| Because nature of segment record the of portion of the input text it does not result the parallel processing decoding algorithm, when a is created to describe a in the destruction of other records describing the same portion or parts of it. Local ambiguities caused by multiple word senses, idioms and the like may result in more than one segment record being created to describe a particular portion of the text, but usually only one of them is Context-free phrasestructure grammars have been known to be inadequate DIAGNOSIS AS A NOTION OF GRAMMAR for describing natural languages for many years, and context-sensitive phrase Mitchell Marcus structure grammars have not been found Artificial Intelligence Laboratory to be very useful, either. M.I.T. Augmented phrase structure grammars, however, appear to be able to express the facts of a natural language in a very concise and convenient manner, This paper will sketch an approach to they have the natural language parsing based on a new power of computer programs, while conception of what makes up a recognition maintaining the appearance of grammars. grammar for syntactic analysis and how such able to combine with its neighbors to become part of the analysis for an entire sentence. IV. SYNTHESIS OF TEXT (ENCODING) Encoding is the process by which strings of text are produced from record Although APSG was used successfully a grammar should be structured. This theory to of syntactic analysis formalizes a notion implement one fairly large system (NLPQ), it is too early to do a thorough appraisal very much like the psychologist's notion of of its capabilities. \"perceptual strategies\" [Bever \"70] and Through the extensive use anticipated in the next year however, makes this formalized notion -which will be its strengths called the notion of wait-and-see and weaknesses should become more diagnostics -a central and integral part of apparent. a theory of what one knows about the | From the examples it can also creation specifications take the form be seen of short procedures consisting of statements that for setting the values of attributes. Each element in a creation specification is basically of the form attribute=value (where \"=\" means replacement), but again this is not obvious becuase of the notational shortcuts used. For example, SUP(VERBSTEM) is short for SUP=SUP(VERBSTEM), PRESPART is short for PRESPART:I (note that this form has a different meaning when it is used in a condition specification), and -SUBJECT is short for SUBJECT=0. In all of the examples here, the attribute whose value is set would be in the segment record being built, but that need not always be the case. If, for example, there were some reason to want to give the PRESPART indicator. When the VERB segment shown above comes off the stack, a rule would be applied to put the string \"are\" into the output. Then, after application of a couple more rules, the top of the stack would have the four pairs VERBSTEM [ SUP \"SERVIC\" i I null N null G null which would result in the string \"servicing\" being produced after four cycles of the algorithm. Complete encoding examples may be found in Reference 3. | ! i | |
| The set that the named record read \"in the somewhere in the the current record. is appear $'ACTION\" and means must chain of In previous notation \"ACTION'\" \"ACTION\" SUPerset the section the named record \"SERVIC\" was defined to have a SUP of \"ACTIVITY'. If the named record \"ACTIVITY\" were similarly defined to have a SUP of \"ACTION', the segment record discussion here would satisfy the condition $'ACTION'. From the above examples it can be seen that the condition specifications take the form of logical expressions involving the values of attributes. Each element in a condition specification is basicaly of the form value.relation.value, but this is not obvious because there are several notational shortcuts available in the rule language. For example, \"BE\" is short for SUP.EG.'BE',PRESPART is short PRESPART.NE.0, and -~SUBJECT is short for SUBJECT.EQ.0. The elements are combined by and's (commas) and or's (vertical bars). In most cases the attribute whose value is being tested is to be found in the segment record associated with the constituent, but that is not always the case. For example, ING* tests the value of the ING indicator in the named record pointed to by the SUP of the segment record, and could be written ING(SUP) or ING(SUP).NE.0. Another example is NUMB(NOUNPH) which was used to refer to the value of the NUMB indicators in the NOUNPH the PROG and FORM indicators and setting the segment in one of the rules above. resulting in the following two pairs being put on the top of the stack: VERB PRES,P3,PLUR VERBPH I SUPPRESPART'SERVIC\" 1 The above rule says that a VERBPH segment with a PROGressive indicator should be expanded into a VERB segment with a SUP of \"BE\" and the same VERBPH, followed by a new VERBPH which begins as the old one and then is modified by deleting a copy (automatically) of segment FORM indicators as the Co., San Francisco, Calif., 1973, 63-113. THOUGHT AND LANGUAGE, R.C. Schank and K.M. Colby (Eds.), W.H. Freeman and English sentences,\" in COMPUTER MODELS OF computation and use for understanding 6. Simmons, R.F., \"Semantic networks: their ISUP \"BE\" for natural language analysis,\" COMM. ACM 13, 10 (Oct. 1970), 591-606. for following pair were to come off the top of the stack: VERBPH I SUP \"SERVIC\" I PRES,P3,PLUR,PROG the following encoding rule could be applied: VERBPH(PROG) --> VERB('BE',FORM:FORM(VERBPH)) 5. Woods, W.A., \"Transition network grammars VERB(-PROG,-FORM,PRESPART) Level Languages, SIGPLAN NOTICES 9,4 (April 1974), 91-100. programming,\" Proc. Symp. on Very High level language for simulation California, Oct. 1972. 4. Heidorn, G.E., \"English as a very high to a simulation programming system,\" Technical Report NPS-55HD72101A, Naval Postgraduate School, Monterey, 3. Heidorn, G.E., \"Natural language inputs under \"I\") is put out. Eventually the stack becomes empty and the algorithm terminates, having produced the desired output string. For example, if at some point the 24th NAT'L CONF., ACM, NY, 1969, 399-417. B., Deverill, R.S., \"REL: a rapidly extensible language system,\" In PROC. 2. Thompson, F.B., Lockemann, P.C., Dostert, structures of the sort already shown. The manner in which this processing is to be done is specified by APSG encoding structure of language. By recognition grammar, we mean here what a speaker of a ACKNOWLEDGEMENTS language knows about that language that rules. The right side of an encoding rule specifies what segments a segment of the type on the left side is to be expanded into. Conditions and structure-building actions are included in exactly the same manner as in decoding rules. The encoding algorithm begins with a single segment record and its associated type side-by-side on a stack. At each cycle through the algorithm, the top pair is removed from the stack and examined. If there is a rule that can be applied, it results in new pairs being put on the top of the stack, according to its right hand side. Otherwise, either the character string value of the NAME attribute of the SUP of the segment record (e.g. \"servic\") is put out, or the name of the segment type itself (e.g. I. Balzer, R.M., and Farber, D.J., \"APAREL -a parse-request language,\" COMM. ACM 12, 11 (Nov. 1969), 624-631. REFERENCES I am indebted to my former students allows him to assign grammatical structure at the Naval Postgraudate School for their efforts on the original implementation and to the word strings that make up utterances in that language. application, my colleagues at IBM Research --Martin Mikelsons, This theory of grammar is based on the Peter Sheridan, Irving Wladawsky and hypothesis that every language user knows as Ted Codd --for their interest, ideas part of his recognition grammar a set of and work on the current implementatons highly specific diagnostics that he uses to and applications, and my wife, Beryl, decide deterministically what structure to for her build next at each point in the process of typing assistance and general helpfulness. parsing an utterance. | record structures of APSG. available by being obtained from stream, and a non-terminal segment becomes the input available whenever a rule is applied.) Then the rule instance record \"waits\" for a segment which can be the next constituent of the associated rule to become available. When such a segment becomes available, the rule instance record is \"extended\". When a rule instance record becomes complete (i.e. all of its constituents are available), the associated rule is applied (i.e. the segment record specified on the right is built and made available). There may be many rule instance records in existence for a particular rule at any point in time. structures of Simmons are very close to the records in APSG, and the semantic network correspond closely to attributes of segment APSG clearly has much in common other current computational theories, with the ideas of procedural specification and arbitrary conditions and strucutre-building actions being popular at this time. It would seem to be most similar to Woods\" augmented transition networks (ATN) [5], especially as used by Simmons [6]. Registers in the ATN model very linguistic with VI. CONCLUDING REMARKS The decoding algorithm used with APSG is basically that of a bottom-up, left-to-right, parallel-processing, syntax-directed compiler. An important and novel feature of this algorithm is something called a \"rule instance record\", which primarily maintains information abut the potential applicability of a rule. A rule instance record is initially created for a rule whenever a segment which can be the first constitutent of that rule becomes available. (A terminal segment becomes work has been documented yet. the expressions, is itself written as a set of APSG rules. This work is part of a project at IBM Research to develop a system which will produce appropriate accounting application programs after carrying on a natural language dialogue with a businessman about his requirements. APSG is also used in the development of alaaguage query system for relational data bases and is being considered for use in other projects at IBM. None of this recent a natural being information in parentheses) into lambda AGENT record of an action an ABC attribute equal to one more than the XYZ attribute of the concept record associated with that action (i.e. the named record pointed to by its SUP), the following could be included in the last rule shown: ABC(AGENT)=XYZ(SUP)+I which can be read as \"set the ABC attribute of the AGENT of this record to the value of the XYZ attribute of the SUP of this record plus I.\" Although in the example rules given here the conditions are primarily syntactic, semantic constraints can be stated in exactly the same manner. Much of the record building shown here can be considered semantic (and somewhat case oriented). The important point, however, is that the of condition testing and structure building done is at the discretion of the person who writes the rules. Complete specifications for the APSG rule language are given in Reference 3. same input and does the same processing the FORTRAN version. An interesting feature of this new version is that the compiler part, whose primary task condition and creation specifications (i.e. is to translate as kind V. IMPLEMENTATIONS AND APPLICATIONS As part of the original work on APSG a computer system called NLP (~atural Language ~rocessor) was developed in 1968. This is a FORTRAN program for the IBM 360/370 computers which will accept as input named record definitions and decoding and encoding rules in exactly the form shown in this paper and then perfor m decoding and encoding of text [3]. A set of about 300 named record definitions and 800 rules was written for NLP to implement a specific system (called NLPQ) which is capable of carrying on a dialogue in English about a simple problem [3,4]. More recently a LISP implementation of NLP has been done, which accepts exactly the in the GPSS simulation language to solve the queuing problem and then producing a program | ! D ! ! ! I ! ! | ! ! p! |