File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/93/e93-1039_metho.xml
Size: 21,914 bytes
Last Modified: 2025-10-06 14:13:18
<?xml version="1.0" standalone="yes"?> <Paper uid="E93-1039"> <Title>apos;Structure and Usage of Current Bell Labs Text</Title> <Section position="5" start_page="332" end_page="332" type="metho"> <SectionTitle> (3) FUNCTIONAL APPLICATION: </SectionTitle> <Paragraph position="0"/> <Paragraph position="2"> CCG extends this strictly context-free categorial base in two respects. First, all arguments, such as NPs, bear only type-raised categories, such as S/(S\NP). Similarly, all functions into such categories, such as determiners, are functions into the raised categories, such as (S/(S\NP))/N. For example, subject NPs bear the following category in the full notation:</Paragraph> <Paragraph position="4"> The derivation of a simple transitive sentence appears as follows in the abbreviated notation: 1 Second, the combinatory rules are extended to include functional composition, as well as application. The following rule will be relevant below:</Paragraph> </Section> <Section position="6" start_page="332" end_page="334" type="metho"> <SectionTitle> (6) FORWARD COMPOSITION (>B): X/Y Y/Z ~B X/g </SectionTitle> <Paragraph position="0"> This rule allows a second syntactic derivation for the above sentence, as follows: 2 using the full semantic notation that this derivation yields an S with the correct interpretation include' sprockets' widgets'. At first glance, it looks as though type-raising will expand the lexicon Marmingly. One way round this problem is discussed in \[14\]. The reasons for making this move, which concern the grammar of coordinate constructions, the general class of rules from which the composition rule is drawn, and the problem of processing in the face of such associative rules, are discussed in tile earlier papers, and need not concern us here. The point for present purposes is that the partition of the sentence into the object and a non-standard constituent S : include' z' widgels'/NP : z makes this theory structurally and semantically perfectly suited to the demands of intonation, as exhibited in example (1). 3 We can therefore directly incorporate intonational constituency in syntax, as follows (cf. \[12\], \[13\], and \[15\]). We assign to all constituents an autonomous prosodic category, expressing their potential for combination with other prosodic categories. Then we lock these two structural systems together via the following principle, which says that syntactic and prosodic constituency must be isomorphic: (S) PROSODIC CONSTITUENT CONDITION: Combination of two syntactic categories via a syntactic combinatory rule is only allowed if their prosodic categories can also combine via a prosodic combinatory rule.</Paragraph> <Paragraph position="1"> One way to do this is to make the boundaries arguments and the pitch accents functions over them. The boundaries are as follows: 4</Paragraph> <Paragraph position="3"> As in CCG, categories consist of a structural type, here b for boundary, and an interpretation, associated via a colon. The pitch accents have the following functional types: 5</Paragraph> <Paragraph position="5"> We further assume, following Bird \[2\], that the presence of a pitch accent causes some element(s) in the translation of the category to be marked as focussed, a matter which we will for simplicity assume occurs at the level of the lexicon. For example, when includes bears a pitch accent, its category will be as follows: (11) (S:(*include')xYkNP:y)/NP:x The categories that result from the combination of a pitch accent and a boundary may or may not constitute entire prosodic phrases, since there may be a prenuclear null tone. There may also be a null tone separating the pitch accent(s) from the boundary.</Paragraph> <Paragraph position="6"> aA similar argument in a related categorial framework is made by Moortgat \[6\].</Paragraph> <Paragraph position="7"> 4These categories slightly depart from Pierrehumbert. 5 Here we are ignoring the possibility of multiple pitch accents in the same prosodic phrase, but cf. \[13\].</Paragraph> <Paragraph position="8"> (Both possibilities are illustrated in (1)). We therefore assign the following category to the null tone, which can thereby apply to the right to any non-functional category of the form X : Y, and compose to their right with any function into such a category, including another null tone, to yield the same category: null</Paragraph> <Paragraph position="10"> It is this omnivorous category that allows intonational tunes to be spread over arbitrarily large constituents, since it allows the pitch accent's desire for a boundary to propagate via composition into the null tone category (see the earlier papers).</Paragraph> <Paragraph position="11"> In order to allow the derivation to proceed above the level of complete prosodic phrases identifying themes and rhemes, we need two unary categorychanging rules to mark the interpretation a of the corresponding grammatical category with that discourse function and change the phonological category, thus: 6</Paragraph> <Paragraph position="13"> These rules change the prosodic category either to p, or to an endocentric function over p. (These types capture the fact that the LL% boundary can only occur at the end of a sentence, thereby correcting an overgeneration in the version of this theory in Steedman \[13\], noted by Bird \[2\]). The fact that p is an atom rather than a term of the form X : Y is important, since it means that it can combine only with another p. This is vital to the preservation of the intonation structure/ The application of the above two rules to a complete intonational phrase should be thought of as precipitating a side-effect whereby a copy of the category E is associated with the clause as its theme or rheme.</Paragraph> <Paragraph position="14"> (We gloss over details of how themes and rhemes are associated with a particular clause, as well as a number of further complications arising in sentences with more than one theme).</Paragraph> <Paragraph position="15"> In \[13\] and \[15\], a related set of rules of which the present ones form a subset are shown to be well-behaved with a wide range of examples. Example (15) gives the derivation for an example related to (7) (since the raised object category is not crucial, it has been replaced by NP to ease comprehension): s Note that it is the identification of the theme and ther composition into the null tone achieved in the earlier papers by a restriction on forward prosodic composition. SNote the focus-marking effect of the pitch accents.</Paragraph> <Paragraph position="16"> (15) Widgets include sprockets ( L+H* LHT, ) ( H* LLT~ ) ......................................</Paragraph> <Paragraph position="17"> S: s/(S: s\NP: *eidget ~) (S: include ~ x y\NP: y)/NP: x NP: *sprockets ' p: theme/b: lh b: lh P: rheme S: include ' * ,eldget ~/NP : * p: theme rheme at the stage before the final reduction that determines the information structure for the response, for it is at this point that discourse elements like the open proposition are explicit, and can be used in semantically-driven synthesis of intonation contour directly from the constituents.</Paragraph> <Paragraph position="18"> Of course, such gushingly unambiguous intonation contours are comparitively rare in normal dialogues. Even in the context given in (7), a more usual response to the question would put low pitch - that is, the null tone in Pierrehumbert's terms - on everything except the focus of the rheme, sprockets, as in the following: (16) Widgets include SPROCKETS Such an utterance is of course ambiguous as to whether the theme is widgets or what widgets include. The earlier papers show that such &quot;unmarked&quot; themes, which include no pitch accent because they are entirely background, can be captured by a &quot;Null Theme Promotion Rule&quot;, as follows: 9</Paragraph> </Section> <Section position="7" start_page="334" end_page="334" type="metho"> <SectionTitle> 3 Parsing </SectionTitle> <Paragraph position="0"> Having established the relationship between prosody, information structure and CCG syntax, we can now address the computational problem of automatically directing the synthesis of intonation contours for responses to database queries. Our computational model (shown in Figure 1) starts with a prosodically annotated wh-question given as a string of words with associated Pierrehumbert-style pitch accent and boundary markings. We employ a simple bottom-up shift-reduce parser of the kind presented in \[14\], making direct use of the CCG-Prosody theory described above, to identify the semantics of the question. The 9See the next section concerning the nondeterminism inherent in this rule.</Paragraph> <Paragraph position="1"> inclusion of prosodic categories in the grammar allows the parser to identify the information structure (theme and theme) within the question as well. The focus and background information within the theme and theme (if any) is further marked by the focus predicate * in the semantic representation. For example, given the question (18) below, the parser produces the semantic and information structure representations shown in (19). 1deg (18) I know that widgets contain cogs, but what parts do WODGETS include?</Paragraph> <Paragraph position="3"> The nondeterminism inherent in unmarked themes is handled by default: the present implementation of Null Theme Promotion delivers the longest unmarked theme that the syntax permits. 11</Paragraph> </Section> <Section position="8" start_page="334" end_page="337" type="metho"> <SectionTitle> 4 Strategic Generation </SectionTitle> <Paragraph position="0"> The strategic phase of generating a response is somewhat simplified in the current implementation, and we have cut a number of corners. In particular, we currently assume that the question is the sole determinant of the information structure in the answer. This is undoubtedly an oversimplification. The complete specification of the semantic and information structures provided by the parser is used by the generator to determine the intelligible and prosodically natural response. For a wh-question, the semantic representation corresponds to a lambda expression in one or more variables ranging over individuals in the database, and has the structure of a Prolog query which we can evaluate to determine the possible instantiations of the open proposition. The instantiated proposition determines the semantic proposition to be conveyed in the response. For the example above, this is part(sprockets)&include( .wodgets, .sprockets) &quot;Wodgets include sprockets&quot;.</Paragraph> <Paragraph position="1"> Note that the derived semantics includes the necessary occurrences of the focus predicate *, determined as follows. All terms that are focused in the question semantics are focused in the response semantics. Intuitively, the instantiated variable in the response semantics must also be focused since it represents the information which is new in the response. For more complex rhemes such as quantified NPs with modifiers, we focus those elements of the semantic representation that are new in the current context. (That is, ones which did not figure in the interpretation of the original query). Thus, given a question such as (1), we choose to focus the modifier &quot;fast&quot; rather than the noun &quot;processor&quot; in the rheme. Similary, in the exchange below, we focus &quot;processor&quot; instead of &quot;fast&quot; because of its newness in the context. (20) Q: What fast component does the widget include? A: The widget includes a fast PROCESSOR. null To determine an appropriate intonation contour for the utterance, we must further determine the appropriate information structure. Fortunately, for the simple question-answering task, the information structure of the response can be assumed to be completely determined by the original query. The theme of a question corresponds to &quot;what the question is about&quot; - in this case, &quot;parts&quot;. The theme of a question corresponds to &quot;what the speaker wants to know about the theme&quot; - here, &quot;What wodgets include&quot;. It follows that we expect the theme of the question to determine the theme of the response. For example (18), the theme of the response should be S : include(.wodgets, x)/NP : x, as in (21) below.</Paragraph> <Paragraph position="2"> Note that we simplify the strategic generation problem by including the syntactic category in our represention of the theme (as determined by the syntactic category of the theme of the original question)) 2 Given the syntactic and semantic representation of the theme of the response, the CCG combination rules can easily be invoked to determine the theme of the response. The rheme is simply the complement 12Here we are cutting another corner: the theme, and hence the rheme, are fully specified syntactically, as well as semantically, as a result of the analysis of the question: in a more general system, we would presumably need to specify syntactic type from scratch, starting from pure semantics.</Paragraph> <Paragraph position="3"> of the theme with respect to the overall semantics of the response, as in (21) below, obtained by instantiating the result and one input of the appropriate combinatory rule (cf. \[7\]): la (21) prop: s: include(,wodgets, *sprockets) theme: s : include(.wodgets, x)/np : x theme: np : ,sprockets 5 Tactical Generation and CCG Just as the shift-reduce parser sketched above can readily be made to construct the interpretations and information structures shown in the examples, specifically marking themes, rhemes and their foci, so it is relatively easy to do the reverse---to generate prosidically annotated strings from a focus-marked semantic representation of themes and rhemes.</Paragraph> <Paragraph position="4"> For simplicity, we start by describing the syntactic and semantic aspects of the generator, ignoring prosody for the moment. In constructing a tactical generation schema, several design options are available, including bottom-up, top-down and semantic head-driven models (\[3\], \[10\]). We adopt a hybrid approach, employing a basic top-down strategy that takes advantage of the CCG notion of &quot;functional head&quot; to avoid fruitless search. While this technique exhibits some inefficiencies characteristic of a depth-first search, it has several significant advantages. First, it does not rely on a specific semantic representation, and requires only that the semantics be compositional and representable in Prolog. Thus the generating procedure is independent of the particular grammar. This modular character of the system has been very useful in developing the competence grammar proposed in the preceding section, and offers a basis for proving the completeness of the implementation with respect to the competence theory.</Paragraph> <Paragraph position="5"> The tactical generation program is written in Prolog, and works as follows. Starting with a syntactic constituent (initially s) and a semantic formula, we utilize the CCG reduction rules to determine possible subconstituents that can combine to yield the original constituent, invoking the generator recursively to generate the proposed subconstituents. The base case of the recursion occurs when a category we wish to generate unifies with a category in the lexicon.</Paragraph> <Paragraph position="6"> For example, suppose we wish to generate an utterance corresponding to the category s:walks'(mary'). Since the given category does not unify with any category in the lexicon, the program proposes possible subconstituents by checking the CCG combination rules in some pre-determined order. By the backward function application rule, we might hypothesize that the categories x and s:walks'(mary')\z are the subconstituents of s:walks'(mary'), where x is 13Again the example is simplified by the use of a nonraised category for the object.</Paragraph> <Paragraph position="7"> (22) gen(s:def(x, ((engine(x)~new(x))&shiny(x))~ def(y, ((gear(y)&rotating(y))&largest(y))~contains(x,y)))). RESULT: the shiny nee engine contains the largest rotating gear. (23) genCs:exists(z, (engineerCz)~brilliantCz))kexists(x,(design(x)~revolutionary(x))& def(y, (engine(y)~new(y))~gave(z,y,x))))).</Paragraph> <Paragraph position="8"> RESULT: a brilliant engineer gave the nee engine a revolutionary design. (24) gen(s:def(x,(widget(x)~*new(x))~probably(contains(x,y)))/np:y @ p:theme). RESULT: the new@lhstar vidget probably containsQlhb.</Paragraph> <Paragraph position="9"> (25) gen(np:(x's)'def(x,(processor(x)&*fastest(x))ls) @ ph:rheme). RESULT: the fastest@hstarprocessor@llb.</Paragraph> <Paragraph position="10"> (26) gen(s:def(x,(widget(x)&new(x))& *probably(contains(x,y)))Inp:y @ p:theme). RESULT: the new widget probably@lhstarcontains@lhb.</Paragraph> <Paragraph position="11"> (27) gen(s:def(x,(*widget(x)~new(x))~contains(x,y))/np:y @ p:theme). RESULT: the new widgetQlhstar contains@lhb.</Paragraph> <Paragraph position="12"> some variable. If we recursively call the generator on s:walks'(mary')kx, we find that it unifies with the category s:walks'(y}knp:y in the lexicon, corresponding to the lexical item walks. This unification forces the complementary category z to unify with np:mary', which yields the lexical item mary when the generator is recursively invoked. Concatenating the results of generating the proposed subconstituents therefore gives the string &quot;Mary walks.&quot; The top-down nature of the generation scheme has a number of important consequences. First, the order in which we generate the postulated subconstituents determines whether the generation succeeds. Had we chosen to generate x before s:walks'(mary'}kx, we would have entered a potentially infinite recursion, since x unifies with every category in the lexicon. For this reason, our generator always chooses to recursively generate the sub-constituent that acts as the functional head before the subconstituent that acts as the argument under the CCG combinatory rules. By strictly observing this principle, we ensure that as much semantic information as possible is deployed, thereby constraining the search space by prohibiting spurious unifications with incorrect items in the lexicon. For this reason, we refer to our generation scheme as a &quot;functional head&quot;-driven, top-down approach.</Paragraph> <Paragraph position="13"> One disadvantage of the top-down generation technique is its susceptibility to the non-termination problem. If a given path through the search space does not lead to unification with an item in the lexicon, some condition which aborts the path in question at some search depth must be imposed.</Paragraph> <Paragraph position="14"> Note that whenever the CCG function application rules are used to propose possible subconstituents to be recursively generated, the subconstituent acting as the functional head has one more curried argument than its parent. Since we know that in English there is a limit to the number of arguments that a functional category can take, we can abort fruitless search paths by imposing a limit on the number of curried arguments that a CCG category can possess. The current implementation allows categories with up to three arguments, the minimum needed for constructions involving di-transitive verbs. Note that this strategy does not prohibit the generation of categories whose arguments themselves are complex categories. Thus, we allow categories such as ((s\np)/np)\(((s\np)/np)/np) for raised indirect objects, but not categories such as (((s\ np}/np)/np)/np.</Paragraph> <Paragraph position="15"> When the CCG composition rule is used to propose possible subconstituents, the subconstituents do not have more curried arguments than their parent. Consequently, imposing a bound of the type described above will not necessarily avoid endless recursion in all cases. Suppose, for example that we wish to generate a category of the form s/x, where s is a fully instantiated expression and x is a variable. If the function application rules fail to produce sub-constituents that generate the category, we rely on the CCG composition rule to propose the possible subconstituents s/y and y/x. Since s/x and s/y are identical categories to within renaming of variables, the recursion will continue indefinitely. We rectify this situation by invoking the composition rule only if the original category has an instantiation for both its argument and result. Such a solution imposes limitations on the types of derivations allowed by the system, but retains the simplicity and transparency of the algorithm. Merely imposing a limit on the depth of the recursion provides a more general solution. Examples of the types of sentences that can be generated appear in (22) and (23).</Paragraph> <Paragraph position="16"> This procedure can immediately be applied to the prosodically augmented grammar. To do so, we merely enforce the Prosodic Constituent Condition at each step in the generation. That is, whenever a pair of subconstituents are considered (by reversing the CCG combination rules), a pair of prosodic subconstituents are also considered and recursively generated using the prosodic combinatory rules. Examples (24) and (25) illustrate the generation of intonation for the theme and theme of the utterance &quot;The NEW widget probably contains the FASTEST processor&quot; .14 Examples (26) and (27) manifest the intonational results of moving the thematic focus among the various propositions in the semantic representation of the theme &quot;The new widget probably contains... &quot;.</Paragraph> </Section> class="xml-element"></Paper>