) (s (conj-s conj)) (s-comp (s fnp)) (s-rel (np nmod-s) (conj -s-rel conj)) (conj-s-rel (np nmod-s))

Table Η {{fnp vfin-main endmark) (subj)) ((fnp vfin-main smod endmark) (subj)) ((smod vfin-main fnp endmark) (subj)) ((smod vfin-main fnp smod endmark) (subj)) ({vfin-main fnp endmark) (obj)) ((vfin-main fnp smod endmark) (obj)) ((fnp vfin-main fnp endmark) (subj obj) (obj subj)) ((fnp vfin-main smod fnp endmark) (subj obj)) ((smod vfin-main fnp fnp endmark) (subj obj)) ((fnp vfin-main fnp smod endmark) (subj obj) (obj subj)) ((vfin- main fnp fnp endmark) (indobj obj)) ((vfin-main fnp smod fnp endmark) (indobj obj)) ((vfin-main fnp fnp smod endmark) (indobj obj)) ((vfin-main smod fnp fnp endmark) (indobj obj)) ((fnp vfin-main fnp smod smod endmark) (subj obj) (obj subj)) ((fnp vfin-main smod fnp smod endmark) (subj obj)); etc ((fnp vfin-main fnp fnp endmark) (subj indobj obj) (indobj subj obj) (obj subj indobj)) ((fnp vfin-main smod fnp fnp endmark) (subj indobj obj)) ((smod vfin-main fnp fnp fnp endmark) (subj indobj obj)) ((fnp vfin-main fnp smod fnp endmark) (subj indobj obj) (indobj subj obj) (obj subj indobj)) ((fnp vfin-main fnp fnp smod endmark) (subj indobj obj) (indobj subj obj) (obj subj indobj)); etc ((fnp vfin-main) (subj)) ( (fnp fnp vfin-main) (subj obj)) ((fnp fnp fnp vfin-main) (subj indobj obj))

Table I

(s ((subj vfin-main) (sing3 sing2 singl plu3 pl u2 plul)) ((subj) (nominative definite indefinite proper)) ((obj) (not-nominative definite indefinite proper)) ((indobj) (not- nominative definite indefinite proper))) (s-rel ((subj vfin-main) (sing3 sing2 singl plu3 plu2 plul)) ((subj-rel vfin-main) (sing3 plu3 singl sing2 plul plu2)) ((subj) ( nominative definite indefinite proper)) ((obj) (not-nominative definite indefinite proper)) ((indobj) (not-nominative definite indefinite proper))) (s-comp ((subj vfin-main) (sing3 sing2 singl plu3 plu2 plul)) ((subj) (nominative definite indefinite proper)) ((obj) (not-nominative definite indefinite proper)) ((indobj) (not-nominative definite indefinite proper)))

Claims (15)

1. Method for parsing a phrase, expressed in natural language, in phrases to be described with functional indications on the basis of word units which have been translated into verbal categories and application-oriented categories, which method is used by the following, per word unit steps to be taken from the sentence are characterized: a. Determining per word unit and per constituent the functional word category within the constituent and / or within a new constituent to be created, on the basis of information about the verbal category of the relevant word unit and the category of the constituent, b. Describing for each constituent on the basis of data about the category of the constituent and the category of the constituent dominating this constituent a measure concerning the closure of the constituent, as well as assigning a functional label, whether or not provisional, constituent to be closed, c. Testing the present constituent against syntax-based rules regarding the coherence of words and / or constituents within this constituent in at least one of the latter two steps, and, if necessary, revalue an expectation factor assigned to the sentence representation, as well as selecting any sentence representation whose expectation factor is above a certain threshold. Method for parsing a sentence in natural language according to claim 1, characterized in that determining the functional category per word unit per constituent also takes place on the basis of the functional category of the previous word unit. Method for parsing a sentence in natural language according to claim 1, which method, further to testing the present constituent and, if necessary, revaluing the expectation factor, is further characterized by replacing a provisional functional label of a closed constituent by a final functional label. Method for parsing a natural language sentence according to claim 3, which method, in addition to the replacement of a provisional functional label, is characterized by revaluing the expectation factor assigned to the relevant sentence representation and selecting the sentence representation with the highest expectation factor. Method for parsing a natural language sentence according to claim 1, characterized in that after determining that there is no expectation factor assigned to a sentence representation which is above the threshold value used, this threshold value is lower value is set. Method for parsing a natural language sentence according to claim 1, which method is characterized by determining the functional word category within that constituent and / or the associated word component per word unit, per word structure concerning that word unit and successively for each generated constituent new constituent to be created. Method for parsing a natural language sentence according to claim 6, characterized in that after determining the functional word category within the constituent and / or new constituent to be created for all word structures relating to a word unit, the description of a functional word category is defined. the conclusion of the constituent concerning measure takes place. Method for parsing a sentence in natural language according to claim 6, characterized in that each time after the per word structure concerning a word unit, a determination of the functional word category within the constituent and / or a new constituent to be created is determined each time. description of the conclusion of the constituent concerning measure takes place. Method for parsing a natural language sentence according to claim 1, which method is characterized by determining the functional word category within the constituent and / or new for each word constituent, per word structure concerning that word unit and for each generated constituent. constituent to be created and description of the conclusion of the constituent concerning measure. A method for parsing a natural language set according to claim 1, which method is characterized by determining the functional word category within that constituent and / or for each word structure concerning that word unit, successively for each word structure concerning that word unit and / or the newly created constituent. Method for parsing a sentence in natural language according to claim 10, characterized in that after determining the functional word category for all constituents within that constituent and / or new constituent to be created, the description of a the constituent concerning measure is closed. Method for parsing a sentence in natural language according to claim 10, characterized in that each time after the determination of the functional word category within that constituent and / or a new constituent to be created per constituent, a description is given of a the conclusion of the constituent concerning measure takes place. Method for parsing a natural language sentence according to claim 1, which method is characterized by determining the functional word category within the constituent and / or new for each word structure, per constituent created and for each word structure concerning that word unit. constituent to be created and description of the conclusion of the constituent concerning measure. Method for parsing a natural language sentence according to claim 1, which method for determining a functional category is characterized by the following steps: a. Using information regarding the word structure and that of the offered selecting a constituent in a memory of information related to the functional label to be assigned the word unit, b. Adding the word unit with selected functional label to the offered constituent, c. Examining each constituent for the presence of incorrect combinations, incorrect order and incorrect number of the assigned functional label, and d. Testing the constituent's probability factor against the applicable threshold value. A method of parsing a natural language phrase according to claim 1, which method relates to