5"OLING 82, J. Horecky, (ed.) 
North-Holland Publishing Company 
© Academia, 1982 
FLEXIBLE PARSING OF DISCRETELY 
UTTERED SENTENCES 
Luigi Borghesi Chiara Favareto 
Elettronica San Giorgio - ELSAG S.p.A. 
Via Hermada, 6 
16154 Genova 
ITALY 
In this paper we describe a syntactic semantic parser of spo-. 
ken sentences pertaining to a subset of natural Italian lan- 
guage. Error-free and fast analysis, partial interpretation 
ability, man-machine dialogue trend, different semantic envi- 
ronment adaptability and natural language usage are its main 
characteristics. All of these features are supported by a 
technique of input reliability evaluation. Particular atten- 
tion is devoted to the description of the knowledge internal 
representation and of the mechanism that manages, at diffe- 
rent points of the analysis, the whole process. 
I. PARSER QUALIFICATIONS 
The parser to be described performs its analysis starting from an intrinsically un- 
reliable input that is the result of an isolated word speech recognizer. The lack 
of certainty on the single items of the input sentence is one of the main problems 
in such a vocal parser. The representation of each uttered word, following the re- 
cognition stage, is, in fact, an ordered list of possible interpretations with as- 
sociated dissimilarity measures. As a consequence, it is possible to have doubts 
not only about every single word of the sentence, but also on complete sentence 
parts. Moreover, irrecoverable recognition errors may require the capability of 
parsing incomplete sentences. 
Spoken Sentences: TOGLI TUTTO DALLA STANZA 
Parsing Input : 
word dis word dis ~ord dis word dis 
I ~IOgLI I32. 1' I UN I46.~I ; DALLA 116.7: I RAQOIO :42.41 
I OTTO 134.7\[ ,; GUANTO;4&.8\[III DAL II?.4I I NOVANTA 142.81 
I DAMMI 134. 81 lil UNA ;46, 9:II : AL I20. 9 " STANZA 142. 9~I 
I IO~LIERE 135. 11 VENTI 147.0III \[ TAVOLO'23. 71 I GUADRATOI44. l,I 
I ~COSTRUISC1135. 91 OTTO 148. 11 I \[ I I I l I 
I I I I UNO I48. 51 I I I I I I 
I I I I TUTTO 4.9. 5 I I I I I I 
I I I I I I I I I I I I 
I I .... I I --~" .... I .... I I ....... I .... I I ' I .... I 
Figure l 
EXAMPLE OF A TYPICAL INPUT OF THE PARSER 
Fig. 1 shows an example of a typical parsing input where each input word is repla- 
ced by a complete list of possible alternatives with associated distance score. It 
37 
38 L. BORGHESI and C. FAVARETO 
is interesting to notice that not only the sentence that was actually spoken "TO- 
GLI TUTTO DALLA STANZA" (= remove everything from the room), is reported but, in 
this case, also some other correct sentences can be found by the parser (for ex- 
ample "COSTRUISCI UN TAVOLO QUADRATO" (build a square table). 
An efficient parser must also be able to solve other problems not strictly connec- 
ted to a particular kind of input. In fact it should, of course, achieve fast ope- 
rations; that requires the ability to minimize the number of alternative parses\[l\]. 
Furthermore the parser should be designed in such a way as to satisfy the "genera- 
lity" expectations; that is it should be easily adaptable to any semantic domain at 
least in the limited semantic domain cases. Since the parser results should be fol 
lowed by the execution of some operation in any practical application it is requi T 
red that it produced trusty results and, in particular, that it always included the 
right sentence interpretation within all the output ones. 
Finally, to allow a graceful dialogue with its users the parser must be able toana 
lyse also partial sentences (for example elliptical or fragmentary ones), thus ma Z 
king it possible to use naturally expressed sentences C2\]. 
2. MAIN PARSER'S CHARACTERISTICS 
The main features of our parser, that permit to satisfy the above mentioned requi- 
rements, are the following: 
I) representation of the language in terms of a network whose elements are syn- 
tactic groups and syntactic features; 
2) definition of a confidence measure of the recognition results and its exten- 
sion also to groups of words (syntactic groups); 
3) adoption of a recursive working strategy which anchors the parsing on the 
most reliable words in a first step and on the most reliable groups in a se- 
cond one. 
We selected the furnishing of a living room as the discourse domain and we defined 
a vocabulary of a I16 words. 
This vocabulary, although limited, leads to a total number of over 10 5 possible 
sentences that include commands for constructing or moving pieces of furniture, as- 
signment of labels, def~of unit lengths, inquiries about mutual distances~-- 
e-TCT. 
2.1. Language representation 
We describe each sentence of the language by a sequence of syntactic groups. A sy~ 
tactic group is defined as a sentence part with a well precise semantic meaning. 
Often, but not necessary, a syntactic group corresponds to a classical grammatical 
object. For example, we defined the verb, the direct object, the location object, 
etc. 
In this way (see Fig. 2) each sentence of the langudge can be described by a se- 
quence of some of these groups. 
On the whole we introduced only 9 groups; in our opinion this set of syntactic 
groups is enough to describe, at a syntactic level, all the possible sentences pe~ 
taining to this semantic environment. 
FLEXIBLE PARSING OF DISCRETELY UTTERED SENTENCES 39 
o.< Verb > 
< Location Obj.> 
< Indirect Obj.> 
< Dimension Obj.> I~ 
/ <Ouant.> III 
II < s ec. Obj.> ~II 
<Direct Obj.> ~<L°c" Obj'> ~ 
/ <Quant. > <Direct Obj.> x~ <.oo  obj .> 
o <Mood Obj.> <Direct Obj.> /7 
~\ <Quant.> / 
I\\\ < Location Obj. > /// 
I\£<Loc. Ob.i.> o <Conj.> o <Loc. Obj.>// 
I~ <Indirect Obj.> / 
L. < Direct Obj. > 
Figure 2 
LANGUAGE REPRESENTATION 
Each syntactic group is, in turn, represented by a number of possible word sequen- 
ces, or, mere precisely, of sequences of associated syntactic features. Figure 3 
shows, for example, how the direct object is represented. 
<l - 20 > • < FurnitvFe> 
< Name > 
/ < Furniture> 
// < Furniture > 
<Furniture> -~ 
<Article> ~V<Furniture>_ <ShaDe>x~ <Label> \ <l - 20> 
<Shape> j 
<Label> <l - 20> j 
< Unit > / 
<Everything> ..... 
Figure 3 
DIRECT OBJECT REPRESENTATION 
It is important to notice ~n~L u,u Teature can represent mere than one word and 
every new word does not always need a new feature definition. So the present voca- 
bulary can be easily increased to a certain degree within the semantic domain, with out any change in the grammar. 
2.2. Reliability evaluation 
The doubts connected with the vocal input suggested the need for a tool that mea- 
sured the goodness of each word recognition. To this purpose a method to evaluate the reliability of recognition results was defined \[3\]. 
40 L BORGHESlandC. FAVARETO 
By this method every word, in the ordered list, is associated with a confidence 
score indicating its probability of being the correct one (see Fig. 4). 
~GUADRATO ~DIVANO 
output oF the ~NOVANTA isolated wo~d :GUATTRO 
recognition ILATO For the ~o~d lRAggIO 
GUADRATO LQRADI IDIAMETRO 
DISSIMILARITY MEASURE 
823.96; 826. 318 
;27.09; the same I28.88' a@te~ 
the I29. 321 ~eliability 
'29. 581 evaluation 129, 828 
129. 848 
1 CONFIDENCE SCORE 
Figure 4 
RELIABILITY EVALUATION 
I l ..... I |QUADRATO 1.326 
IDIVANO 1.326 INOVANTA 1.326 
I(I~JATTRO I.OO1 ILATO 1.001 
IRA~glO 1.001 :QRADI 1.001 
IDIAMETRO |.001 I 
8 I 
In this way, as described below, the most reliable words of the sentence can be se 
lected and the parser anchored to them. The same reliability score is also used to 
evaluate the syntactic groups found and to decide which, among alternative groups, 
is the most probable one. 
2.3 Island driven workin 9 strategy 
All the operations of the parser are centered around the concept of reliability 
score. In fact, in a first step, the parser anchors its analysis to the most re- 
liable word of the sentence (that we named "guide word") and searches, both to the 
right and to the left ofitforall the syntactic groups that include the features as- 
sociated to the guide word. Each of these syntactic groups is named "island". 
Not only the first word in the ordered list can be used for this aim, but sometimes 
also the second and the third ones are taken into consideration. For each island 
a cumulative reliability score, function of the single word scores, is computed. 
The same procedure is then applied to the remaining words untilthe whole sentence 
has been examined and there are no more guide words; at this point a lattice of 
island, possibly overlapping, is obtained. 
In a second step the parser, in an almost identical fashion as before, searches for 
the most reliable island (that we named "guide island"), anchoring to it the explo- 
ration of the language network to get a match with one of the possible sentences. 
When this is not possible, because of very unreliable recognition of a whole syn- 
tactic group, the partial sentence recovered is proposed in output together with a 
hypothesis about the missing constituent. 
At this point a module for graceful man-machine interaction could be activated, in 
order to obtain the needed information by meansofan appropriate dialogue. 
In addition there are some parameters, specifying the number of retained alternati- 
ves at various points of the parsing, that allow to control parser's performance 
both in terms of speed and confidence. These parameters allow the parser to work 
with different degrees of flexibility and so, they must be carefully selected, ac- 
cording to the application, i.e. according to the risk that can be tolerated when 
accepting an acoustically unclear sentence. 
FLEXIBL~ PARSING OF DISCRETELY UTTERED SENTENCES 4\] 
3. RUNNING EXAMPLES 
In Fig. 5 the main steps of the analysis of a particular sentence are summarized. 
To make the comprehension easier we report a simulated english example that corre- 
sponds to a real italian sentence processed by the parser. 
(lth ~uide word) (2th auide word) 
PUT .B5 TWO .32 ROUND .91 TABLE .41 BOX .03 THE .32 . TABLES .38 
BUILD .03 TABLE •15 , 
• CHAIR .15 . ° 
PUT .B5 THE ROUNO TABLE ,36 
TWO ROUNO TABLES ,34 THE ROUND TABLE NUHBER TWO .19 
~(~uide island) 
outeut---~ P U T THE ROUND TABLE 
Figure 5 
RUNNING SIMULATED EXAMPLE 
(3th ~uide word) 
NUNBER .25 THE .82 ROOM .42 INTO .23 TWO •04 CHAIR .40 
ON .23 . TABLE .04 
* • • 
THE CHAIR .40 |islands 
INTO THE ROOM ,31 J ON THE TABLE ,11 
INTO THE ROOM 
The input sentence is: PUT THE ROUND TABLE INTO THE ROOM. In the first step, star- 
ting respectively from the I st, 2 nd and 3rd guide word (PUT, ROUND, THE), the par- 
ser finds some possible islands with associated reliability score. In a second 
step, starting from the guide island the parser searches a match between a path in 
the language network and the islands. The final result is the correct interpreta- 
tion of the sentence even if there were three recognition errors. 
Sometimes the parser outputs are not univocal as in the previous example. In fact, 
if the reliability score of a whole island is too low, the parser provides an out- 
put in which, instead of a detailed word-by-word interpretation, an hypothesis 
about the type of the missing syntactic group appears as shown in the example be- 
low: 
PUT < direct object > IN THE MIDDLE OF THE ROOM 
If, on the contrary, tnere are two or more words with approximately the same relia 
bility score and the same syntactic role, then the parser supplies in the output - 
those alternatives with their associated reliability scores and the whole decision 
will be deferred to a following pragmatic module or dialogue component. For exam- 
ple we can have an output like this one: 
PUT THE I TABLE CHAIR .32.29 I NUMBER TWO IN FRONT OF THE DOOR 
4. RESULTS 
The present parser has been tested on a'set of 50 sentences spoken by three diffe- 
rent speakers (two males and one female). A poor word recognizer was adopted in or 
der to stress parser's capabilities. We compare in table l the parser performanceT 
and those of the recognizer a~one. For each speaker the first column reports the 
percent of success of the recognizer alone, i.e. how many times all the words of a 
sentence were in the first position. The second column reports the percent of suc- 
ces of the parser (i.e. how many times the parser was able to interpret the sen- 
tence). 
Each row corresponds to a case in which there are, respectively, none,l, 2, and 3 
lost islands whose reliability was not sufficient to take any decision. 
42 L. BORGHESI and C. FAVARETO 
We want to notice that for the I st speaker the parser locates the correct sentence 
in the 92% of the cases and achieves a 96% correct interpretation if it assumes 
that there is one lost island. For the 2 nd speaker these values increase more slow 
ly because of a very unreliable input (10% of success for the recognizer). 
However the main result is that the parser never took a decision that did not con- 
tain the correct interpretation. 
I ; I ................. ; ................. I 
I# oF I SPEAKER N. 1 I SPEAKER N. 2 I SPEAKER N. 3 I 
Ilost I I I I 
lislands Ire¢ogntzerlparslrlrecogntzerlpa~serlrecogntzerIparserl 
I ........ I .......... I ...... I I ...... I .......... I ....... I 
I 0 I 30 I 92 I 10 I 10 I 58 I 8~ I 
I I I I I I I I 
I I I I 96 I I 54 I I 98 I 
I I I I I I I I 
I ~ I I I I 88 I I I00 I 
I I I I I I I I 
I 3 I I I I 94 I I I 
I I I ...... I I ...... I ............ I ...... I 
Table 1 
PERCENT OF SENTENCE RECOGNITION (see text) 
5. CONCLUSIONS 
We have described a parser of spoken sentences in which the design decisions taken 
are a necessary step to satisfy the requirements of a voice interactive system. 
In fact the choice of using features instead of vocabulary entries and of descri- 
bing the sentences in terms of syntactic groups agrees with the needs of improving 
the analysis speed, of making the semantic environment representation oriented to 
a partial interpretation analysis and of allowing the adaptability to different se 
mantic environments. 
Finally we want to notice the characteristics that make this parser oriented to a 
graceful man-machine interaction. They are its ability to provide in the output 
more than one choice with its associated reliability score and its ability to in- 
terpret also partial input. 
The first characteristic, in fact, allows a pragmatic module to make a choice among 
various alternatives, looking at their reliability score or, if in doubt, to acti- 
vate a dialogue module which requests the needed information. The second charac- 
teristic permits the management of a dialogue in which also elliptical sentences 
are allowed or, anyhow, the maximum freedom of expression is permitted. 

REFERENCES

Woods W.A., Syntax, semantics and speech, in: Reddy D.R. (ed.), Speech Recog- 
nition (Academic Press, 1975). 

Hayes P., Reddy R., An anatomy of graceful interaction in spoken and written 
man-machine communication, Report CMU-CS-79-144, Carnege Mellon Univ. (sept. 
1979). 

t~sso G., Borghesi L., Favareto C., Confidence evaluation for an Isolated 
Word Recognizer, 4th F.A.S.E. Proceedings (april 1981) 293-296. 
