File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/86/c86-1138_metho.xml
Size: 29,072 bytes
Last Modified: 2025-10-06 14:11:54
<?xml version="1.0" standalone="yes"?> <Paper uid="C86-1138"> <Title>Forward message Smith CMUA and these would have the nesting structure: \[ForwardAction HeadForm FORWARD MsgObj \[MsgObjDesc HeadForm MESSAGE\] MsgRecipientObj \[MaiIAdrOesc HeadForm SMITH llost \[LocationDesc HeadForm CMUA\]\]\] \[ForwardAct ion HeadForm FORWARD MsgObj \[MsgObjDesc HeadForm MESSAGE\] CCRecipientOb.j \[MailAdrDBsc HeadForm S,M I TIt Host \[Locat.i onDesc HeadEorm CMUA\]\]\] \[ForwardAction tleadForn~ FORWARD MsgObj \[MsgObjDesc HeadForm MESSAGE MsgOriginObj \[MailAdrgesc HeadForm SMITlt Itost \[LocationDesc</Title> <Section position="2" start_page="0" end_page="587" type="metho"> <SectionTitle> 2. Problems with Network-based Parsing of Spoken Input </SectionTitle> <Paragraph position="0"> The case for substantial integration of natural language processing with speech recognition is clear, The issue is how to adapt natural language parsing techniques to cope with the special problems of spoken input as described above, Most such adaptation efforts until new have been based on transition network parsing. Essentially, they encode the expectations of the parser in a transition network whose arcs are labelled by syntactic or semantic categories of words or constituents. An input is analyzed by finding a path through the network the4 corresponds to the sequence of words in the input.</Paragraph> <Paragraph position="1"> Constituent labels on arcs are associated with their own subnetworks, and traversing the arc in the top-level network is accomplished by traversing the corresponding subnetwork.</Paragraph> <Paragraph position="2"> Typically, transition net parsers operate by traversing the network from left to right in step with the input, exploring subnetworks in a top.down manner as they go. Well known examples of transition-net parsers include ATN \[14\] parsers (as used in the LUNAR system \[15\]), the RUS parser\[I\], and the parser used in LIFER \[8\]. The HARPY system \[9\] used an integrated network encoding for linguistic and acoustic information.</Paragraph> <Paragraph position="3"> A major problem with transitionmet parsers for speech recognition lies in the difficulty they have in handling input that does not meet their grammatical expectations. Frequently a word may be missing due to acoustic misrecognition or actual omission. If a network is being explored left to right, finding the correct path through the network would then involve skipping over the arc that corresponded te the rnissing word. If simple skipping were all that was involved, the problem might well be tractable, but the problem is compounded by 'the typical n'rilt\]pHcity of possible parses, especk, l!y if the word lattice contains many alternative words for the same speech segment. The method tlsed to detect a non-viable parse in tile search is inability to follow any arc from the current node --. precisely the situation most likely with a missing word, Thus, network parses can no Ioqger cise the standard halting criteria for non.productive (constraint violating) searches. A furi.her compounding of the pi'oblem arises if the word after the missing word allows a spurious arc to be followed from the network node at which the missing word should have been recognized. In this case, it will generally he very hard to find out where the errer really occurred. Other forms of ungrammaticatity, either actually spoken or mia-recoonition artifacts, result in similar problems. \]he absence of consistent word boundalies from the acoustic analysis I~hase complicates things further.</Paragraph> <Paragraph position="4"> Various methods have been tried to adapt network parsing to these problems, including on-demand insertion of extra arcs (e.g. \[13, 12\]). Perhaps the most promising modification for speech input is the replacement of left-to-right tracing techniques by center-out techniques that work from words with high certainty according to the acoustic component \[16\]. However, semantic importance has never been combined with acoustic certainty in selecting these islands.</Paragraph> <Paragraph position="5"> Island growing, attractive in theory, presents serious practical problems for ATN parsers, not the least of which is the requirement of running ATNs from right to left. This method of interpreting the networks, necessary with center-out teehr4ques, fails when tests depend on registers that have not yet been set, No modifications to network-based techniques have been totally successful.</Paragraph> </Section> <Section position="3" start_page="587" end_page="591" type="metho"> <SectionTitle> 3. Semantic Caseframe Parsing </SectionTitle> <Paragraph position="0"> Our approach is quite different from the transition network approach and is derived from recent work at Carnegie-Mellon University by Carbonell, Hayes, and others \[3, 7, 6, 2\] on understanding typed, restricted domain natural language, with a particular concentration on handling ill-formed input. The technique that makes it possible to process sensible but potentially imperfect or incomplete uttere.nces is called semantic caseframe instantiation.</Paragraph> <Paragraph position="1"> Unlike network-based techniques, caseframe methods er~able a parser to anchor its interpretation on the most significant input parts, and to grow its islands of interpretation to the less significant segments. Since the more significant words tend to be longer and therefore more likely to be recognized reliably, the islands of significance are correlated with islands of certainty. In the process, semantic and syntactic expectations generated from the more meaningful parts of the input can be used to discriminate arid hypothesize the meaning o\[ troublesome segments.</Paragraph> <Paragraph position="2"> The essential difference between caseframe and transition network techniques is the level of encoding of the syntactic and semantic information they both use. Caseframe techniques encode the information at a more abstract level and thus are able to interpret it in multiple ways. Network techniques &quot;compile&quot; the information into netwmks at a much lower an d more rigid level, and thus do not have nearly as niuch fleedom in iriterpreting the same knowledge in multiple ways, As we will show, the ability to apply syntactic and semantic information in an interpietive way is the key to the successful integration 0! speech and naturaI language processing.</Paragraph> <Paragraph position="3"> The central notion behind a caseframe is that of a head concept modified by a set of related concepts or cases, bearing well-defined semantic relations to the head concept. The original linguistic concept of a caseframe as first described by Fillmore \[4\], relied on a small set of universally applicable cases. The recent work at CMU adapts this idea to restricted domain situations by allowing specialized cases for each concept related to a head concept.</Paragraph> <Paragraph position="4"> Consider, for instance, the caseframe shown in Figure 1.</Paragraph> <Paragraph position="5"> The notation is that of the casefrLtme speech pa.is~'r described later. Without going into notational dmtails, the caseframe is identified as a verb or clausal caseframe corresponding to the verbs (HeadForms) &quot;forward&quot; or &quot;resend&quot;. It also has four cases: Agent (the person doing the sending), MsgObj (the message being forwarded), MsgRecipientObj (the person the message is beMg forwarded to), and CCRecipientObj (the people who get a copy of the forwarded message). The MsgObj case must be filled (hlstanceOf) by a MsgObjDesc (defined by another caseframe, see below), and the other cases must be filled by a MailAchDesc (the caseframe representing a peison or &quot;mail address&quot;). All the cases are required, except CCRecipientObj, which is optional. In addition, to this pLIrely semantic information, the casefr~me contains sorne ,&quot;,yutu.cUc information: the Agent case is \[mt~d\[ested as ~l~e .';yntactic subject; MsgObj as the direct object; MsgRecipientObj as either the indirect object or as the object (PrepO) of a prepositional phrase, whose preposition (CaseMarker) is &quot;to&quot;; CCRecipientObj as a prepositional phrase with &quot;prepositions&quot; either ccing or copying.</Paragraph> <Paragraph position="6"> Figu re 2: Caseirame for message In addition t,_) actions, we also use caseframes to describe objects Figure 2 shows a nominal caseframe for the message object of our electronic mall system. This has the same fern\] as the verb caseframe, except that its HeadForms correspond to the head nouns of a noun phrase describing an electronic mail message. In addition, the Descriptors case has a new SyntaxCase, prenominal, which implies that the elements of Pattern (new, re;-ent, etc.) may appear in the adjective position in this caseframe.</Paragraph> <Paragraph position="7"> With a suitable caseframe for MailAdrPesc and knowledge of what things like clause, noun phrase, direct object, adjective position, etc. mean, the above caseframes clearly contain enough information to produce analyses of sentences like: Forward to Jones at CMUA the messages from Smith.</Paragraph> <Paragraph position="8"> Did Brown resend any new me.ssage,.~ to Oreen at BI3N? What mail did Jones forward to ,PSmith? Brown is forwarding the re, cent mes@oges to Green, l~he central question is how to combine the information in the caseframe definitions with syntactic knowledge and thus analyze the sentences into a set of caseframe instances.</Paragraph> <Paragraph position="9"> The approach taken in earlier caseframe work at CMU has been to embed the syntactic knowledge in the parser code and let the parser interpret the caseframes using that knowledge. E.g. the algorithms in \[3\] use semantic caseframes and focus on prepositions as casemarkers as well as the order of subject, verb, indirect object and direct object for parsing. Unfortunately, prepositions tend to be small function words that are often poorly enunciated and recognized.</Paragraph> <Paragraph position="10"> Therefore we have adopted the same general approach for our speech parsing work, but modified the parsing algorithms. The same caseframes are used, but with a somewhat different interpretation process.</Paragraph> <Paragraph position="11"> The ability to apply multiple recognition methods is a central advantage of caseframe parsing. Since the restricted-domain language description embodied in the caseframes is at such a high level of abstraction, we are free to interpret it in a way appropriate to the particular situation. The caseframes tell us whet componen.ts to look for and constrains where we can look for them. But exactly how we look for them is adaptable so that it can be driven by the most reliable information we have.</Paragraph> <Paragraph position="12"> 4. Applying casel'rames to speech iqput We can summarize the previous two sections as follows: ~, caseframes of the kind we have described contain the right amount of information at the right level of abslraction to parse restricL~d-domain spoken input; ,,the algorithms that have been developed for using such caseframes in parsing typed natural language input are unsuiiable for spoken input because the algorithms rely on the presence of small function words that are recognized at best unreliably by word hypothesizers.</Paragraph> <Paragraph position="13"> The trial implementation o\[ our approach applies caseframes to the input, but does it in a novel way by: 1. examin!ng the lattice of words !typothesized by the speech recognizer for these that correspond to caseframe headers&quot; 2. combining all the case.frames correzponding to the words found in all semantically and syntactically plausible ways 3. for each caseframe eornbinatien thus formed, attempting to account for the gaps between the cqseframe header words that were involved in its formation by parsing words from the gaps against empty semantic and syntac\[.ic roles in the caselrame cembinatiorl 4. selecting as the final parse those c~seframe instances that best account for the tel)at, based on hew much input they cover and the acoustic scores of the words in that parse.</Paragraph> <Paragraph position="14"> This multi-stage approach avoids the problems of the caseframe parsing algorithms for typed input by anchoring the parse on caseframe headers. Caseframe headers are verbs (for clausal caseframes} and nouns (for nominal caseframes). These are content bearing words that tend to be stressed in speech and are often multisyllabic. This improves their chances of recognition above that of short, unstressed function werds. The anchor points are thus correlated to the most acoustically certain words.</Paragraph> <Paragraph position="15"> The idea of forming one or more parses at a skeleton level and instantiating the one (or ones) that satisfy all constraints down to the lexical level is akin to the ABSTRIPS \[10\] and NOAH \[t 1\] planners that first established a general plan and later worked in all tile detail called for in the situation. That way, the parser does not waste time in. hypothesizing local details that cannot possibly fit into a global parse.</Paragraph> <Paragraph position="16"> An additional advantage associated with working from caseframe headers is that tile resulting caseframe combinations form a ready. made semantic interpretation of the input. The interpretation is typically incomplete until it is filled out in the subsequent gap-filling stage. However, if the recognition of some or all of the rernaining words is so poor that the semantic interpretation is never fully completed; then the parser still has something to report. Depending on the application domain, a skeleton interpretation could be sufficient for the application, or would at least form the basis of a Iocussed t~eguest for confirmation or clarification to the user \[5\]. In the remainder of this section, we examine irJ more detail our current implemental.ion of,.the approach outlined above, starting first with a description of the word lattice that drives our casefl'am(.'-based parser for spoken input. This parser operates in the context of&quot; a complete speech understanding system Hint handles sp~aker independent continuous speech with a 200 word vocabulary in an electronic mail domain.</Paragraph> <Paragraph position="17"> 4.1. The word lattice Tile input to our caseframe speech parser can be viewed as a two-dimensional lattice of words. Each word has a begin time, an end time, and a likelihood score. The begin/end times slate where the word was detected in the utterance. The score indicates hew certes.in we are that the word is correct, based on acoustic-phonetic information. In the sample lattice below, the I~erizontal din~eP.sion is time, and the vertical dimension corresponds to certainly of recognition of individual words by the speech recognizer (F;nerating the lattice. This word lattice was consbucted by har=d for demonstration purposes.</Paragraph> <Paragraph position="18"> To start its processing, the parser selects from the word lattice all header words above a recognition likelihood threshold. These headers correspond to caseframes, but only some combinations of the hypothesized caseframes are possible in the domain. To calculate the legal caseframe combinations, a set of phrase structure rules were derived that apply at the frame level (rather than at the more detailcd word level).</Paragraph> <Paragraph position="19"> To make matters more concrete, let us refer to the sample lattice above. In this lattice, the underlined header words would be combined to form the nuclei of sentences like: &quot;Forward message Smith CMUA&quot; and &quot;Print message lineprinter.&quot; Caseframes can combine in this way if one is of the right type (as defined by the InstanceOf attribute for the case) to fill a case of another. When combining caseframes associated with header words, the parser also uses knowledge about word order to limit the possible combinations. In our example, the forward caseframe (as defined in Figure 1) has a slot for a MsgObjDesc as a DirectObject. Tile order restrictions built into the parser only allow for the direct object after the verb. The message caseframe (Figure 2) fulfills these requirements. It is a MsgObjDesc, whose HeadForm &quot;message&quot; occurs after the forward caseframe HeadForm &quot;forward&quot; in the lattice. Thus the two can be combined, ~s long as the constraint of the required MsgRecipientObj can be satisfied (by &quot;Smith&quot;).</Paragraph> <Paragraph position="20"> Each time a valid sequence of headers is found, it is given an overall likelihood score and merged with the previous ones. At the end of the header combination phase, we have a list of ordered partial phrases, containing all the legal sequences of header words that can be found in the word lattice. Each partial phrase is represented as a set of nested caseframe instances. For instance, three combinations would be formed from the header words: Forward message Smith CMUA and these would have the nesting structure: where square brackets indicate c.~seframo instances and tile r~esting is convoyed by textual inclusion, A routine,t() check word iunctures is used (:luring the header combination phase. Whenever two header words are combined for a p~rtial phrase, tile juncture between these words is chocked to ascertain whether they overlap (indicating an illegal combination), abut, or have a gap between them (indicating .,significant intervening speech events). This check also enables the p~;r'ser to deal efficiently with co-articuI~ted phonemes as in &quot;some messages&quot;. These pho~lemes are merged in pronunciation, resulting in a pair of ow:rlapping but valid word candidates. These word juncture checks comprise a tap-down feedback mechanism to improve the speech recognition.</Paragraph> <Section position="1" start_page="589" end_page="590" type="sub_section"> <SectionTitle> 4.3. Casemarker connection </SectionTitle> <Paragraph position="0"> Once caseframe combinations have been formed, the next step is to fill in the gaps between the words of the corresponding partial phrase. We take each combination in turn, starting with the one with maximal-likelihood. The caseframe speech parser first tries to fill In casemarkers, which are usually prepositions.</Paragraph> <Paragraph position="1"> Let us continue our example with the first header combination formed from the phrase &quot;Forward message Smith CMUA&quot;. For this phrase, casemarkers may appear before the prepositionally marked cases &quot;Smith&quot; and &quot;CMUA'. The requirement that the casemarkers must appear between the header words of the containing and contained caseframes is a strong constraint on the possible locations of.the casemarkers. Thet:e are generally strong limitations on what words could possibly serve as markers for these cases. In our example, using the caseframe definitions of the previous section, the parser would thus try to verify one of the words &quot;to&quot;, &quot;from &quot;, &quot;ccing&quot; or &quot;copying&quot; between &quot;message&quot; and &quot;Smith&quot; and one of the words &quot;on&quot; or &quot;at&quot; between &quot;Smith&quot; and &quot;CMUA &quot;. Whenever a set of words are predicted by the parser in a given segment, a word verification module is called. This module has knowl#dge of the complete word lattice. A word that matches the prediction is sought from the lattice in the specified gap. In addition, the acoustic-phonetic data is consulted to give an indication whether ~e word is a perfect fit for the gap, a left or right anchored fit, or if there are intervening significant speech events on the left or right.</Paragraph> <Paragraph position="2"> This information allows the parser to determine how much input has been accounted for by a given partial phrase hypothesis.</Paragraph> <Paragraph position="3"> Every successfully verified casemarker causes the parser to spaCvn another partial phrase hypothesis. The word could be a spuriously hypothesized word, i.e. one that was &quot;recognized&quot; even though it was never spoken (also known as a false alarm). Therefore we leave the old partial phrase without the casemarker in the ordered list of partial phrases and merge a new partial phrase into the list. The new partial phrase is a copy of the old one, with the casemarker also filled in. A new likelihood score is computed for this phrase.</Paragraph> <Paragraph position="4"> The score for a partial phrase is currently computed as the sum of the time normalized probabilities of each word divided by the time of the total utterance. Thus the probability of each word is multiplied by the duration of the word, summed over all words and divided by the duration of the utterance. This favors longer partial phrases over shorter ones. However, even extremely low scoring long phrase candidates are favored over well scoring shorter phrases. We are currently al';o exploring other alternative scoring procedures for</Paragraph> </Section> <Section position="2" start_page="590" end_page="590" type="sub_section"> <SectionTitle> 4.3. Casemarker connection </SectionTitle> <Paragraph position="0"> Once caseframe combinations have been formed, the next step is to fill in the gaps between the words of the corresponding partial phrase. We take each combination in turn, starting with the one with maximal-likelihood. The caseframe speech parser first tries to fill in casemarker,% which are usually prepositions.</Paragraph> <Paragraph position="1"> Let LIS continue our example with the first header combination formed from the phrase &quot;Forward message Snffth CMUA'. For this phrase, cas~;markers may appear before the prepositionally marked cases &quot;Smith&quot; and &quot;CMUA'. The requirement that the casemarkers must appear between the header words of the containing and contained caseframes is a strong constraint on the possible locations of. the casemarkers. rhet:e are generally strong limitations on what words could possibly serve as markers for these cases. In our example, using the caseframe definitions of the previous section, the parser would thus try to verify one of the words &quot;to', &quot;from&quot;, &quot;ccing&quot; or &quot;copying&quot; between &quot;message&quot; and &quot;Smith&quot; and one of the words &quot;on&quot; or &quot;at&quot; between &quot;Smith&quot; and &quot;CMUA &quot;. Whenever a set of words are predicted by the parser in a given segment, a word verification module is called. This module has knowledgc-' of the complete word lattice. A word that matches the prediction is sought from the lattice in the specified gPp. In r-~dditlon, the acoustic-phoneiic data is consulted to give an indication whether the word is a perfect fit for the gap, a left or right anchored fit, or if there are intervening significant speech events on the left or right.</Paragraph> <Paragraph position="2"> This information allows the parser to determine how much input has been accounted for by a given patti,t1 phrase hypothesis.</Paragraph> <Paragraph position="3"> Every succ~:ssfully w~rified casemarker causes the parser to spawn another partial phrase Ilypothesis. 1he word could be a spuriously hypothesized word, i.o. one that was &quot;recognized&quot; even though it was never spoken (also known as a false alarrn). Therefore we leave the old partial phrase without the cus(.~marker in tile ordered list of parti&l phr~tses and merge a new p~,.t tial phr.t,..se into the list. The new partial i)hrase is a copy of the old one, will\] the cas,:.marker also filled in. A new likelihood score it computed for this phrase.</Paragraph> <Paragraph position="4"> The score for a partial phrase is cunently computed as the ann of the time normalized probabilities or each word d!vided by the time of the total utterance. Th~s the probability of each word is multiplied by the dHt alien of tt~o word, summed over all words and divided by the duration of the utter~.'.nce. This favors longer pmtial phrases ever si~oHt.,r enos. However, even exhenlely low scoring long phrase c~tHdi:!al,~?S are favored over w~ll :~ccring shelh'~r phrases. We are eurr,?ntty also ~:~xFl,),ing other aht:!lDativ9 SCOIJll(J ptocedLtres for partial phrases. These methods will recognize the tradeoff between long, low scoring utterances that seem to account for all the input and short phrase hypotheses with excellent scores that leave gaps in the utterance unaccounted for. An ideal scoring function would also use semantic and syntactic wellformedness as criteria.</Paragraph> <Paragraph position="5"> Sometimes, none of the case markers being verified are found.</Paragraph> <Paragraph position="6"> This may moan that: + the speech recognizer failed to detect the marker. Unvoiced co-articulated monosyllabic words (such as prepositions) often go undetected; + or, the most-likely parse at the case-header level was indeed incorrect, and a lower likelihood parse should be explored to see if it is more consistent with the acoustic data.</Paragraph> <Paragraph position="7"> At present only the second choice is considered, but we are.</Paragraph> <Paragraph position="8"> exploring the possibility of an enhanced verifier to re-invoke the lower level processes (acoustic analysis or word hypothesizer modules) with shong expectalions (one or two words) at a prespecified window in the input. We llope that such a process can detect words missed in a more cursory general scan --- and thus use sen\]antic and syntactic expectations to drive the recognition of the most difficult segraents of the input. If the verifier were \[o return with a recognized case rnarl-~er, but too low u hkelihood, the overall likelihood value o\[ the next parse couM rnal,~e it the preferr(~d one.</Paragraph> <Paragraph position="9"> 4.4, Pronominal filling The next phase fills it\] the prenornirml sections of the partial i;,hrasos, The parser looks for prcnominals in the following order:</Paragraph> </Section> <Section position="3" start_page="590" end_page="591" type="sub_section"> <SectionTitle> Predeterminer DeterminP.r Ordinal Cardinal Adjective * </SectionTitle> <Paragraph position="0"> A lexicon associates each potential prenominal word with the correct type. Thus we first look for all possible predeterminers to.g, 'all') within the available gal) before the corresponding header word, Again the succe.':',sful verification of s,lch ~ prediction spqwns a new partial phrase, just as described for casemarker& the old partial phrase relnains ill the list as a precaution against false al,'u'nls. It should be notcd that remaining old phrases ac~;ounting for les::~ input receive a lower global likelihood value because unaccounted for input is penalized, Then deiernliners are examined. In our exanlple, the determiner &quot;th,.:&quot; will succu:s~,\[ully be foHnd to modify the nI,P, SS~i,~O ca:~,~\[rLulle, The other prenumirml types are lilted it\] the same way. Post.nominal modifiers (i.e., pruposJtiolial phr4ses) ~re parsed by the ca.';eframe instantiation inclhod above, as nominal an,I sunteadal caselrames ~.,.re treated in mucit lhe same w!G.</Paragraph> <Paragraph position="1"> 4,5. Extending coverage to simple questions Although we have not made completeness of syntactic coverage a focus in this work (see next section), we made some simple extensions to gain some idea of the difficulty in syntactic extension. In particular, we extended the system to deal with simple interrogatives as welt as imperatives and declaratives. No changes to the casehames themselves were necessary, just to the parsing algorithm. We introduced a separate stage in processing to look exclusively for question words. These words may be,'the standard wh-words (who, what, when, ...) or sentence-initial auxiliary verbs to indicate a yes/no question (do, does, is, will, ...).</Paragraph> <Paragraph position="2"> Tile word order rules in the header combination phase also required extension. These rules now have to allow fronted cases What messages did Smith send and questions where the HeadForm of the case is collapsed into a question word Who sent this message * Finally, we added a new module to fill auxiliary verbs in the correct locations. It operates just like the casemarker connection module and will not be described further here. By providing the parser with constraints governing the agreement of subject/verb, of auxiliary verb/main verb, and of pronominal/noun, the number of plausible alternatives is kept low.</Paragraph> </Section> </Section> class="xml-element"></Paper>