<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1148">
  <Title>ON KNOWLEDGE-BASED MACHINE TRANSLATION</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
ON KNOWLEDGE-BASED MACHINE TRANSLATION
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
ABSTRACT
</SectionTitle>
    <Paragraph position="0"> This paper describes the design of tile knowledge representation medium used for representing concepts and assertions, respectively, in a subworld chosen for a knowledge-based machine u'anslation system. This design is used in the TRANSLATOR machine translation project. The kuowledge representation language, or interlingua, has two components, DIL and TIL. DIL stands for 'dictionary of interlingua' and descibes tile semantics of a subworld. TIL stands for 'text of interlingua' and is responsible for producing an interlingua text, which represents tile meaning of an input text in tile terms of trte interlingua. We maintain that involved analysis of various types of linguistic and eucyclopaedic meaniug is necessary for the task of autx)matic translatiou. The mechanisms for extracting and nlanipnlating and reproducing the nteaning of te~ts will be reported in detail elsewhere. The linguistic (inchlding tile syutactic) knowledge about source altd target languages is used by the nlechanisnls that translate texts into aud from the btterlingua. Since interlingua is an artificial langnage, we can (and do, through TII,) control tile syntax and semantics of the allowed interlingua elements. The interlingua, snggesled for TRANSI.ATOR has a ln'oader coverage than other knowledge re, presentation schemata for natural language. It involves the knowledge about discourse, speech acts, focus, thne, space and other facets of the overall meaning of texts.</Paragraph>
    <Paragraph position="1"> to Delimiting file Problem.</Paragraph>
    <Paragraph position="2"> TRANS/,AfOR explores the knowledge based apln'oach to machine translation. &amp;quot;File basic translation strategy is to extract nleaniug froul tile inlmt text in source language, SL, represent this nmaning hi a language iudependeut senlantic representation and tlmn render this meauh~g in /, tw'get language, TI,. The knowledge representation language used in such a set-up is called, for historical reasons, interlingua (henceforth, ILl.</Paragraph>
    <Paragraph position="3"> TRANSLATOR'S ultima~ ainl is achieving good quality an/el/latin translation in n non-trivial snbworld and its corresponding sublangnage.</Paragraph>
    <Paragraph position="4"> The philosophy of 'rltANSI.ATOR ailns at tile independence of tile process of trauslafion froln human intervention in tile fcnnl of the traditional preand/or post-editing, hlteraction during tit/,* process of tra~lslation can be accommodated by this philosophy, but only as a temporary measure.</Paragraph>
    <Paragraph position="5"> Interactive modules will be phlgged into the system pendhlg the develop ment of autonlatic modules for perfbrnling tile various tasks as well as more powerful inference engines and representation schemata. This is a device that facilitates early testing of a system Even tlefbre all the modules are actually built. Another advantage of this strategy is that the systnlu becomes 'dynamic', in the sense that its knowledge is growing with use.</Paragraph>
    <Paragraph position="6"> This strategy is an exteusion of one of the approaches discussed, for example, in Carbonell and Tomita (1985) since it implies knowledge acquisition during the exploitation stage a/~d also involves a broader class of texts as its inlnlt. Johnson and Whitelock (1985) are also proponents of the interactive approach, lint their motivation is different, in that they perceive the human to be an integral part of their system even in its final incarnation. In any case, interactlvity is not tile central design feature of TRANSLATOR.</Paragraph>
    <Paragraph position="7"> Before proceeding to describe the knowledge chlsters in TItANSI.A..</Paragraph>
    <Paragraph position="8"> &amp;quot;fOR we would like to colnnlent very briefly on a number of methodologi cal points concerning MT research. It seems that some of file opinions more or less commonly hekl by some members of the MT con/munity may need rethinking. In what follows we list some of these opinions, together with our comments. A more detailed treatment of these topics will be given elsewhere.</Paragraph>
    <Paragraph position="9"> l- Thin paper is based upon work suptx~rted by the National Science Foundation under Grant DCR-8407114.</Paragraph>
  </Section>
  <Section position="3" start_page="0" end_page="627" type="metho">
    <SectionTitle>
* Colgate University
** Purdue University
</SectionTitle>
    <Paragraph position="0"> Opinion. It is nnnecessary to extract tile full meaning from the SL text in order to achieve adequate MT.</Paragraph>
    <Paragraph position="1"> Conunent, An MT system can do well withont (involved) semantics in nlany cases, bnt has to USE meaning in tile rest (or rely on hnlnau interyen/ion). Machines, unlike humans, cannot on demand prodnce interpretations of text at all arbitrary depth sufficient for understanding. Therefore, if one aims at fully automatic, one has to prepare tile system for tile treatment of even very semantically involved text. One Call, of course, think of designing a systenl that can decide how deeply each sentence can be analyzed semantically in an atlempt to minimize selnantic analysis. We maintain that tile decision nlaking involved is as complex as the initial problem of deep senlantic analysi:;.</Paragraph>
    <Paragraph position="2"> Opiniondeg II is not necessary to finish lnocessing the inlmt sentence before starting the translation. Indeed, people very often do this (consider interpleu~.rs) with very good re, suits.</Paragraph>
    <Paragraph position="3"> {2nlnment. This Opinion is based on iutrospEcdon. The \[eal thought processes that gt, on ill tile trans\[atols' or thE interprEtErs' heads are uot known. The (quite considerable) knowledge that the translators i/ave about the subject of the text (speech) and about tilt: speech situation itself prorupts them to preempt the text by following their expectations concerning the most probable set of meanings fbr the tcxl and deciding before tile final eorloboration arrives, biveu if' hi a majority of cases this strategy works (as it is supposed to, because otherwise humans, being intelligent creatures as thi:y are, Would not have had tile above expectations in the first place!), them is nothing unusual in making an crier of&amp;quot; judgenlent. Those ot us who worked as translators surely remeulbct nmltiple instances of this kind. Of course, tills disEussion is relative to the quality of product desired in tile Iranslatioii.</Paragraph>
    <Paragraph position="4"> Opiniondeg Apln'oaches to MT based on AI do not pay sufficient attention to the syntactic analysis of SI,, while syntactic information is important for MT.</Paragraph>
    <Paragraph position="5"> Cllnlnlexll. Syntactic structure of inpnt conveys meaning; this nmanmg is extracted by the semantic analyzer with the help of syntactic knowledge.</Paragraph>
    <Paragraph position="6"> All clues are indeed used. No resnlts of' syntactic analysis are, storEd because they are not needed. Any approach that attempts to relate directly various syntactic slyucturn trees between SE, lad T\]~, strikes us as quite nnpromising. It is only some early Al-otiented MT systems that were vuh/erable to this criticism.</Paragraph>
    <Paragraph position="7"> Opiniondeg lL-based approaches Inad to an overkill because no peculia,i.</Paragraph>
    <Paragraph position="8"> ties of SI, (and of the relationship between, or contrastive knowledge of, SI, and TI,) can be used in translation. Some languages have quite a lot in conunon in their syntax and meauing dislributiou. It is wastefltl not to USE this additional infbrmatiou iu translatiou.</Paragraph>
    <Paragraph position="9"> Comment. While snch insights cau sometimcs bE detected and nsed, hies/ of them comes fronl h/uuan intnitinu, and cannot be taken advautage of in an MT systeel, which can hardly he considered a model of human performance. It is also totally wrong to imply, in our opinion, that discovery and implenlentation of those pieCES of contrastive knowledge can be simpler or, in fact, distinct from invoiw?d semantic analysis.</Paragraph>
    <Paragraph position="10"> Opinion. With l\[,, the process of translation beconms one of interpretatiou, The structure of the SL text, whert used in addition to It, in MT, governs tile choice of one of tile paraphrases. Moreovm, again, II, is an overkill, because tile paraphrases are not needed and add an elemeut of ambiguity.</Paragraph>
    <Paragraph position="11"> Comment. thnnan translators always have a few practically Equally acceptable paraphrases for virtually every St. sentence. The degree of meaning similarity among the acceptable paraphrases is determined by external parameters. The translation is executed according to the human translator's intuitive understanding of these parameters. Only in II.</Paragraph>
    <Paragraph position="12"> approaches can one control tile required degree of sinlilarity among the acceptable paraphrases as la'anslatious of all SL sentence.</Paragraph>
    <Paragraph position="13"> Opiniou. Generation of TL is a relatively simple problem for which very little or no knowledge other than lexical or syntactic is needed.</Paragraph>
    <Paragraph position="14">  Cmnment. Generation requires non-trivial decision making, for instance, in the light of tim discussion in the previous paragraph, or, for that matter, as regards the computational stylistics, which will have to be a part of the choice-making mechanisms in building TL texts.</Paragraph>
  </Section>
  <Section position="4" start_page="627" end_page="627" type="metho">
    <SectionTitle>
2. Configuration of TRANSLATOR
</SectionTitle>
    <Paragraph position="0"> The background of the TRANSLATOR MT project at Colgate is presented in Tucker and Nirenburg (1984). This paper focuses on the static knowledge clusters of TRArqSLA'roR. The latter are identified as follows: null  There are also dynamic knowledge clusters in TRANSLATOR: tile parser and the generator modules as well as the inferencing mechanism (known as the Inspector) used to derive additional knowledge from IL representations when troubleshooting becomes necessary.</Paragraph>
    <Paragraph position="1"> In this paper we will describe the structure of the IL dictionary and the IL grammar, the central components of the system. These two structures are actually knowledge representation languages. IL dictionary is written in a language for describing tim types of concepts that can appear in the subworld of translation. IL grammar is written in a language for representing the assertions about tokens of those types that actually appear in texts. We will call these languages DIL (for Dictionary Interlingua) and TIL (Text Interlingua), respectively. The distinction between DIL and TIL is similar, for instance, to that between the description and the assertion languages in KL-ONE (cf., e.g., Brachman and Schmoltze, 1985). After discussing these languages we will briefly discuss the structure of knowledge about SL (the SL grammar and the SL - IL dictionary), enough only to help us through an illustration of how the IL dictionary and grammar are used.</Paragraph>
  </Section>
  <Section position="5" start_page="627" end_page="629" type="metho">
    <SectionTitle>
3. The IL Dictionary.
</SectionTitle>
    <Paragraph position="0"> The IL dictionary serves as the database where TRANSLATOR stores its knowledge about rite subworld of translation. It is purely semantic, conceptual. The IL dictionary is a source of information for representing the meanings of SL texts. In it one does not find any information pertaining to any particular SL or TL. Thus, it is pure coincidence that most of the entry heads in this dictionary, as well as most of the members of the property sets (cf. below) look like English words. This choice was made with the dictionary writers in mind. The other possibility would have been to assign non-suggestive identifiers to entries and values in the IL dictionaries. This would have slowed doffn the process of dictionary compilation. The dictionary writers must do their best not to mix the semantics of an IL dictionary entry with that of an English word whose graphical form coincides with that of the IL dictionary entry head.</Paragraph>
    <Paragraph position="1"> There are two kinds of entities in DIL: concepts and properties.</Paragraph>
    <Paragraph position="2"> Concepts are IL 'nouns' (objects) and IL 'verbs' (events). IL 'adjectives', 'adverbs' and 'numerals' are represented by properties. These are organized as sets of property values indexed both by the name of the property set (e.g., 'color', 'time' or 'attitude') and by the individual values, to facilitate retrieval. Property values are applicable to specific concept types. Their tokens do not appear on their own in IL texts, but only as fillers of slots in the frames for concept tokens. Thus, for example, 'red' will be a potential filler for the 'color' property of a token of every physical object. An explanation of the relationship between IL word types and tokens follows.</Paragraph>
    <Paragraph position="3"> The IL dictionary is organized as a set of entries (concept nodes) interconnected through a number of link types (properties). However, the structural backbone of the dictionary is the familiar isa hierarchy with property inheritance. Note that most of the time the translation system will be working with terminal nodes in this hierarchy. But the nonterminal nodes play a special role in it. By representing sets of entries, thereby providing a link among a number of (related) concepts, they serve as the  basis for a variety of inference-making procedures. Even more importantly, these 'nonterminal entries' constitute, together with tile sets of various property values, the schema of the dictionary, the set of terms that arc used to describe the semantics of the rest of the dictionary entries.</Paragraph>
    <Paragraph position="4"> Just as all other nodes in the hierarchy, nonterminal nodes represent dictionary entries, which means that they can also have tokens.</Paragraph>
    <Paragraph position="5"> This device comes bandy when, on analyzing a segment of input, we con ciude that a certain slot filler is unavailable in the text. At the same time, if we know the identities of other slot fillers in the frame, we can come to certain conclusions about the nature of an absentee. For instance, if the Agent slot of a certain mental process is not filled, we, by consulting the 'agent-of' slot of the nonterminal node 'mental-process', can infer (or, rather, abduce) that, whatever it is, it must be a 'creature'. This knowledge helps in finding referents for anaphoric phenomena.</Paragraph>
    <Paragraph position="6"> The dictionary entries represent IL concept and property types; IL texts consist of IL concept tokens (as well as \[L clause and sentence tokens). Every token of an IL concept stands in the is-token-of relationship to its corresponding type. Structurally both IL concept types and IL concept tokens are represented as frames. The frame for a type and the frame for a corresponding token are not identical in structure, though the intersection of their slot names is obviously non-zero. One must note, however, that even in this case the semantics of the slots in the dictionary frames is different from that of the corresponding slots in the text frame.</Paragraph>
    <Paragraph position="7"> Some of the slot names in the type frames refer to the paradigmatic relationships of this concept type with other concept types. These are the type parameters of an IL dictionary entry. The rest of the information in an entry describes syntagmatic relationships that tokens of this particular type have with tokens of other types on an IL text. These are called token parcaneters. Among the type parameters one finds the pointers in the isa hierarchy, relationships like part-of, belangs-to, etc.</Paragraph>
    <Paragraph position="8"> The token-parameter slots in the dictionary entries contain either default values for the properties (the 'no-value' value is among the possible default choices) or acceptable ranges of values, for the purpose of validity testing. IL concept tokens, which are components of IL text, not its dictionary, have their slots occupied by actual values of properties; if information about a property is not forthcoming, then the default value (if any) is inherited from the corresponding type representations.</Paragraph>
    <Paragraph position="9"> In what follows we will describe DIL, the IL dictionary language.</Paragraph>
    <Paragraph position="10"> We will do this by presenting the top levels of the isa hierarchy of concepts in our world and listing the frames for high-level nodes. Next, we'll present examples of IL dictionary frames, including one complete path in the isa hierarchy, from the root to a terminal node.</Paragraph>
    <Paragraph position="11"> The actual contents of the tree are, as we already said, idiosyncratic: it may be overdeveloped in some of its branches and underdeveloped in  many others. This state of affairs corresponds to the strategy of working within a subworld.</Paragraph>
    <Paragraph position="12"> 3.1. Frames.</Paragraph>
    <Paragraph position="14"> This is the root of the isa hierarchy. The three slots present here mean that every node in the tree has an id; every node features some properties (which exactly, will be shown in lower-level nodes); and every node represents a concept that belongs to one or more subworlds.</Paragraph>
    <Paragraph position="16"> At this level we meet the 'isa' slot for the first time. This is the,pointer to a node's parent in the hierarchy. Events divide into processes and states.</Paragraph>
    <Paragraph position="17"> The only overtly mentioned property common to all events is the conceptual case of 'patient' (this reflects our opinion that in the sentence (1) John is not an agent, but rather a patient). Note that 'patient' in DIL subsumes the semantics of 'beneficiary'.</Paragraph>
    <Paragraph position="19"> In addition lo the conceptual case slots, the process frame contains information about preconditions and effects. These are states that must typically hold before and after the process takes place, respectively. A process can also be a part of other processes. Thus, for instance, move is a part of travel and, at the same time, of fetch or insert. The 'is' slot of a process frame contains either tile constant primitive, if the process is not furfller analyzable in DIL, or the description of file seqnence of processes which comprise the given process. The process-sequence is a list of process names connected by tile operators sequential, choice and shuffle. In other words, a process may be a sequence of subprocesses (sequential), a choice among several subprocesses (choice), a temporally unordered sequence of subprocesses (shuffle) or any recursive combination of file above. This treatment of processes is inspired by Nirenburg et al., 1985.</Paragraph>
    <Paragraph position="20"> For the purposes of machine translation it seems unnecessary to introduce a more involved t~nporal logic into consideration for the 'is' slot.</Paragraph>
    <Paragraph position="22"> Only creatures can be fillers for the 'agent' slot. Mental objects classify into reaction processes (cf. the English 'please' or 'like'), cognition  processes ('deduce') and perception processes ('see'). Objects of mental processes can be either objects, as in (2) or events, as in (3). (2) I know John (3) 1 know that John has traveled to Tibet.</Paragraph>
    <Paragraph position="24"> Speech processes are primitives. The speech processes recognized by D1L include assertions (that further subdivide into definitions, opinions, facts, promises, etc.) and requests (questions or commands). The 'agent' slot filler has file semantics of the speaker. The 'patient' is the hearer. Note that there is a possibility for the hearer to be a group or an organization, as in (4).</Paragraph>
    <Paragraph position="25"> (4) I promised the band to let them have a ten-minute break every hour.</Paragraph>
    <Paragraph position="26"> The 'agent' is the 'source' and the 'patient' is the 'destination' of a speech process.</Paragraph>
    <Paragraph position="27"> state ::= ('skate' ('isa' event) ('part-of' state*)) The actant in states, which is the patient rather than the actor, is inherited from the event frame.</Paragraph>
    <Paragraph position="29"> The '...-of' slots are used for consistency checks.</Paragraph>
    <Paragraph position="30"> 3,2. Properties.</Paragraph>
    <Paragraph position="31"> Property values are primitive concepts of IL used as values for slots in concept frames. We give here just an illustration of these. Many more exist and will be used in the imphrmentation.</Paragraph>
    <Paragraph position="32">  size-set :: = nil I infinitesimal \[ ... I huge color-set :: = nil t black \] ... \[ white shape-set :: : nil \] flat I square \] spherical ... material-set :: = nil I (gold (specific-gravity 81) (unit-value 228))1 ... subworld-set :: = nil I computer-world \[ business-world \[ everyday world boolean-set :: = nil I yes \] no texture-set :: = nil I smooth \] ... \[ rough properties :: = ('properties'</Paragraph>
    <Paragraph position="34"> 3.3. From the Root to a Leaf.</Paragraph>
    <Paragraph position="35"> A path of concept representations fi'om the root to a leaf node is presented below.</Paragraph>
    <Paragraph position="36"> all-&gt; object-&gt; pobject-&gt; +alive-&gt; creature-&gt; person-&gt; computer-user Frames for 'all' and 'object' see above.</Paragraph>
    <Paragraph position="38"> The '+' sign in slots means all inherited information plus the contents of tile current slot.</Paragraph>
    <Paragraph position="40"> The complete frame of the leaf of this patb, 'computer-user', including all inherited slots and default values is listed below. In reality frames like tMs do not exist, because the tokens of this type do not contain all the possible slot fillers.</Paragraph>
    <Paragraph position="42"/>
  </Section>
class="xml-element"></Paper>