File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/86/c86-1059_metho.xml
Size: 24,114 bytes
Last Modified: 2025-10-06 14:11:48
<?xml version="1.0" standalone="yes"?> <Paper uid="C86-1059"> <Title>l)isambiguation and Language Acquisition through the Phrasal Lexicon * Abst,'act</Title> <Section position="3" start_page="0" end_page="247" type="metho"> <SectionTitle> 11.2 The Modeled Phenomenon </SectionTitle> <Paragraph position="0"> The program R1NA \[7mmik85c\] is designed to parse English sentences by reading text and producing tile corresponding concepts in the context. However, RINA's lexicon is incomplete, and unknown phrases in the text might be encountered and * This research wa',; supported in part by a grant fiom the rl'A Fotmdation. processed. RINA's objective is to acquire these phrases from examples in context. Thereafter, RINA's lexicon, augmented by these dynamically acquired phrases, facilitates parsing and generation of further sentences. In the scenm'io below, RINA encounters the new phrase to throw the book: User: The mobster eluded prosecution for years.</Paragraph> <Paragraph position="1"> Last month, they threw the book at him for income-tax evasion.</Paragraph> <Paragraph position="2"> RINA is familiar with the words throw and book; however she does not know the figurative phrase itself. RINA engages in a dialog with a user in order to acquire that phrase.</Paragraph> <Paragraph position="3"> RINA: The prosecutor propelled a book at him? User: No. The Judge threw the book at him.</Paragraph> <Paragraph position="4"> RINA: The Judge threw the book at him. He found him guilty.</Paragraph> <Paragraph position="5"> Initially, RINA attempts to interpret the text using the literal phrase to throw an object at a person. When this interpretation fails, RINA forms a hypothesis about the new phrase. At each stage, RINA spells out her new hypothesis to let the user know her state of knowledge and provide counterexamples if necessary. As the user wovides examples, RINA's hypothesis about the new phrase is upgraded. Later on, RINA is able to use this phrase in parsing further sentences: User: The dean of the school threw the book at John.</Paragraph> <Paragraph position="6"> RINA: He punished him.</Paragraph> <Paragraph position="7"> Four issues must be addressed in modeling this behavior.</Paragraph> <Paragraph position="8"> (1) The syntax of the new phrase. How can the program determine the scope and variability of the new phrase? For example, should the new phrase for throw the book at him accept the sentence throw a volume at him? Is the prepositional phrase for income-tax evasion incorporated as a mandatory part of the new phrase? (2) Literal interpretation. What is the contribution of senses of single words in forming the meaning of a new phrase? Would the learner come up with the same meaning if he heard the hypothetical phrase the judge threw the book to him? (3) The context. What is the contribution of the context in forming both the syntax and the meaning of the new phrase? (4) Disambiguation. Once acquired, what facilitates the selec null tion of appropriate meanings for lexical phrases? For example, how can the program come up with two different meanings for the following pair of sentences: S2: The Judge decided to throw the book at him.</Paragraph> <Paragraph position="9"> $3: The judge decided that Mary threw the book at him.</Paragraph> <Section position="1" start_page="247" end_page="247" type="sub_section"> <SectionTitle> 1.3 The Program </SectionTitle> <Paragraph position="0"> The program consists of four components: (1) Phrasal lexicon: This is a list of phrases where each phrase is a declarative pattern-concept-situation triple. (2) Case-frame parser: In parsing, case-frames \[Carbonel184\] match the text with syntactic and semantic phrase properties. Unification \[Kay79\] accounts for phrase interaction. (3) Pattern Constructor: Learning of phrase patterns is accomplished by analyzing parsing failures. A patternmodification action is associated with each failure situation. (4) Concept Constructor: Learning of phrase concepts is accomplished by a set of strategies which are selected accord null ing to the context.</Paragraph> <Paragraph position="1"> Schematically, the program receives as input a sequence of sentence-context pairs from which it refines its current patternconcept-situation hypothesis. The pattern is acquired from the sentence, while the concept and the situation are derived from the context.</Paragraph> </Section> <Section position="2" start_page="247" end_page="247" type="sub_section"> <SectionTitle> 1.4 The Pattern Representation </SectionTitle> <Paragraph position="0"> Three sample phrasal patterns as they are represented in RINA's declarative lexicon are given below.</Paragraph> </Section> </Section> <Section position="4" start_page="247" end_page="248" type="metho"> <SectionTitle> 2. Phrase Disambiguation </SectionTitle> <Paragraph position="0"> There are syntactic patterns which are not unique to one lexical entry. Therefore, phrases cannot be selected unambiguously merely through syntactic patterns. Thus, ambiguity must be resolved by semantic means. For example, for the pattern run into there are two distinct senses, as shown in the following sentences $4 and $5: $4: For years I tried to locate my high school teacher, but I could not find her address.</Paragraph> <Paragraph position="1"> Last week Iranint0 her in the street.</Paragraph> <Paragraph position="2"> S5: My client was driving carefully. Your client rsnlnto his Mercedes on a red light.</Paragraph> <Paragraph position="3"> These senses are referred to mnemonically as fortuitous-encounter and vehicle-collision respectively, tlowever, also a literal interpretation exists for run into as shown in $6 below: $6: Driving home, I became hungry so I ran into a McDonald's.</Paragraph> <Paragraph position="4"> The meaning of the clause in $6 is produced literally out of its constituents, meaning &quot;move-quickly to the interior of a setting&quot;. Although ambiguity in sentences $4-$6 does not pose any problem for a human reader (who knows English phrases), a computer program (and a second language speaker) might experience difficulty in discriminating the appropriate sense in each case. Ambiguity is resolved in each case by matching phrase situations with the context. The first lcxical phrase depicts fortuitous-encounter: pattern ?x:person run:verb <into ?y:person> situation (act (ptrans (actor ?x) The situation of this phrase is given in terms of phms and goals. It requires that tim context satisfy three points: (1) ?x moves to the proximity of ?y.</Paragraph> <Paragraph position="5"> (2) As a result, ?x is able to see ?y.</Paragraph> <Paragraph position="6"> (3) ?x has not planned that outcome (seeing ?y is t, ncxpected). The concept is given in terms of the act itself (concept is (act-of situation)). The second phrase depicts vekiclecollision: null pattern ?x:vehicle run:verb <into ?y:thing> situation ($vehicle-oollision (vehicle ?x) (thing ?y)) concept (propel (object ?x) (to (location-of ?y))) $vehicle-collision represents the following situation: (2) ?x is involved in a moving script (Scar-riding, $ice-skating, etc.) (2) ?x moves into the location of ?y.</Paragraph> <Paragraph position="7"> (3) As a consequence of (2), a preservation-goal of ?x (citber preserve-health or preserve-social-control) is thwarted. Again, the concept itself points to a single element within the situation. null</Paragraph> <Section position="1" start_page="248" end_page="248" type="sub_section"> <SectionTitle> Partial Matching of Situations </SectionTitle> <Paragraph position="0"> When several lexical patterns match a sentence, the phrase whose situation matches the context is selected. However, the situation prescribed by the phrase does not always exist in the context prior to reading the text. For example, consider the way the situations of the phrases P4 and P5 above match tim context in this paragraph: $4: For years I tried to locate my high school teacher, but I could not find her address.</Paragraph> <Paragraph position="1"> Last week I ran into her in the street.</Paragraph> <Paragraph position="2"> The plan/goal scenario established in tim context prior to the reading of the phrase in the sentence is: (1) There is an active goal to see the teacher.</Paragraph> <Paragraph position="3"> (2) No plan is available to achieve this goal.</Paragraph> <Paragraph position="4"> This context does not match the situation of vetticle-collision since there is no indication of an accident (riding a car on the one hand, or accident repurcussions on the other hand). Itowever, tbe other situation is not perfectly matched either (i.e. tile outcome of the encounter is not in the context). Thus, phrase selection must be accounted for by partial matching. The best match is selected-fortuitous-encounter in this example.</Paragraph> </Section> </Section> <Section position="5" start_page="248" end_page="249" type="metho"> <SectionTitle> 3. Phrase Interaction </SectionTitle> <Paragraph position="0"> Besides the surrounding context, phrase disambiguation is also influenced by phrase constituents. We describe this aspect in two cases. Interaction of a phrase with its cornplement and interaction of a phrase with an embedded reference.</Paragraph> <Section position="1" start_page="248" end_page="248" type="sub_section"> <SectionTitle> 3.1 Indirect References-the Utility of Selectional Restric- tions </SectionTitle> <Paragraph position="0"> The selection of a phrase sense could rely on an embedded reference. For example, consider the sentence: 87: AyellowMercedes ran into my car oi'~ a red light.</Paragraph> <Paragraph position="1"> being read relative to two lexical phrases P4 and t'5 (fortuitousencounter and vehicle-collision respectively): P4: ?x:person run into ?y:person PS: ?x:vehicle run into ?y:thing Due to the type of the reference (i.e., a yellow Mercedes is a vehicle), the phrase P5 (vehicle collision) is selected, iIowever, the type of the reference might lead to an incorrect selection. For instance consider the sentence: $8: Yourclient ran into my ear on a red light.</Paragraph> <Paragraph position="2"> Here the type of the reference your client is l)erson. This type causes the inappropriate selection of P4 (where we know flom the context that P5 is the appropriate selection). Tbis reference is an example of the phenomenon we call indirect reference (see also \[Hershkovits85\] and \[Fanconnier85\] ) Although the speaker in $8 said your client he actually meant your c\].ient's car, which refers to an instance of a car. Thus, the referent is refen'ed to indirectly through another object. Two pairs of examples illustrating indirect references are: S9: She likes horses.</Paragraph> <Paragraph position="3"> Sl0: She likes tuna fisll.</Paragraph> <Paragraph position="4"> Slh John listened toMozart.</Paragraph> <Paragraph position="5"> g12: Usually, I listen to my parents .</Paragraph> <Paragraph position="6"> In $9, the reference horses could mean any activity related to horses such as riding hor.~es or watching horses. On tile other hand, in S10 the activity is probably eating tuna fish. in S 11, the reference Mozart does not refer to the person Mozart himself, rather it refers to the. sound of his music. On the other hand, ill S12 my parents refers actually to my parent's advice. Thus, a reference might refer to an object which merely represents the intended referent.</Paragraph> <Paragraph position="7"> Therefore, categories of lexical items as they appear in the text, or selectional restrictions \[Chomsky65\] present a very weak method for disambiguation, and generally should not provide the only key for phrase selection, in fact, the identity of referents for indirect references inust be resolved through the embedding phrase. The identity of the phrase itself is determined by the context, as shown in the previous section.</Paragraph> </Section> <Section position="2" start_page="248" end_page="249" type="sub_section"> <SectionTitle> 3.2 A Coml)lement-Taldng Phrase </SectionTitle> <Paragraph position="0"> Ambiguity appears also in complement-taking phrases such as ask, promise, instruct~ inform, etc. Consider tile following set of sentences: S13: John asked her if she was having hulch .</Paragraph> <Paragraph position="1"> S14: The judge asked Mary to approach the bench .</Paragraph> <Paragraph position="2"> S15: Mary asked the judge to give her a break .</Paragraph> <Paragraph position="3"> Although they all involve the word ask, in each sentence ask appears in a different sense (query, command, and appeal, respectively). What is the representation of the phrase for ask, and how are these sentences disambiguated? Two extreme approaches to lexical representation are: (1) Maintain only a single lexical entry which contains all the knowledge of the word ask and all of its possible interactions. This model is best presented by the word-expert approach \[Rieger77\]. In this model, general linguistic knowledge is encoded repeatedly in individual words.</Paragraph> <Paragraph position="4"> (2) Maintain a lexical entry for each phrase sense. Clearly this second extreme is not feasible since it is impossible to predict all possible situations in which ask might appear. In RINA, there are two lexical entries, representing two basic senses of ask.</Paragraph> <Paragraph position="6"> is a goal of ?x P7: pattern: ?x ask ?y ?z:infinitive-phrase concept: ?x inform ?y that ?z is a goal of ?x Accordingly, the meaning of a sentence is constructed in two steps: (a) The syntax of the complement determines whether to select P6 or PT. In S13, P6 is selected due to the question form of its complement. In S14 and S15, P7 is selected due to the infinitive form of its complement.</Paragraph> <Paragraph position="7"> (b) The meaning of the phrase is constructed by interpreting the concept relative to the context.</Paragraph> <Paragraph position="8"> The context in both $14 and S15 is the asymmetrical authority relationship. Thus, the asking act in S14 is taken as an authority-decree, while in S15 the same asking act is interpreted as authority-appeal. It is important to distinguish between these two meanings so that the parser can maintain the appropriate expectations in each case.</Paragraph> <Paragraph position="9"> In conclusion, the lexicon must ensure that linguistic clues (such as the syntactic form of the complement) be exploited in constructing the meaning. However, the lexicon should not contain meanings which could be inferred by general world knowledge.</Paragraph> </Section> </Section> <Section position="6" start_page="249" end_page="250" type="metho"> <SectionTitle> 5. Phrase Acquisition </SectionTitle> <Paragraph position="0"> So far, we have assumed the existence of necessary phrases in the lexicon. However, in reality a program might encounter new phrases in the text. Thus, the program must accomplish two objectives: (a) pm'se the text in spite of the unknown element, and (b) acquire the unknown element for future encounters.</Paragraph> <Paragraph position="1"> Consider the situation in which the figurative phrase is first encountered. null User: The mobster eluded prosecution for years.</Paragraph> <Paragraph position="2"> Last month, they threw the book at him for income-tax evasion.</Paragraph> <Paragraph position="3"> RINA: The prosecutor propelled a book at him? User: No. A judge threw the book at him.</Paragraph> <Paragraph position="4"> RINA: The judge threw the book at him. He found him guilty.</Paragraph> <Paragraph position="5"> And later on: User: The dean of the school threw the book at John. RINA: He punished him.</Paragraph> <Paragraph position="6"> There are three stages in the acquisition process: (1) Apply the literal interpretation.</Paragraph> <Paragraph position="7"> (2) Acquire the figurative phrase.</Paragraph> <Paragraph position="8"> (3) Generalize the new phrase beyond the specific context.</Paragraph> <Section position="1" start_page="249" end_page="250" type="sub_section"> <SectionTitle> Literal Interpretation </SectionTitle> <Paragraph position="0"> In the absence of the appropriate phrase in the lexicon, RINA utilizes other available knowledge sources, namely (a) the literal interpretation and (b) the context. The literal interpretation is given by the phrase: pattern ?x:person throw ?y:phys-obj <at ?z:person> concept (act (propel (actor ?x)</Paragraph> <Paragraph position="2"> This phrase describes propelling an object in order to hit another person. Notice that no situation has been specified. General phrases such as take, give, catch, and throw do not have a specified situation since they can be applied in many situations.* The literal inteq3retation fails by plan/goal analysis. In the context laid down by the first phrase (prosecution has active-goal to punish the criminal), &quot;propelling a book&quot; does not serve the prosecution's goals. In spite of the discrepancy, RINA spells out that interpretation, The prosecutor propelled a book at him? to notify the user about her state of knowledge.</Paragraph> <Paragraph position="3"> The Trial Context The user's second sentence, (a Judge threw the book at. him), and specifically the reference a judge, brings in $trial (the trial-script). This script involves five entities: (a) Judge, (b) Prosecutor, (c) Defendant, (d) Alleged-Crime, (e) the Law regarding that crime and its punishment. The script involves a sequence of events: (a) The Prosecutor says (mtrans) his arguments.</Paragraph> <Paragraph position="4"> (b) The Defendant says his arguments.</Paragraph> <Paragraph position="5"> (c) The Judge decides (select-plan) either: (i) Punish (thwarts a goal of) Defendant.</Paragraph> <Paragraph position="6"> (2) Do not punish him.</Paragraph> <Paragraph position="7"> This script is used in forming the elements of the lexical phrase.</Paragraph> <Paragraph position="8"> (a) The phrase situation is taken as the script itself. (b) Thepattern is extracted from the sample sentence. (c) The concept is extracted from the script.</Paragraph> <Paragraph position="9"> Forming the Pattern Four rules are used in extracting the linguistic pattern fl'om the sentence: $13: Last month, they threw the book at him for income-tax evasion.</Paragraph> <Paragraph position="10"> (a) Initially, use an existing literal pattern. In this case, tile ini null * Notice tile distinction between preconditions and situation. While a precondition for &quot;throwing a ball&quot; is &quot;first holding it&quot;, this is not part of the phrase situation. Conditions which are implied by common sense or world knowledge do not belong in the lexicon.</Paragraph> <Paragraph position="11"> (b) Examine other cases in the sample sentenece, and include in the pattern cases which could not be interpreted by general interpretation. There are two such cases: (1) Last month could be interpreted as a general time adverb (i.e.: last year he was still enrolled at UCLA, the vacation startedlast week, etc.).</Paragraph> <Paragraph position="12"> (2) For income-tax evasion can be interpreted as a element-paid-for adverb (i.e.: he paid dearly for his crime, he was sentenced for a murder he did not commit, etc.).</Paragraph> <Paragraph position="13"> Thus, both these cases are excluded.</Paragraph> <Paragraph position="14"> (c) Variablize references which can be instantiated in the con- null text. In our case ?x is the Judge and ?y is the DefEndant. They are maintained as wtriables, as opposed to the other case: (d) FrEEze references which cannot be iaslantiatcd in the context. No referent is fonnd for the reference the book. Therefore, that reference is taken as a frozen part of tile pattern instead of the case ?z :phys-obj.</Paragraph> <Paragraph position="15"> The resulting pattern is:</Paragraph> </Section> <Section position="2" start_page="250" end_page="250" type="sub_section"> <SectionTitle> Forming the Concept </SectionTitle> <Paragraph position="0"> In selecting the concept of the phrase, there arc four possibilities, namely the events shown in Figure 4. The choice of the appropriate one among these four events is facilitated by linguistic clues. AS opposed to the phrase they threw the book to him which implies cooperation between the characters, the phrase they threw the book at him implies a goal conflict between the characters. Since this property is shared among many verbs, it is encoded in the lexicon as a general phrase: pattern ?x:person ?v:verb ?y:physobj <at ?z:person> concept (act (propel (actor ?x) Notice that rather than having a specific root, the pattern of this phrase leaves ont the root of the verb as a variable. Using this phrase concept as a search pattern, the &quot;punishment-decision&quot; is selected f,'om $trial. Thus, the phrase acquired so far is: pattern ?x:person throw <the book> <at ?y:person></Paragraph> </Section> <Section position="3" start_page="250" end_page="250" type="sub_section"> <SectionTitle> Phrase Generalization </SectionTitle> <Paragraph position="0"> Although RINA has acquired the phrase in a specific context, she might hear the phrase in a different context. She should be able to transfer the phrase across specific contexts by generalization. RINA generalizes phrase meanings by analogical mapping. Thus, when hearing the sentence below, an analogy is found between the two contexts.</Paragraph> <Paragraph position="1"> SI6: The third time he caught John cheating in an exam, the professor threw tile book at him.</Paragraph> <Paragraph position="2"> The trial-script is indexed to a general authority relationship. The actions in a trial are explained by the existence of that relationship. For example, by saying something to the Judge, tile Defendant does not dictate the outcome of the situation. He merely informs the Judge with some facts in order to influence the verdict. On the other hand, by his decision, tile Judge does determine the outcome of the situation since he presents an authority. null ThrEe similarities are found between Ihe $trial and the scene involving John and the professor.</Paragraph> <Paragraph position="3"> (a) The authority relationship between ?X and ?y.</Paragraph> <Paragraph position="4"> (b) A law-violation by ?x.</Paragraph> <Paragraph position="5"> (c) A decision by ?x.</Paragraph> <Paragraph position="6"> Therefore, the phrase situation is generalized fiom the specific trial-script into the general authority-decree situation which encompasses both examples.</Paragraph> <Paragraph position="7"> 6. Presupposilions as a Phrase Situallon A message might be conveyed by an utterance beyond its straightforward illocntion. That message, called tile presul)position of the utterance, is described by Keenan (1971) as follows: (see also \[Grice75\] and \[Fauconnier85\] Ch. 3): The presuppositions of a sEntEnce are those conditions that the world must meet in order for the sentence to make literal sense. Thus if some such condition is not met, for some sentence S, then either S makes no sense at all or else it is understood in some nonlitcral way, for example as a joke or metaphor.</Paragraph> <Paragraph position="8"> Despite this definition, presupposition has been studied as a means for generation and propagation of implications \[Gazdm'79, Karttunen79\]. In general, the effort is to compute the part of the sentence which is already giwm, by applying &quot;backward&quot; reasoning, i.e.: from the sentence the king of France is bald determine if indeed there is a king in France, or fro,n the sentence it was not John who broke the glass, determine whether somebody indeed broke the glass. Rather than use presuppositions to develop further inferences, we investigate how presuppositions are actually applied according to Keenan's definition above, namely, in determining appropriate utterance interpretations. In the phrasal lexicon, we equate presul)position with phrase situation.</Paragraph> </Section> </Section> class="xml-element"></Paper>