File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/80/p80-1030_metho.xml
Size: 29,579 bytes
Last Modified: 2025-10-06 14:11:22
<?xml version="1.0" standalone="yes"?> <Paper uid="P80-1030"> <Title>PHRAN - A Knowledge-Based Nature\] Language Understender</Title> <Section position="4" start_page="117" end_page="117" type="metho"> <SectionTitle> 3.0 PHRAN </SectionTitle> <Paragraph position="0"> P~AN (PHRasal ANalyzer) is an English language understanding system which integrates both generative and non-productive language abilities to provide a relatively flexzble and extenstble natural language understanding facility. While PHRAN does have knowledge about individual words, it is not limited to such knowledge, nor ms its processing capability constrained by a word-based bias.</Paragraph> <Paragraph position="1"> Here are some examples of sentences PHRAN can understand: e 0i%men are encouraged by the amount of oil discovered in the Baltimore Canyon, an undersea trough 100 m$1es off the shore of New Jersey. (Newsweek, Feb 1980) (The previous sentences are analyzed by an uncompiled version of PHRAN on the DEC-20/4Q system at UC Eerkeley in from 2 to 9 seconds of CPU time).</Paragraph> <Paragraph position="2"> At the center of PHRAN is a knowledge base of phrasal patterns. These include literal strings such as &quot;so's your old man&quot;; patterns such as &quot;<nationality> restaurant&quot;, and very ~eneral phrases such as &quot;<person> <give> <person> <object> .</Paragraph> <Paragraph position="3"> Associated with each phrasal pattern is a conceptual template. A conceptual template is a piece of meanln~ representation with possible references to pieces of the associated phrasal pattern. For example, associated with the phrasal pattern &quot;<nationality> restaurant&quot; is the conceptual template denoting a restaurant that serves <nationality> type food; associated with the phrasal pattern &quot;<person~> <give> <personJ> <object>&quot; is the conceptual template that denotes a transfer of possession by <person1> of <object> to <personJ> from <person1>.</Paragraph> </Section> <Section position="5" start_page="117" end_page="118" type="metho"> <SectionTitle> ~.O HOW PH~AN WORKS ~.1 Overall Algorithm </SectionTitle> <Paragraph position="0"> FHRAN is made up of three parts - a database of pattern-concept pairs, a set of comprehension routines, and a routine which suggests appropriate pattern, concept pairs. PHRAN takes as input an English sentence, and as xt reads it from left to right, PHRAN comnares the sentence against patterns from the database. Whenever a matching pattern is found, PHRAN interprets that part of the sentence that matched the pattern as describing the concept associated with the pattern in the pattern-concept pair.</Paragraph> <Paragraph position="1"> When PHRAN analyzes a sentence, it reads the words one at a time, from left to right. It does just enough morphological analysis to recognize contractions and &quot;'s s. The pattern suggesting routine determines if any new patterns should be tried, and PHRAN checks all the new patterns to see if they agree with that part of the sentence already analyzed, discarding those that don't. A word's meaning is determined simply by its matching a pattern consisting of that literal word. Then a term is formed with the properties specified in the concept associated with the word, and th:s term is added to a list PHRAN maintains. PHRAN checks if the term it just added ~ the list completes or extends patterns that had alread3 been partially matched by the previous terms. If a pattern is completely matched, the terms matching the* pattern are removed and a new term, specified by th, concept part of the nattern-conceDt pair, is formed and replaces the terms the pattern matched.</Paragraph> <Paragraph position="2"> When PHRAN finishes processing one word it reads the next, iterating thls procedure until it reaches the end of e sentence. At this point, it should end up with a single term on its list. This term con%sins the conceptualization representing the meaning of the whole sentence.</Paragraph> <Paragraph position="3"> A pattern-concept pair consists of a specification of the phrasal unit, an associated concept, and some additional information about how the two are related. When PHRAN instantiates a concept, it creates an item called a term that includes the concept as well as some additional information.</Paragraph> <Paragraph position="4"> A pattern is a sequence of conditions that must hold true for a sequence of terms. A pattern may specify optional terms toq, the place where these may appear, ana what effect (if any) their appearance will have on the properties of the term formea if the pattern is matched.</Paragraph> <Paragraph position="5"> For example, consider the following informal description of one of the patterns suggested by the mention of the verb 'to eat' in certain contexts.</Paragraph> <Paragraph position="6"> { pa;tern to recognize |<first term: represents a person> * <second term: is an actlve form of EAT> <OPTIONAL third term: represents food>\] term to form (INGEST(ACTOR <first term>) (OBJECT <third term, if present, else FOOD>)) } Notice that the third term is marked as optional. If it is not present in the text, PHRAN will fill'the OBJECT slot with a default representing generic food.</Paragraph> <Paragraph position="7"> 4.1.. ~ Simple Example The following is a highly simplified example of how PHRAN processes the sentence &quot;John dropped out of school&quot;: First the word &quot;John&quot; is read. &quot;John&quot; matches the patter~ consisting of the literal &quot;John&quot;, and the concept associated with this pattern causes a term to be formed that represents a noun phrase and a particular male erson named John. No other patterns were suggested. ~his term is added on to *CONCEPTS, the list of terms PHRAN keeps and which will eventually contain the meaning of the sentence. Thus *CONCEPT* looks like < \[JORNI - person, NP\] > &quot;Dropped&quot; is read next. It matches the literal &quot;dropped&quot;, and an appropriate term is formed. The pattern suggesting routine instructs PHRAN to consider %he 'basic pattern associated with the verb 'to drop', which is: I \[<person> <DROP> <object>\] \[ ... I 1 Its initial condition is found to be satisfied by the first term in *CONCE ~PT e -- this fact is stored under that term so that succeeding ones will be checked to see if this partial match continues. The term that was formed after reading &quot;dropped&quot; is now added to the list. *CONCEPT* is now < \[JOMNI - person, NP\] , \[DROP - verb\] > PHRAN now checks to see if the pattern stored under the first term matches the term just added to CONCEPT too, and it does. This new fact is now stored under the last term.</Paragraph> <Paragraph position="8"> Next the word &quot;out&quot; is read. The pattern suggestion mechanism is alerted by the occurence of the verb 'drop' followed by the word 'out', and at this point It instructs PHRAN to consi ;r the pattern I \[<person> <DROP> &quot;out&quot; &quot;of&quot; <school> I \[ ... \] ! The list in *CONCEPT* is checked against this pattern to see if it matches its first two terms, end since that is the case, this fact is stored under the secord term. A term associated with 'out' is now added to *CONCEPT*: < \[JOHNI - person, NP\] , \[DROP - verb\] , lOUT \] > The two patterns that have matched up to DROP are checked to see if the new term extends them. This is true only for the second pattern, a~d this fact is stored unde~ the next term. The pattern l<person> <DROP> <object>) is discarded.</Paragraph> <Paragraph position="9"> Now the word &quot;of&quot; is read. A term is formed and added to *CONCEPT*. The pattern that matched to OUT is extended by OF so %he pattern is moved to ~e next term.</Paragraph> <Paragraph position="10"> The word &quot;high&quot; is read and a term is formed and added to *CONCEPt. Now the pattern under OF is compared against HIGH. It doesn't satisfy the next condition. PHRAN reads &quot;school&quot;, and the pattern suggestion routine presents PHRAN with two patterns: I. I \[ &quot;high .... school&quot; \] \[ represention denoting a school $o~ IOth through 12th</Paragraph> <Paragraph position="12"> noun modified by adjectiveJ 1 Both patterns are satisfied by the previous term and this fact is stored under it. The new term is added to *CONCEPT*, now: < ~JOHNI - person ~V2 \] ,~\[DROP - verb\] , \[OUT\] \[0FT , \[HIGH - sdjl , \[SCHOOL - sch6ol, noun\]'> The two patterns are compared against the last term, and both are matched. The last two terms a~'e removed from *CONCEPT*, and the patterns under 0F are checked to determine which of the two possible meanings we have should be chosen. Patterns are suggested such that the more specific ones appear first, so that the more specific interpretation will be chosen if all patterns match equally well.. 0nly if the second meanin~ (i.e. a school that is high) were explicitly specifled by a previous pattern, would it have been chosen.</Paragraph> <Paragraph position="13"> A term is formed and added to *CONCEPT*, which now contains < \[JOHNI - person, NP~ . \[DROP - verb\] \[OUT\] , \[0FI , \[HIGH-SCHOOLI - school, NPJ > The pattern under OF is checked against the last term in *CONCEPT ~. PHRAN finds a complete match, so all the matched terms are removed and replaced by the concept associated with this pattern.</Paragraph> <Paragraph position="14"> *CONCEPT* now contains this concept as the final result:</Paragraph> <Paragraph position="16"/> <Section position="1" start_page="118" end_page="118" type="sub_section"> <SectionTitle> 4.2 Pattern-Concept Pairs In More Detail </SectionTitle> <Paragraph position="0"> The pattern portion of a pattern-concept pair consists of a sequence of predicates. These may take one of several forms: 1. A word; which will match only a term representing this exact word. 2. A class name (in parentheses); will match any term ~epresenting a member @f this class (e.g. &quot;(FOOD)&quot; or &quot;(PHYSICAL-OBJECT)&quot;).</Paragraph> <Paragraph position="1"> ~. A pair, the first element of which is a property name end the second is a value; will match any ~e rm hav%ng the required valge of the property e.g. &quot;(Part-0f-Speech VERB)&quot;). In addition, we may negate a condition or specify that a conjunction or disjunction of several must hold. The following is one of the patterns which may be suggested by the occurrence of the verb 'give' in an To indicate the presence of optional terms, a list of pattern concept-pairs is inserted into the pattern at the appropriate place. These pairs have as their first element a sub-pattern that will match the optional terms. The second part describes how the new term to be formed if the maxo pattern is found should be modified to reflect the existence of the optional sub-pattern.</Paragraph> <Paragraph position="2"> The concept corresponding to the optional part of a pattern zs treated in a form slightly different from the way we treat regular concept parts of pattern-concept pairs. As usual, it consists of pairs of expressions. The first of each pair will be places as is at ~he end of the properties o~ the term to be formed, end the second will be evaluated first and then placed on that list.</Paragraph> <Paragraph position="3"> For example, another pattern suggested when 'give' is seen is the following:</Paragraph> <Paragraph position="5"> The terms of this pattern describe a person, the verb give, and then some pnysical object. The last term describes the optional terms, consisting of the word to followed by a person description. Associated with th~ pattern is a concept part that specifies what to do with the optional part if it is there. Here it specifies that the second term in the optional pattern should fill in the TO slot in the conceptualization associated with the whole pattern.</Paragraph> <Paragraph position="6"> This particular pattern need not be a separate pattern in PHRAN from the one that looks for the verb followed by the recipient followed by the object transferred. We often show patterns without all the alternatives that are possible for expositional purposes. Sometimes it is simpler to write the actual patterns separately, although we attach no theoretical significance to thxs disposition.</Paragraph> <Paragraph position="7"> When a pattern is matched. PHRAN removes the terms that match zt from *CONCEPT* and replaces them with a new term, as defined by the second part of the pattern-concept pair. For example, here is a pattern-concept pazr that may be suggested when the verb &quot;eat' is encountered: (\[(PERSON) (BOOT EAT) (\[((FOOD)) (FOOD (OPT-VAL I CD-FOBM))\])\]</Paragraph> </Section> </Section> <Section position="6" start_page="118" end_page="118" type="metho"> <SectionTitle> \[P-O-S 'SENTENCE CD-FORM '(INGEST (ACTO~ ?ACTOR) (OBJECT ?FOOD)) ACTOR (VAL~ I CD-FORM) FOOD 'FOOD\]) </SectionTitle> <Paragraph position="0"> The concept portion of this pair describes a term covering an entire sentence, and whose ~eaning is the action of INGESTing some food (Schank, 1975). The next two descriptors specify how $o fill in vaTiable parts of this action. The expression (VALUE n prop) specifies the 'prop' property of the n'th term in the matched sequence of the pattern (not including optional terms). OFT-VAL does the same thing with regards to a matched optional sub-pattern. Thus the concept description above specifies that the actor of the action is to be the term matching the first condition. The object eaten will be either the default concept food, or, if the optional sub-pattern was found, the term corresponding to this suo-pattern.</Paragraph> <Paragraph position="1"> Sometimes a slot in the conceptualization can be filled by a term in a higher level pattern of which this one is an element. For example, when analyzing &quot;John wanted to eat a cupcake&quot; a slight modification of the previous pattern is used to find the meaning of &quot;to eat a cupcake&quot;. Since no subject appears In this form, the higher level pattern specifies where it may find it.</Paragraph> <Paragraph position="2"> That is, a pattern associated with &quot;want&quot; looks like the following: { ~<person> <WANT> <in$initive>\] ,infinitive DFOHM This specifies that the subject of the clause following want is the same as the subject of went.</Paragraph> <Section position="1" start_page="118" end_page="118" type="sub_section"> <SectionTitle> 4.5 Pattern Manipulation In More Detail </SectionTitle> <Paragraph position="0"/> </Section> </Section> <Section position="7" start_page="118" end_page="119" type="metho"> <SectionTitle> 4.~.I Reading A Word - </SectionTitle> <Paragraph position="0"> When s word is read PHRAN compares the ~atterns offered by the pattern suggestinC/ routine with the list *CONCEPT* in ~ne manner aescrioea in ~ne example in section 4.1.3. It discards patterns that confllct with *CONCEPT* and retains the rest. Then FH~AN tries to determine which meaning ?f the word to choose, using the &quot;active&quot; patterns (those that have matched up to the point where PHRAN has read). It checks if there is a particular meaning that will match the next slot in some pattern or if no such definition exists if there is a meaninC/ that might be the beginning of a' sequence of terms -whose meaning, as determined via a pa~tern-concept pair, will satisfy the next slot in one of the active patterns. If this is the case, that meanin~ of the word is chosen. Otherwise PHRAR defaults to the fzrst of the meanings of the word.</Paragraph> <Paragraph position="1"> A new term is formed and if it satisfies the next condition in one of these patterns, the appropriate ~atzsrn Is moved to the pattern-list of the new term. If zhe next condition in the pattern indicates that the term speczfled is optional, %hen PHRAN checks for these Optlonal terms, and if it is convinced that they are not present, it checks to see if the new term satisfies the condition following the optional ones in the pattern.</Paragraph> <Section position="1" start_page="119" end_page="119" type="sub_section"> <SectionTitle> a.3.2 A Pattern Is Matched - </SectionTitle> <Paragraph position="0"> When a pattern has been matched completely, PHRAN continues checking all the other patterns on the pattern-list. When it has finished, PHRAN will take the longest pattern that was matched and will consider the concept of its pattern-concept pair to be the meaning of the sequence. If there are several patterns of the same length :hat we re matched PHRAN will group all their meanings together.</Paragraph> <Paragraph position="1"> New patterns are suggested end a disembiguation process follows, exactly as in the case of a new word being read. For example, the words &quot;the big apple&quot;, when recognized, will have two possible meanings: one being a large fruit, the other being New York Clty, PHRAN will check the patterns active at that time %0 determine if one of these two meanings satisfies the next condition in one of the patterns. If so, then that meaning will be chosen, Otherwise 'a large fruit' will be the default, as it is the first in the list of possible meanings.</Paragraph> </Section> </Section> <Section position="8" start_page="119" end_page="119" type="metho"> <SectionTitle> 4.~ Adverbs And Adverbial Phrases </SectionTitle> <Paragraph position="0"> In certain cases there is need for slightly modified notions of pattern and concept, the most prominent examples being adverbs and adverbial phrases. Such phrases are also recognized through the use of patterns. However, upon recognizing an adverb, PHRAN searches within the active patterns for an action that it can modify. When such an action is found the concept part of the pair associated with the adverb is used to modify the concept of the original action.</Paragraph> <Paragraph position="1"> Adverbs such as &quot;quickly&quot; and &quot;slowly&quot; are currently defined and can be used to modify conceptualizations containing various actions. Thus PHRAN can handle constructs like: John ate slowly. Ouickly, John left the house.</Paragraph> <Paragraph position="2"> John left the house quickly.</Paragraph> <Paragraph position="3"> John slowly ate the apple. John wanted slowly to eat the apple.</Paragraph> <Paragraph position="4"> Some special cases of negation are handled by specific patterns. For example, the negation of the verb want usually is interpreted ss meaning &quot;want not&quot; - &quot;~ didn't want to go ~o school&quot; means the same thing as &quot;Mary wanted not to go:to school&quot;. Thus PHRAN conzains the specifi~ pattern \[<person> (do> &quot;not&quot; <want> <inf-phrase>! which Is associated with this interpretation.</Paragraph> </Section> <Section position="9" start_page="119" end_page="119" type="metho"> <SectionTitle> ~-5 Indexing And Pattern Suggestion </SectionTitle> <Paragraph position="0"> Retrieving the phrasal pattern matching a particular utterance from PHRAN's knowledge base is sn important problem that we have not yet solved to our complete satisfaction. We find some consolation in the fact that the problem of indexing a large data base is a neccesary and familiar problem for all Enowledge based systems.</Paragraph> <Paragraph position="1"> We have tried two pattern suggestion mechanisms with PHRAN: I. Keying oatterns off individual words or previously matched patterns.</Paragraph> <Paragraph position="2"> 2. Indexing patterns under ordered seouences of cues go%ten from the sentence a~d phras~T paz~erns recognized in it.</Paragraph> <Paragraph position="3"> The first indexing mechanism works but it requires that any pattern used to recognize a phrasal expressions be suggested by some word in it. This is unacceptable because it will cause the pattern to be suggested whenever the word it is triggered by is mentioned. The difficulties inherent in such an indexing scheme can be appreciated by considering which word in the phrase &quot;by ana large&quot; should be used to trigger it. Any choice we make will cause the pattern ~o be suggested very often in contexts when it is not appropriate. ~nthis form, FHRAN's ~rocessing roughly resembles ELI's (Riesbeck et el, 19V59.</Paragraph> <Paragraph position="4"> We therefore developed the second mechanism. The ~ atterns-concapt pairs of the database are indexed in s ree. As words are read, the pattern suggesting mechanism travels down this tree, choosing branches according to the meanings of the words. It suggests to PHRAN the patterns found at the nodes it has arrived at. The list of nodes is remembered, and when the next word is read the routine continues to branch from them, in addition to starting from the root. In practice, the number of nodes in the list is rather smsll.</Paragraph> <Paragraph position="5"> For example, whenever a noun-phrase is followed by an active form of some verb, the suggesting routine instructs PHRAN to consider the simple declarative forms of the verb. When a noun-phrase is followed by the vero 'to be' followed by the perfective form of some verb, the routine instructs PHRAN to consider the passive uses of the last verb. The phrasal pattern that will recognize the expression &quot;by and large&quot; is found st the node reaches only after seeing those three woras consecutively. In this manner this pattern will be suggested only when neccessary.</Paragraph> <Paragraph position="6"> The main problem with this scheme is that it does not lend itself well to allowing contextual cues to influence the choice of patterns PHRAN should try. This is one area where future research will be concentrates.</Paragraph> </Section> <Section position="10" start_page="119" end_page="120" type="metho"> <SectionTitle> 5.O COMPARISON TO OTHER SYSTEMS </SectionTitle> <Paragraph position="0"> There are a number of other natural lenguage processing systems that either use some notion of patterns or produce meaning structures as output. We contrast PHRAN w~th some of these.</Paragraph> <Paragraph position="1"> An example of a natural language understanding system that produces declarative meaning representations Ss Riesbeck's &quot;conceptual analyzer&quot; (Riesbeck, 1974). Riesbeck's system (and the various systems that have descended from it) works by attaching routines to ind~vidusl words. These routines are generally responsible for building pieces of s meaning reprDsentstion. When a word is reed by the system, the routines associated with that word are used to build up a meaning structure that eventually denotes the messing of the entire utterance.</Paragraph> <Paragraph position="2"> While our sims are much in the spirit of Riesbeck's analyzer, we believe there ere both practical and theoreticsl d~fficulties inherent in his approach. For example, in R~esbeck's conceptual analyzer, specific understanding routines are needed for each word known to the system. Thus extending the system's vocabulary requires the creation and *debugging of new code. In addition, these routines function only in the understanding process. The knowledge they embody is inaccessible to other mechanisms, in particular, to production procedures.</Paragraph> <Paragraph position="3"> Moreover, because Riesbeck's approach is word-oriented, it is difficult to incorporate phrssel structures into his model. Some word of the phrase must have a routine associated w~tb it that checks for that phrase. At best, this implementation is awkward.</Paragraph> <Paragraph position="4"> One of the earliest language understanding systems to incorporate phrasal patterns is Colby's PARRY. PARRY is 8 s~mulation of a paranoid me~tal patient that contains a natural language front and (Psrklnson st al, 19~). It receives a sentence as input and ,na\]yzes it in several separate &quot;stages&quot;. In effect, PARRY replaces the input wi~h sentences of successively simpler form. In %he simplified sentence PARRY searches for patterns, of which there ere two bssic types: patterns used to interpret the whole ~entence, snd those used on~y to interpret parts of ~t {relative clauses, for example).</Paragraph> <Paragraph position="5"> For PARRY, the purpose of the natural language analyzer is only to translate the input into a simplified form that a model of a paranoid person may use to determine an appropriate response. No attempt Js made to model the analyzer itself after a human language user, as we are doing, nor are claims made to this effect. A system attempting to model human language analysis could not permit several unre\]e+ed passes, the use of s transition network grsmmsr to interpret only certain sub-strings in the input, or a rule permitting it to simply ignore parts of the input.</Paragraph> <Paragraph position="6"> This theoretical shortcoming of PARRY - hsving separate grammar rules for the complete sentence ~nd for sub-parts o&quot; it - is shsred by Henarix's LYFER (Hendrix. IO77).</Paragraph> <Paragraph position="7"> LIFER is designed to enable a database to be queried usJn~ 8 subset of the English language. As is t~_ case for PARRY, the natural language ansAysis done by ~Ar~R is not meant to model humans. Rather, its function is to translate the input into instructions and produce s reply as efficiently es possible, and nothing resembling s representation of tne meaning of the input is ever l ormea, u: course the purpose of LIFE~ is not to be th ~ front end of a system that understands coherent texts and which must therefore perform subsequent inference processes. Wh~le LIFER provides s workable solution to the natural language problem in a limited context I msny general problems of language analysis are not adoresseo in that context.</Paragraph> <Paragraph position="8"> SOPHYE (Burton, 1976) was designed to assist students in learning about simple electronic circuits. It can conduct a dialogue with the user in a restricted subset of the English language, and it uses knowledge about patterns of speech to interpret the input. SOPHIE accepts only certain questions and instructions concerning a few tasks. As is the case with LI-FER. the langusge utterances acceptable to the system are restricted to such an extent that many natural language processing problems need not be deelt with and other problems have solutions appropriate only to this context. In addition, SOPHIE does not produce any representation of the meanin~ of the input, and it makes more than one pass on the Input i~morlng unknown words, practices that nave already been crlticized.</Paragraph> <Paragraph position="9"> The augmented finite state transition network (ATN) has been used by a number of researchers to aid in the analysis of natural language sentences (for example, see Woods 1970). However, most systems that use ATN's incorporate one feature which we find objectioneble on both theoretical and practical grounds. This is the separation of analysis into syntactic and semantic phases. The efficacy and psychological validity of the separation of syntactic and sementicprocessing has been argued at lengthelsewhere (see Schar~ 1975 for example).</Paragraph> <Paragraph position="10"> In addition, most ATN based systems (for .xample Woods' LUNAR program) do not produce represents%ions, but rather, run queries of a data base.</Paragraph> <Paragraph position="11"> In contrast to the systems just described, Wilks' English-French machine ~ranslstor do~s not share several of their shortcomings (Wilks, 197~). It produces a representation of the meaning of an utterance, and it attempts to deal with unrestricted natural language. The maxn difference between Wilk's system and system we describe is that Wilks' patterns are matched against concepts mentioned in a sentence. To recognize these concepts he attaches representations to words in e dictionary.</Paragraph> <Paragraph position="12"> The problem is that this presupposes that there is a simple correspondence between %he form of a concept and the form of a language utterance. However, it is the fact that this correspondence is not simple that leads to the difficulties we are addressing in our work. In fact, since the correspondence of words to meanings is complex, it would appear ~hat a program like Wilks' translator will even~ually need %he kind of knowledge embodied in PHRAN to complete its analysis.</Paragraph> <Paragraph position="13"> One recent attempt at natural language analysis that radically departs f~om pattern-based approaches is Rieger ' and Small's system (Smell, 1978). This system uses word experts rather than patterns as its basic mechsnxsm. ~nelr system acknowledges the enormity of the knowledge base required for language understanding, and proposes s way of addressing the relevant issues. However, the idea of puttin~ as much information as possible under individual words is about as far from our -conception of language analysis as one can get, and we would argue, would exemplify all the problems we have described in word-based systems.</Paragraph> </Section> class="xml-element"></Paper>