to the whole or - PDF document

1 to the whole or the Department set of t
to the whole or the Department set of this grammar. the grammar in which very powerful of movinga new AcknowledgmentsI am very grateful to the many people who have influenced this research and madethis thesis possible. First and foremost, I would like to thank my thesis supervisors,Dr. David Goddeau and Dr. James Glass. DG provided me with an immeasurableamount of help, ideas, and encouragement, and he patiently listen

2 ed to countless"quick questions" while I
ed to countless"quick questions" while I worked on this thesis. Jim provided invaluable guidanceand advice, making sure I concentrated on the important ideas and "big picture"overall, as well as making sure I had the necessary resources and data. I cannotthank either of them enough.I also want to thank DIGITAL's Cambridge Research Laboratory for granting thefunding and support for this thesis research, as well as providi

3 ng me with excitingresearch projects dur
ng me with excitingresearch projects during the past few summers. Dr. Robert lannucci, the directorof DEC CRL, always ensured that other interns and myself were always providedwith intellectually stimulating projects. Likewise, the administrative and systemstaff furnished me with the necessary resources during each summer. I have alsothoroughly enjoyed interacting and conversing with the research scientists in this lab.I

4 would also like to thank MIT's Spoken L
would also like to thank MIT's Spoken Language Systems group for providingme with the data I needed for my experiments. I enjoyed getting advice from andconversing with the staff and students in that group.During the past few years, my supervisors have helped me considerably. ChrisWeikart taught me many valuable skills regarding system design and developmentduring my first two summers. Dr. David Goddeau provided great s

5 upervision duringthis past summer and wh
upervision duringthis past summer and while I worked on my thesis, introducing me to natural lan-guage. Dr. William Goldenthal brought me into the VI-A program at DEC CRL andopened my eyes to the fascinating field of speech recognition, while inspiring me tochallenge myself continually.Finally, I want to thank my friends and family, in particular my parents, for all oftheir support during the past five years at MIT, espe

6 cially during this last year. Thecontinu
cially during this last year. Thecontinual support of my parents made everything I have done possible. ...... .. .................................... .. .. ...................................................S ................. ........I I I --I.. Major components ............... parse tree flight from ................... associated semantic .......... ................... .21during parsing. ................... entries. ..

7 ........................ ...............
........................ ................ ........... ................................................. . ................... ................... ................... of piece-count ............. .... a spoken if necessary);of words Major components a speech generate meaning provide models models and these systems, system mayand subsequently understanding the a sentence a spoken for new parsing with grammar induction

8 the resulting focus on as how parsing ex
the resulting focus on as how parsing experiments. one in are derived. automatic grammar this grammar related grammar Chapter 2BackgroundOur research investigates the development of a new grammar induction techniqueusing semantics and semantic parsing as its basis. This chapter provides an overviewof related research as well as other relevant background information. We first discussseveral grammar induction techniques. N

9 ext, we loosely define semantics and re-
ext, we loosely define semantics and re-view other works aimed at extracting semantics directly from an utterance. Finally,we conclude this chapter by describing how some existing systems parse sentencessyntactically and subsequently compute semantics from the corresponding parse tree.2.1 Grammar InductionGrammar induction is an extensively researched field that attempts to derive andgeneralize the rules necessary to acc

10 ept utterances from a specified language
ept utterances from a specified language. Nu-merous techniques have been used with varying degrees of success in this field, manyof which are examples of pattern recognition and learning. These include clustering-based techniques [4], domain-specific heuristic methods [14], artificial neural networks(ANNs) [5, 13], hidden Markov models (HMMs) [14], and Bayesian networks [11].Much development and effort revolves around cl

11 ustering-based approaches. Theseapproach
ustering-based approaches. Theseapproaches begin with a simple, overly specific grammar (to a training set), whichis iteratively improved through means of merging syntactic units based on a givendistance metric. Indeed, some inference techniques begin with a grammar containing the training units together together 12].One possible distance metric involves computing distributional similarity, whichmeasures the divergence o

12 f the probability probability One techni
f the probability probability One technique uses this metric to derive a relatively uncomplicated grammarand subsequently acquires phrase structure automatically by using entropy or othermetrics to guess which pairs of units can form a phrase [3]. Another metric uses errorstatistics (differences between learned models and training data) to achieve iterativeimprovement of the acquired grammar [14]. As one can imagine, man

13 y other met-rics and variants exist; fin
y other met-rics and variants exist; finding a good good 13]. While some of these depend on super-vision in order to train the weights in the network, others are mostly or completelyunsupervised. In one such such However, theresearchers behind that work admit that this grammar grammar 12]. Unfortunately, while aesthetic and desirable, the latter, more automatic approachdoes not always produce results these techniques ove

14 rcome some a sentence.sentence.a broad
rcome some a sentence.sentence.a broad abstraction for a devicewhich translates input from one language into output in another one (here, a semanticrepresentation language); HMMs or ANNs can be used in the implementation of thisabstract device. One HMM-based approach is trained using the utterances utterances An interesting ANN-based approach consists of a system which is simply suppliedwith the words in an input senten

15 ce and a semantic category into which th
ce and a semantic category into which the sentencefalls [6]. The system then automatically computes associations (the weights) betweenvarious words and categories. The simple, one-layer version of this network actuallyhas no dependencies on word order and only computes whether enough words whichare associated with a certain class are present in the input utterance. The more com-plicated two-layer network does contain som

16 e type of phrase structure [7]; however,
e type of phrase structure [7]; however,the meaning or utility of this type of semantic phrase is not entirely straightforwardor apparent other than for use in performing simple semantic classification.Therefore, neither of these approaches of some some It is evensuggested (though not tried or implemented) that this technique allows for learning ofsyntax from the semantics used to complete a parse. Accordingly, our thes

17 is researchattempts to develop a differe
is researchattempts to develop a different technique for semantic parsing -one which can provideuseful information for driving grammar induction systems.2.3 Basic NLU System2.3.1 Components and StagesTo extract the semantics of an utterance, many systems first perform a syntactic parseand generate a syntax tree for the utterance. Next, they compute the semantics ofthe different components of the syntax tree, working up t

18 o the root of the tree togenerate its ov
o the root of the tree togenerate its overall semantics. This section will describe each of the various aspectsof such a system in more detail.Syntactic Parsing A syntactic parsing system typically uses a set of context freegrammar (CFG) rules, as shown in Figure 2-1, and through a series of rewrites,transforms a sentence into its syntactic be used + NP a noun as a of key-value shown in such as semantics produced to the

19 with thoseadjective (index parsing. Th
with thoseadjective (index parsing. The a set its arguments can combine, semantic functions to compute embedded noun of words words to perform book-keepingof partial results and keep track of matches between the input words and phrases andthe defined syntactic rules. In addition, the parser also records its current positionin each rule (usually denoted by a "dot" in output; the dot and contain of speech to the } ) .No

20 f complete of speech, been considered th
f complete of speech, been considered the chart. items other than rules. Indeed, in our implementation, edges simply contain a state,which allows the edge to contain words (items defined in the lexicon), rules, or otherpieces of information, as described later in this thesis. This provides us with theflexibility we need to parse utterances without using syntax, while still utilizing thesimplicity and power of the chart

21 parsing mechanism. the grammar semantic
parsing mechanism. the grammar semantic functions whatever combinations semantic functions general mechanisms to the be ignored. be considered any valid This drastically ignores word semantic functions a good to the such as under the can be to other being appliedgood example serve as serve as semantically. The can form a number representing different forms, by checkingand months. Semantic Function Example Inputs Out

22 put Semanticsadd_topic {a},{b} {a :topic
put Semanticsadd_topic {a},{b} {a :topic {b}}addtopic2 {a},{b} {a :topic2 {b}}add_agent {a},{b} {a :agent {b}}add_marked {a},{b :topic {c}} {a :b {c}}add_pred {a},{b} {a :pred {b}}add_name {a},{b} {a :name {b}}add_nnrel {a},{b} {a :nnrel {b}}Table 3.1: "Phrasal" semantic functions.This table lists all of the "phrasal" semantic functions we used in our experiments. Thesefunctions handle concepts of adding topics, agents,

23 predicates, and so forth, to other words
predicates, and so forth, to other wordsand phrases.certain types of numbers, as well as a possible "a_m" or "p_m" modifier, and attemptsto create semantics for a time. Table 3.2 lists all of these functions and gives examplesof how they work.3.3 ConstraintsAs mentioned, the semantic constraints represent a critical component of semanticparsing. These constraints provide restrictions on the output of semantic parsing,dic

24 tating which combinations are meaningful
tating which combinations are meaningful and preventing illogical meaning rep-resentations from being considered or created. The structure of the constraints aretightly coupled with the form of the semantic functions themselves, denoting whetheror not a certain concept can serve as a topic or as an agent (for use by add_topicand add_agent, respectively) of another concept, and so forth. The constraints arespecified in a

25 file which lists each of the major conce
file which lists each of the major concepts in a given domain. Undereach of these concepts, the constraints list a set of valid arguments, specifying thetype of the argument, as well as the possible keys under which the argument may beattached.The constraints typically specify this through the type and role arguments. Thetype argument ensures that only arguments of the specified type (or children of the semantic function

26 s.dates, and specified type) can be bou
s.dates, and specified type) can be bound under the appropriate key. Similarly, the role argumentdetermines what key (and hence what semantic function) can be used to combine apair of semantics. For example, when add_topic tries to combine a pair of semantics,it checks the constraints of the first argument to see if it has an entry with a role of"topic" and a type of the same type as the second argument. Likewise, add_a

27 gentchecks its first argument's constrai
gentchecks its first argument's constraints for an entry with a role of "agent." Mostof the other types of functions and constraints follow a similar pattern, with a fewexceptions. In particular, add_marked checks the constraints of its first argumentfor an entry with a role of the head of its second argument and a type correspondingto the semantics bound under the key of "topic" in the second argument. add_predsearches

28 the constraints of the second argument,
the constraints of the second argument, checking if it has an entry with arole of "subj" and a type corresponding to the first argument (checking if the secondargument can serve as a predicate for a subject of the specific type).In addition to listing these argument entries, the constraints can also includelists of nn-modifiers. These nn-modifiers refer to words describing objects which canmodify a target object, forming

29 a new object or concept. The nn-modifie
a new object or concept. The nn-modifiers for aword declare which other semantic concepts can be combined to form an object-object pair, such as "butter" and "knife" combining into "butter knife." Accordingly,the function add_nnrel checks its first argument's constraints to see if the secondargument is listed as a possible nn-modifier. Finally, the constraints for a word canalso list a possible parent of the word to all

30 ow for hierarchical lookup in determinin
ow for hierarchical lookup in determiningthe validity of predicates and arguments. Examples of constraints for selected wordsare shown in Figure 3-2.Thus, the constraints for "flight" allow addmarked to place a location under thekey of "from" or "to" and allow add_nnrel to modify the flight with airline or classobjects (i.e., "american airlines flight," "first class flight"). Similarly, the constraintsfor "from" allow ad

31 d_topic to give it a topic of a location
d_topic to give it a topic of a location. Finally, the constraintsfor "cost" allow addpred to apply it to a flight, and the constraints for "desire"allow addagent to give it a speaker as its agent (the one who desires something, ascompared to the item being desired). } )} )( { ( { :role subj The arguments If these are satisfied, indicating the arguments to "action" with compatible argument topic and number and to deter

32 mine its inputs type "action" solely on
mine its inputs type "action" solely on in all statement; the of a effort of to capture a user the three show me flights from words which contribute nothing to the show meto the as "filler words." More they still have their own semantics but can also be considered filler words for otherpurposes.The parsing system can therefore attempt to compute sentence-level semantics byskipping over the fillers and passing the non-fil

33 ler semantics of an utterance to thesent
ler semantics of an utterance to thesentence-level functions, which then decide into which sentence category, if any, thatutterance falls.3.5 Log FileIn addition to recording the generated meaning representation as it parses utterances,this system also computes and logs other information that may be of use in laterprocessing. Specifically, it logs the parts of speech of the words that combine as wellas the semantic funct

34 ion that allows for their combination.Th
ion that allows for their combination.Thus, the semantic parse of "cheapest flight from boston to philadelphia" containsthe information displayed in Figure 3-3. One can see that "cheapest" (an adjective)and "flight" (a noun) combine together by the addpred semantic function, and"from" (a preposition) and "boston" (a name) combine together by the add_topicfunction. In turn, these two units get bracketed together by addmar

35 ked, and thatunit combines with "to phil
ked, and thatunit combines with "to philadelphia" by the same function.The information recorded in semantic parsing also illustrates how the systemignores filler words. For example, the parse of "please show me flights from bostonto philadelphia" computes its final semantics by skipping over "please" and using thecomputed semantics of "show me flights from boston to philadelphia" as the sentence'soverall semantics.3.6 Im

36 plementation DetailsWe implemented the p
plementation DetailsWe implemented the program used for semantic parsing (shown in Figure 3-4) in C++for the WIN32 platform. In fact, we used the same program for syntactic parsingexperiments; the parsing mode can be set by a menu option. We chose the WIN32 (0 1):[addmarked (0 1):[add_pred (1 0): Adj N][addtopic (0 1): P Name][addtopic (0 1): P Name]utterance: which are flights from from (0 2): WhObj Cop [...]]utterance:

37 please show me flights from the actual
please show me flights from the actual a user thesis made made domain, a set of utterances involving airlinetravel queries and statements. In addition, some subsequent portability experimentsutilized the Jupiter [15] domain, a set of utterances regarding information about theweather.We divided these a speakerexperiment. The tested the for each these domains 2ABC [[A B] C] 1Table 3.4: Illustration of piece-count measu

38 rement.When the parser processes an utte
rement.When the parser processes an utterance consisting of the words A, B, sets. Semantic of each the utterances a set in coverage a piece-count the spectrum, ATIS TEST Histogram (Semantic Parsing)Set- Piece-Count 2 4 8 10 16 18the majority Grammar Inductionand phrases simple techniques semantic parse and subsequently make more these techniques Extraction Phasean example. (0 1):[addmarked (0 1):[add_pred (1 0): Adj

39 N][addtopic (0 1): P Name](0 1): P Name]
N][addtopic (0 1): P Name](0 1): P Name][add_topicUnique BracketingsFigure 4-1: Rules learned from unique bracketings in "cheapest flight from boston tophiladelphia."This figure shows the semantic parse of "cheapest flight from boston to philadelphia" andthe resulting syntactic rules learned via the unique bracketings approach. From the Adj Ncombination, the system extracts the rule X0 -+ Adj N. Similarly, it extracts th

40 e X1 rulefor the P Name combination and
e X1 rulefor the P Name combination and the merged rules.no new a new to the more generalized, extracted syntactic distance metrics a co-occurrence left of the grammar check piece-countof finalOverview of The induction ments on a separate test set of utterances.4.2 Semantic-Head Driven InductionOne potential flaw in the previous approach is that it does not make much use of thesemantic-level information present in th

41 e semantic parse logs. The unique bracke
e semantic parse logs. The unique bracketingsapproach uses just the bracketings provided by semantic parsing, essentially ignoringthe semantic phrasal information also embedded in the logs. Semantic-head driveninduction (SHDI) represents another approach to the grammar induction processthat does make use of this information. Essentially, SHDI builds on top of the uniquebracketings approach and improves it by making more

42 use of the information available.Because
use of the information available.Because of the nature of language, and accordingly the way we define semanticfunctions, the first argument to a function constitutes the head, or major concept,of the resulting semantic frame. For example, when "cheapest" and "flight" combine,"flight" is the first argument to add_pred, and the resulting semantic frame retains'flight" as its head. This can provide very useful information f

43 or grammar induction.Essentially, this o
or grammar induction.Essentially, this observation recognizes how semantic-level phrases are constructed.One can therefore use these semantic-level phrases to influence the learning of syn-tactic structure. By using the part of speech of the first argument to a semanticfunction, the extraction mechanism can create syntactic phrases based on semanticsand generate clean, readable rules directly from the parse logs, rather

44 than assign-ing arbitrary non-terminal l
than assign-ing arbitrary non-terminal labels. Further, the extracted rules are often recursive innature, eliminating the need for a subsequent merging phase; essentially, the learnedgrammar is pre-merged.The example in Figure 4-4 illustrates this mechanism. Given the semantic parsefor "cheapest flight from boston to philadelphia," SHDI begins by recognizing the noun"flight" as the first argument to addpred and extracts

45 a syntactic rule N_0 for theAdj N combin
a syntactic rule N_0 for theAdj N combination. Similarly, it recognizes the preposition "from" as the head in thecombination of "from" with "boston" and creates the P_0 rule and phrase type for thiscombination. Finally, upon encountering the combination of those two phrases, the marked (0 1):[addmarked (0 1):[add_pred (1 0): a set induction scripts syntactic parser,parsing experiment, which are to run or choosein which

46 rules are induction scripts the utteran
rules are induction scripts the utterances (and number of rules). are slight; rules between bracketings, our system generates different rules 15, X89, X90, X99, X146357358 involve each choosing merges parsing experiments, 15 57 67 71 1028 1072 90 , X8999 1000 1500 rules are N N. and parsesthus the N N The next to the the chart and immediately were semantically valid (and were handled completely by the defined s

47 emantics andconstraints).Another enhance
emantics andconstraints).Another enhancement made to the system involved relaxing the definition of"equals" for two edges. Two syntactic edges often had slightly different structurebut contained the same computed semantics, as shown in the "cheapest flight fromboston" example in Figure 5-1. We therefore relaxed the notion of equals so that twoedges with the same semantics were considered equivalent, based on the justific

48 ationthat an interpreter would perceive
ationthat an interpreter would perceive no difference in this set of potentially ambiguousedges and parses (and therefore maintaining a difference was unnecessary and evenwasteful).In addition, sometimes two edges would contain similar semantics with only slightdifferences, such as the locations where embedded arguments might be bound, asshown in the "from boston fare to philadelphia" example in Figure 5-1. Therefore, we

49 also relaxed the notion of equals for se
also relaxed the notion of equals for semantics themselves, so the contents, not thelocations of where arguments were bound, were considered. Again, we felt this wasperfectly reasonable, as the two alternative meaning representations would evaluateto the same thing but slowed down the syntactic parsing, since the parser had tomaintain and consider essentially duplicate edges in the chart data structure.With these enhance

50 ments, the syntactic parser proved to be
ments, the syntactic parser proved to be fast enough to runall of the desired experiments in a reasonable amount of time. Of course, because ofsemantic filtering, the syntactic parser could do no better than the semantic parser.If a phrase did not have valid semantics, the syntactic parser filtered it out, removingit from consideration. Thus, the results of semantic parsing served as an upper boundagainst which we could

51 evaluate the results of our syntactic pa
evaluate the results of our syntactic parsing experiments.5.2 Parsing ExperimentsThe first set of parsing experiments involved assessing how well the learned grammarperformed in the original domain. After doing a semantic parse of the training setof utterances from the ATIS domain and inducing the corresponding grammar, we N] [P Name]] {flight:from { city :name "boston"}:pred { cheap :type "superlative"}}[Adj [N [P Name]

52 ]] {flight:from { city :name "boston"}:p
]] {flight:from { city :name "boston"}:pred { cheap :type "superlative"}}edge: from boston fare to philadelphia[[[P Name] N] [P Name]] {fare:from { city :name "boston"}:to { city :name "philadelphia"}}[[P Name] [N [P Name]]] {fare:to { city :name "philadelphia"}:from { city :name "boston"}}Figure 5-1: Examples of edges with identical noun phrase, these parses tried the grammar on two sets of utterances from ATIS. First,

53 as an initial check,we ran a syntactic
as an initial check,we ran a syntactic parsing experiment in the original set of utterances in which thegrammar was learned (ATIS TRAIN) to ensure we got the same results as before. Infact, we observed slightly improved results! The reason for this improvement relatedto sentence-level rules. Because sentence-level functions required the pre-definitionof all filler words which it should ignore, a certain instance where a

54 non-filler wordactually contributed not
non-filler wordactually contributed nothing to an utterance could not be handled semantically; thesystem could not combine the non-filler word with the rest of the utterance. However,a syntactic sentence-level rule existed where the non-filler slot was ignored (becausethat slot was typically occupied by a filler word), allowing the non-filler word to betreated as a filler in order to compute the overall semantics for th

55 e utterance. Thus,one extra utterance wa
e utterance. Thus,one extra utterance was completely covered that had not been covered in the semanticparsing stage.Also, because sentence-level semantic functions were only applied over an en-tire parse, sentence fragments were never produced in semantic parsing, although anutterance might contain several embedded fragments or sentences. Their correspond-ing sentence-level syntactic counterparts, however, could be appli

56 ed anywhere in theparse, since the parse
ed anywhere in theparse, since the parser treated them like any other grammar rule. Therefore, the syn-tactic parser handled sentence fragments and embedded sentences much more easilythan the semantic parser, resulting in a lower piece-count for the syntactic parser.Overall, in the ATIS TRAIN domain, the syntactic coverage and piece-count resultswere very similar to the semantic coverage and piece-count, as expected.Set

57 Number of Utterances Total Piece-count N
Number of Utterances Total Piece-count Number ofComplete parsesATIS TRAIN 3764 12342 1504ATIS TEST 1033 3428 420Table 5.1: ATIS syntactic parsing results -coverage of sets.The overall syntactic parsing results for the ATIS TRAIN and TEST sets. The experimentsmeasured the piece-counts and coverage (total number of complete parses) of the utterancesin these sets. 2 4 6 8 16 18 the grammar of a were close being completelya

58 different is the Set Number of Utteranc
different is the Set Number of Utterances Total piece-count Number ofComplete ParsesJupiter TEST 1000 2529 541Table 5.2: Jupiter syntactic parsing results -coverage of set.The overall syntactic parsing results for the TEST set of Jupiter.the piece-counts and coverage of this set.Jupiter TEST Set: Piece-Count Histogram500400a300-20200-2 4 6 8 10# PiecesThe experiments measured12 14 16 18Figure 5-3: Jupiter syntactic pars

59 ing results -histogram of piece-counts.T
ing results -histogram of piece-counts.This graph shows the histogram of piece-counts for the utterances in the Jupiter TEST set.Again, the majority of utterances have a low piece-count.experiment, we defined a lexicon and set of semantics for Jupiter and attempted toparse a set of utterances using the grammar learned in the ATIS TRAIN domain.As shown in Table 5.2 and Figure 5-3, the grammar performs reasonably well.It c

60 overs 85% of the semantically valid utte
overs 85% of the semantically valid utterances (54% overall), so performance de-grades somewhat compared to the in-domain parsing experiment. We have theories toexplain this degradation in performance, however, and will address this issue shortly.Again, the piece-count histogram illustrates that the majority of utterances have alow piece-count, as desired.cM| I-1 1n~F. I in useful The system for each the appropriate beca

61 use an Our experiments time and this cha
use an Our experiments time and this chapter.the separate Pro] [[[[N [P Name]] [P Name]] [P Name]][P [Number Clock]]]]meaning representation: {show :topic2 {speaker :who "me" }:topic {flight :from {city :name "philadelphia" }:to {city :name "boston" }:pred {on :topic{weekday :name "monday" }}:pred {after :topic{time :hour 7 :merid "am" }}}}utterance: could you please tell me the cheapest farefrom atlanta to bostonsyntac

62 tic parse: Aux Pro [Please [[V Pro] [Det
tic parse: Aux Pro [Please [[V Pro] [Det [Sup [[N [P Name]][P Name]]]]]]meaning representation: ( {can},{you},{show :topic2 {speaker :who "me" }:topic {fare :from {city :name "atlanta" }:to {city :name "boston" }})utterance: show me roundtrip fares between sanfrancisco andwashington_d_csyntactic parse: [[V Pro] [Adj N]] PComp Name Conj Namemeaning representation: ( {show :topic2 {speaker :who "me" }:topic {fare :pred { t

63 rip_type:name "roundtrip" }{between},{ci
rip_type:name "roundtrip" }{between},{city :name "san_francisco" }, {and}, Cop [Det [[N N] [P Name]]]]meaning representation: {wh_ques :topic {WhObj :pred {quant :wh "what"}}:comp {forecast :for {city :name "boston"}:nnrel {weather}:pred {quant :sp "the" }}}utterance: i would like a weather forecast pleasesyntactic parse: Pro Aux [[V [[Det N] N]] Please]meaning representation: ( {speaker :who "me" }, {will},{desire :topi

64 c {forecast {forecast [P Name]]meaning r
c {forecast {forecast [P Name]]meaning representation: ( {be},{it},{rain :in {state :name "alaska" }})utterance: can you give me the low and high temperaturesin antarcticasyntactic parse: [Aux Pro [V Pro]] Det Adj Conj [[Adj N] [P Name]]meaning representation: ( {show :topic2 {speaker :who "me"}},{quant :sp "the"}, {low},{and},{temperature :in {continent :name "antarctica"}:pred {high}} )utterance: i would like to know t

65 he weather danny getoff the phone in wyo
he weather danny getoff the phone in wyomingsyntactic parse: Pro Aux V INF [V [Det N]] Unk V Unk DetUnk [P Name]meaning representation: ( ... )Figure 5-5: Examples of parses from Jupiter.This displays examples of parses from Jupiter. Again, one parse is complete and containsa reasonable and useful meaning representation. Other parses are incomplete and containvarying amounts of useful original query of speech this unrela

66 ted ATIS TESTO' 0not syntactically ...
ted ATIS TESTO' 0not syntactically ..." to the of rules rules are the grammar as one a generalized the grammar in new this grammarthe amount One can imagine using an ASR can be the grammar to a recognizer the grammar but rather use a set of the grammar. guidance provided syntax alto-handling them integrates syntax on seeded and results.NLU as in covering would be learn these this grammar. to the to the investigat

67 ion into the other results and robust to
ion into the other results and robust to significant improve- Bibliography[1] R. Bobrow, R. Ingria, and D. Stallard. Syntactic/semantic coupling in the BBNDELPHI system. In Proceedings DARPA Speech and Natural Language Work-shop, pages 311-315, Harriman, NY, February 1992.[2] E. Brill, D. Magerman, M. Marcus, and B. Santorini. Deducing linguistic struc-ture from the statistics of large corpora. In Proceedings DARPA Speec

68 h andNatural Language Workshop, pages 27
h andNatural Language Workshop, pages 275-281, Hidden Valley, PA, June 1990.[3] E. Brill and M. Marcus. Automatically acquiring phrase structure using distribu-tional analysis. In Proceedings DARPA Speech and Natural Language Workshop,pages 155-159, Harriman, NY, February 1992.[4] K-S Fu and T.L. Booth. Grammatical inference: Introduction and survey -PartsI and II. IEEE Transactions on Systems, Man, and Cybernetics, SMC-

69 5 No. 1and No. 4, January and July 1975.
5 No. 1and No. 4, January and July 1975.[5] C.L. Giles, C.B. Miller, D. Chen, G.Z. Sun, H.H. Chen, and Y.C. Lee. Extractingand learning an unknown grammar with recurrent neural networks. Advances inNeural Information Processing Systems, 4:317-324, 1992.[6] A.L. Gorin, S.E. Levinson, and A.N. Gertner. Adaptive acquisition of spokenlanguage. In Proceedings ICASSP, pages 805-808, Toronto, Canada, April 1991.[7] A.L. Gorin,

70 S.E. Levinson, L.G. Miller, A.N. Gertner
S.E. Levinson, L.G. Miller, A.N. Gertner, A. Ljolje, and E.R. Gold-man. On adaptive acquisition of language. In Proceedings ICASSP, pages 601-604, Albuquerque, NM, April 1990. L. Hirschman et al. Multi-site data collection for a spoken language system.In Proceedings DARPA Speech and Natural Language Workshop, pages 7-14,Harriman, NY, February 1992.[9] M. Kay. Algorithm schemata and and E. Levin and R. Pieraccini. Concept

71 -based Concept-based H. Lucke. Inference
-based Concept-based H. Lucke. Inference of stochastic context-free grammar rules from example datausing the theory of Bayesian belief propagation. In Proceedings Eurospeech, pages1195-1198, Berlin, Germany, Germany, M. McCandless. Automatic acquisition of language models for speech recogni-tion. Master's thesis, Massachusetts Massachusetts M. Negishi. Grammar learning by a self-organizing network. Advances in NeuralInfo

72 rmation Processing Systems, 7:27-31, 199
rmation Processing Systems, 7:27-31, 1995.[14] E. Vidal, F. Casacuberta, and P. Garcia. Grammatical inference and automaticspeech recognition. NATO ASI series. Series F, Computer and system sciences,147:174-191, 1995.[15] V. Zue, S. Seneff, J. Glass, L. Hetherington, E. Hurley, H. Meng, C. Pao, J. Po-lifroni, R. Schloming, and P. Schmid. From interface to content: Translingualaccess and delivery of on-line information. I