Recognizing Topics through the Use of Interaction Structures 
TAKESHITA, Atsushi 
NTI" Human Interface Laboratories 
1-2356 Take Yokosuka-Shi Kanagawa 238-03 Japan 
E-mail: take@ntthli.ntt.jp 
Abstract 
A crucial problem in topic recognition is how 
to identify topic continuation. Domain knowl- 
edge is generally indispensable for this. How~ 
ever, knowledge-based approaches are imprac- 
tical because not all domain knowledge needed 
for the identification can be prepared in advance. 
This paper presents a topic recognition model 
using dialogue interaction structures. The model 
can deal with both task-oriented and non-task- 
oriented dialogues in any language. Topic con- 
tinuation is identified without domain knowl- 
edge because utterances of relevant topics are 
indicated by certain interaction structures. The 
model avoids the weak point of knowledge- 
based approaches. The model is validated by 
the result of a topic recognition experiment. 
1 Introduction 
An aggregation of sentences having local co- 
herence is called a "discourse segment". Such 
a structure must be recognized to understand 
discourse including dialogues. The structure 
constrains candidates, for example, referents for 
anaphora resolution and plans for plan recogni- 
tion. A topic is a kind of local coherence. Seg- 
ments can be recognized in both task-oriented 
and non-task-oriented dialogues because most 
dialogues have explicit topics. 
Recognized topics can also be used in a topic- 
oriented video retrieval snpport system. The 
system recognizes the topic structures of video 
sequences such as documentaries, and shows a 
topic list. Topic nests are expressed by indenta- 
tion. Users can survey the contents of a video 
library, and play back sequences connected to 
an interesting topic. 
This paper describes how to recognize top- 
ics of both task-oriented and non-task-oriented 
dialogues without domain knowledge. First, a 
basic topic recognition mechanism is discussed. 
Second, identifying topic continuation through 
the interaction structure is presented. Finally, 
coverage of the interaction structure approach is 
discussed. 
2 A Topic Recognition Mechanism 
2.1 What Isa Topic? 
Topics are discourse referents shared by dia- 
logue participants. They are things described 
by noun phrases and events described by verb 
phrases. However, these referents are topic can- 
didates not topics. Those referents recognized 
as topics by persons must be shared by partici- 
pants for a while; presented as topics again, or 
referred to by pronouns or zero pronouns. 
A set of utterances having topic coherence 
is called a "topic segment". Topic structures 
consist of topic segments, topics in the segment 
and relations between the segments: nests or 
conjunctions. Post Office Dialogue in Fig. 1 can 
be segmented into topic segments as follows: 
(A-1 B-1 (A-2 B-2 A-3 B-3 A-4 B-4)) (A-5). 
~gt~i~ (Mail delivery) is talked about from 
utterance A-1 to B-4, N~ (express) from A-2 to 
B-4 and so on. 
There are various types of relations between 
topics. In Post Office Dialogue in Fig.l, the 
topic "NL~(express)" in utterance A-2 is a 
subtopic of the topic "J~9~idi~ (mail deliv- 
ery)" in A-1 because N~ is a subcategory of J 
(mail). In another example where a certain 
person Taro had moved to Kyoto recently, Kyoto 
may be a subtopic of Taro. Non-task-oriented 
dialogues may include various topic relations. 
ACTE$ DE COLING-92, NAMES, 23-28 AOt3"r 1992 l 0 6 4 PROC. OF COLING-92, NANTES, AUO. 23-28, 1992 
A-1 .~9~d~_ov,~-~L?cwo'v~¢o 
I'd like to ask you a question about 
mail delivery. 
B-1 tic,, ~5~o 
Yes, six'. 
I want to send a letter by express. 
B-2 Ni${Od<.~-~ ? 
Special delivery? 
A-3 ~.Ntc_tJ)\]Ll~-c_.m~\]~ ~ 9-/~, ? 
Will the letter reach Kyoto by 
tomorrow? 
I think it will because the next letter 
collection is at noon. 
Can I drop the letter into that mailbox? 
B-4 t.t V,o 
Yes. 
The next question is about a postal 
deposit. 
Figure 1: Post Of lice Dialogue 
TOPIC > EMPATHY > subject 
> object > others 
TOPIC is a noun phrase marked by the postpo- 
sitional particle " ~ (wa)". In the sentence " 
~?. (Tokyo) t.t ~e) (Japanese) ~ (capital) -t:. 
-~- (is)", ~ (qbkyo) is TOPIC. EMPATHY in- 
cludes the subject of mental verbs such as tg 
.~: (yorokobu, be glad), the source of ~< (iku, 
go), etc.. These verbs indicate the speaker's 
perspective. The subject markers include "/,~ 
(gay', and the object markers" ~ (wo)". 
These candidates can be used for topic mark- 
ers. The candidate priority of topics is the same 
as that of focus; if TOPIC exists, it is a topic. 
If TOPIC does not exist but EMPATHY does, 
EMPATHY is a topic. 
Examples of Japanese clue words indicating a 
topic change am shown in Table 2. Correspond- 
ing English clue words are also shown. 
Japanese pronunciation \]English \] 
-fgg--ec (mazu daiichi ni) first 
?~¢- (tsugi ni) next 
-~ ~'~ ~ c~ (sorekara) then 
c 7~ -e (tokorode) now 
This variety of topic relations makes it difticult 
to identify topic relevance by domain knowl- 
edge prepared in advance. Thus, the weak point 
should be avoided by a new approach. 
2.2 Topic Markers and Clue Words 
There are many topic marker expressions in 
Japanese. For example, expressions in Table 1 
indicate topics explicitly. English expressions 
such as "concerning ..." and "as regards ..." 
are similar to these expressions. 
Japanese expression pronunciation 
TOPIC ~l g-c (ni kanshi te) 
TOPIC ~ow-c (ni tsuite) 
TOPIC ~ v, 5 ¢~t~t (to iu no wa) 
TOPIC ti (wa) 
"TOPIC" means an indicated topic. 
Table 1: topic marker expressions 
Focus candidate priority in Japanese has been 
proposed\]l\] \[41: 
Table 2: clue words 
2.3 A ~lbpicStack 
A stack is used to handle discourse segments in 
the discourse model by BJ.Grosz\[2\]. A stack 
element corresponds to a segment, and is called 
a "focus space". Discourse entities such as 
objects are maintained in focus spaces. The 
top stack element holds the most salient enti- 
ties. Discourse segment structures are related 
to the intentional structure. The "dominance" 
and "satisfaction precedence" relation between 
intentions decides pushing and popping of focus 
spaces. 
A "topic segment" is a discourse segment of 
large size, and "topic stack" is used to handle 
topics. However, pushing and popping of topics 
cannot be determined by the intentional structure 
in ore approach because both topic-oriented and 
non-topic-oriented dialogues are treated, and the 
intentional structure may be ill-formed. 
Instead of the intentional structure, only clue 
words are allowed to determine the pushing 
AcrEs DE COL1NG-92. NANTES, 23-28 Aotrr 1992 1 0 6 5 PROC. OF COLING-92, NAN'rE.s, AUG. 23-28, 1992 
and popping. For example, "~ ~g4--~: (first)" 
indicates pushing, and "0~: (next)" popping. 
To recognize local topic structures, a simple 
mechanism is used. Each element of a topic 
stack is treated as a stack called an "inner stack". 
Topics are pushed onto the inner stack. If an 
explicit topic indicated by makers in Table.1 is 
recognized, non-explicit topics are popped from 
the stack. 
3 Identifying Topic Continuation 
3.1 The Basic Idea 
In dialogues, topics can be changed naturally at 
some utterances, but not at others. For example, 
topics unfold naturally in the dialogue in Fig. 1. 
On the other hand, topic expansion is not natural 
in the dialogue in Fig. 2. 
P-1 Will the letter reach Kyoto by 
tomorrow? 
Q-1 The next question is about a postal 
deposit. 
Figure 2: unnatural topic expansion 
3.2 Topic Expansion and Speech Acts 
The unnatural topic expansion in Fig.2 is re- 
lated to speech act purposes called illocutionary 
points. Classification of the illocutionary points 
was proposed by J.R.Searle\[3\]: The "assertive 
point" tells how the world is, e.g. to state and 
to predict. The "commissive point" commits 
the speaker to doing something. A promise is 
an example. The "directive point" tries to have 
the hearer do things. Making a request is an 
example. The "declarative point" changes the 
world by saying so, e.g, to declare and to name. 
The "expressive point" expresses the speaker's 
feeling, e.g. to apologize. 
A hypothesis is built: ira current utterance 
follows a directive utterance, the current topic 
is relevant to the topic in the directive utterance. 
This is called "topic forwarding". The unnatural 
topic expansion in Fig.2 can be explained by this 
hypothesis. The topic of utterance Q-1 must be 
relevant to one topic of P-1 because the utter- 
ance P-1 is directive. However, "0,:~: (next)" 
Speech Acts \[ Japanese Expressions 
ask ~-t;.~;o, (desu ka), ~T~ 
(masu ka) 
request ~I~ J'~,, (itadake masu ka) 
confirm ~ (ne) 
Table 3: Examples of Topic-Binding Speech Acts 
I Speech Acts Japanese Expressions 
inform ~"-,)- ~ ~" (desu ked0), ~'-e 
¢~: (desu ga) 
acknowledge ~ w (hai), ~L L (ee) 
Table 4: Examples of Topic-Nonbinding Speech Acts 
indicates a topic change. This contradiction 
causes unnatural topic expansion. 
Utterance pairs such as "requesting - accept- 
ing" and "asking - informing" will retain a topic 
even if the pairs are nested. For example, in the 
following, R-1 - S-2 have the topic of "restau- 
rant" and S-I and R-2 have the topic of "money 
for restaurant". 
R- 1 Do you know a good restaurant? 
S-1 How much money do you have? 
R-2 My salary is low. 
S-2 That reshaurant is cheap and good. 
However, pairs are not always so formed. In 
Post Office Dialogue in Fig.l, utterance A-3 
performs two speech acts: informing-if and ask- 
ing. Deeper dialogue understanding is needed 
for correct pair identification. Therefore, in this 
work, the pairs are not identified and a directive 
utterance is regarded as forwarding a topic only 
to the next utterance. 
3.3 Utterance Types 
"Topic forwarding" classifies utterances into two 
types: topic-binding and topic-nonbinding ut- 
terances. Topic-binding utterances have the 
directive point but topic-nonbinding ones do 
not. Topic-binding utterance speech acts in- 
clude to ask, to request and to confirm. Topic- 
nonbinding utterance speech acts include to in- 
form and to acknowledge. 
In Japanese, the utterance type can be identi- 
fied by pattern matching with expressions such 
as those shown in Table 3 and 4. 
ACRES DE COLING-92, NANTES, 23-28 AOUT 1992 1 0 6 6 PROC. OF COLING-92. NANTES. AUO. 23-28. 1992 
3.4 ~lbpic Recognition 
A set of consecutive utterances in which the 
same topics continue is called a "topic unit". 
A topic unit can be identitied by using "topic 
forwarding" instead of domain knowledge: 
1. The current utterance belongs to the same 
topic unit as tire previous utterance if the 
previous utterance is topic-binding, or there 
is no topic candidate indicated by a topic 
marker ill tile current ntterance. 
2. Otherwise, a new topic unit is created. 
The unit is used to validate candidate of topics 
and topic changes, and has no clTect on tile topic 
and the inner stack. 
Noun phrases indicated by topic markers arc 
regarded as topic candidates, and utterances 
with clue words are detected as topic change 
candidates. Some of them arc recognized as 
topics. Topic cmldidates are preserved in a 
"candidate list". Recognized topics arc pushed 
onto the inner stack of the topic stack described 
in 2.3. Topics can be identitied by using tire 
topic unit: 
a) A topic candidate indicated by a topic 
marker such as those listed in Table 1 is ina- 
mediately recognized as a topic, and pushed 
onto the inner stack. This is because such 
markers indicate topics explicitly. These 
markers are called "explicit topic markers", 
and the topics "explicit topics". 
b) A topic candidate indicated by other inarkcrs 
such as"/A (ga)" and" ,? (we)" is l/reserved 
in tire candidate list. It is recognized as a 
topic only when tile candidate continues for 
n utterances. If recognized as a topic, it is 
removed from the candidate list, and pushed 
onto the inner stack. The optimum value of 
n is 4 according to the results of a manual 
topic recognition experiment. 
c) lfa new topic unit is generated, the candidate 
list is reset to an empty list. 
d) A topic change candidate is recognized as a 
topic change only when the candidate is in 
the lirst utterance in a topic unit. 
Ira topic change is recognized, the candidate 
list is reset to an empty list and the immr 
stack is pushed onto or popfmd flom the 
topic stack according to clue words. 
This topic recognition algorithm can be used 
for any language because "topic forwarding" is 
not language-specific. Only dictionaries for the 
topic markers, the clue words and ttle utterance 
type identification am unique for each language. 
3.5 All Example of rlbpic Recognition 
In utterance Aol in Post Office Dialogue in Fig. 1, 
"~{~¢.~t~ (yuubin no haitatsu, mail delivery)" 
is identified as a topic candidate by the topic 
marker " m~v,c (ni tsuite)". This candidate is 
immediately recognized as a topic because of the 
explicit marker. Utterance A-1 and B-1 belong 
to the same topic trait because B-I has no topic 
candidate. The system state after processing 
t1-1 is the following. Each element of the topic 
stack is a inner stack. The right-most element 
of thc topic and the inner stack is the top stack 
element. 
CandidaZ~l)isZ = {} 
From utterance A-2 to B-3, a topic marker 
" ~ (we)" is detected in A~2 and B-3, and " 
~c (ni)" in A-3. Therefore, "Ni~ (sokutatsu, 
express)" in A-2, "gg (Kyoto)" in A-3 and" 
{~tfs (yuubin butsu, mail)" in B-3 are identified as 
topic candidates. Ftuthermom, B-3 is detected 
as a topic change candidate because of the clue 
word "?km (tsugi ni, next)". A-2 generates a 
new topic unit because B-1 is topic-nonbinding 
and there is a topic candidate in A-2. As a 
rcsuh of the unit generation, the candidate list 
is reset. Utterances from 11-2 to B-3 belong to 
tire second topic unit. This is because there is 
no topic candidate in B-2, and B-2 and A-3 are 
topic-binding. Therefore, the candidate "Ni~" 
continues for 4 utterances in tire second topic 
unit and is recognized as a topic. The topic 
change candidate in B-3 is dismissed correctly 
because it is not in the first utterance in the topic 
unit. The system state after processing B-3 is: 
(;andidat.cList = {~,N,~}. 
Utterance A~4 generates a new topic unit and 
the candidate list is reset to an empty set. In 
A-4, " ,~ x ~ (posuto, a mailbox)" is detected as 
a topic candidate. B-4 belongs to tire unit. The 
state of the inner stack does not change. 
ACTES DE COI.JNG-92, NANfES, 23-28 AOtn" 1992 l 06 7 PRec. OF COLlNG-92, NAm'~S, AUG. 23-28, 1992 
In utterance A-5, a topic candidate "~1~. 
(yuubin chokin, a postal deposit)" is identified. 
A-5 is detected as a topic change candidate be- 
cause of the clue word "0~tc (tsugini, next)". 
The change candidate is recognized as a topic 
change correctly because A-5 is the first utter- 
ance of a new topic unit. As a result, the inner 
stack is popped from the topic stack . The 
system state after processing A-5 is: 
TopicSt,,.~, : \[\[ \]J 
4 Discussion 
The results from a topic recognition experiment 
using 207 utterances taken from dialogue tran- 
scripts is shown in Table 5. Topics recognized 
by our system are compared with the manually 
recognized topics. 
Recognition and dismissal of topic change 
candidates was peffomaed correctly. This cor- 
rectness has the beneficial effect that wrong 
popping of the topic stack and the reset of the 
candidate list can be avoided. 
2 noun phrases were wrongly recognized as 
topics by the system. These errors occurred 
when current topic T-1 returned to past topic 
T-2, and T-2 was not described explicitly at that 
time. Although a topic change has occurred, T-1 
is regarded as a current topic because no topic 
candidate was presented. 
3 topics were not recognized as topics but 
were wrongly dismissed. This error occurred 
when the current topic was rephrased; "topic 
forwarding" fails in this case. Synonyms such 
as a fridge and a refrigerator are often used. 
Topic recognition accuracy is sufficient for 
a topic-oriented video retrieval support system. 
The recognition method is effective especially 
in dialogues with interaction structures such as 
"asking- asking" and "requesting - asking". 
The experimental results show that such struc- 
tures are included in many dialogues. Mixed- 
initiative dialogues may lbrm the structures. 
To improve topic recognition accuracy, other 
approaches such as a knowledge-based approach 
can be added. For example, a synonym list and a 
thesaurus would contribute to topic continuation 
identification. 
Recognized Dismissed 
correctly I wrongly correctly I ~gly 
ct,~.g~s II 21 0 II 21 0 
Topics I \[ explicit 15 3 -- -- 
other 7 2 5 3 
Table 5: The Number of Recognized and Dismissed 
Candidates 
5 Conclusion 
A topic recognition model has been proposed. 
The model identifies topic continuation by using 
dialogue interaction structures instead of domain 
knowledge. This is based on the hypothesis that 
a directive utterance repeats the same topic until 
the next utterance. The model has been validated 
by the results of a topic recognition experiment. 
References 
\[1\] S. E.Brennan, M. W.Friedman, and 
C.J.Pollard. A centering approach to pro- 
nounss. In 25th Annual Meeting of the 
Association for Computational Linguistics, 
pages 155-162. ACL, 1987. 
\[2\] B.J.Grosz and C. L.Sidner. Attention, inten- 
tion and the structure of discourse. Compu- 
tational Linguistics, 12(3):175-204, 1986. 
\[3\] J. R.Searle. Expression and Meaning. Cam- 
bridge University Press, 1979. 
\[4\] M. W.Friedman, M. Iida, and S. Cote. Cen- 
tering in Japanese discourse. In Papers pre- 
sented to the 13th International Conference 
on Computational Linguistics. ACL, 1990. 
ACRES DE COLING-92, NANTES, 23-28 AOt'rr 1992 I 0 6 8 PROC. OF COLING-92, NANTES, AUO. 23-28, 1992 
