ANALYZING THE STRUCTURE OF ARGUMENTATIVE DISCOURSE 
Robin Cohen 
Department of Computer Science 
University of Waterloo 
Waterloo, Ontario, Canada N2L 3G1 
Consider a discourse situation where the speaker tries to convince the hearer of a particular point of view. 
The first task for the hearer is to understand what it is the speaker wants him to believe - to analyze the 
structure of the argument being presented, before judging credibility and eventually responding. 
This paper describes a model for the analysis of arguments that includes: 
• a theory of expected coherent structure which is used to limit analysis to the reconstruction of particular 
transmission forms; 
• a theory of linguistic clues which assigns a functional interpretation to special words and phrases used by 
the speaker to indicate the structure of the argument; 
• a theory of evidence relationships which includes the demand for pragmatic analysis to accommodate 
beliefs not currently held. 
The implications of this particular design for dialogue analysis in general are thus: 
• structure is an important feature to extract in a representation to control the processing; 
• linguistic constructions can be assigned useful interpretations; 
• pragmatic analysis is crucial in cases where the participants differ in beliefs. 
1 THE PROBLEM AREA 
Consider the task of designing an "argument understand- 
ing system", a natural language understanding system 
(NLUS) where the input is restricted to arguments. 
Consider as well arguments constructed in a dialogue 
situation, where a speaker (S) tries to convince a hearer 
(H) of a particular point of view. The hearer patiently 
listens; hence, the input is "one-way communication". 
The argument understanding system therefore plays the 
role of the hearer, and tries to analyze the structure of 
the argument being presented. This task is isolated as a 
necessary first step for a hearer, in order to be a success- 
ful participant in a conversation. In other words, the 
hearer must have some representation of what it is the 
speaker wants him to believe, before judging credibility 
and eventually responding. 
Note that this language problem is relatively new and 
yet feasible. It is distinct from other NLU endeavors, 
such as story understanding, which appeal to a stereotype 
of content in order to reduce processing. In arguments, 
one is never sure what points the speaker will address; 
content can't be stereotyped. However, arguments have a 
defining characteristic - they are necessarily goal-orient- 
ed. The speaker wants to convince the hearer of some 
overall point. Thus, there is an overall logical structure to 
the input and this fact may be used by a hearer to control 
analysis. 
For our model, the representation for the structure of 
the argument is restricted to an indication of the claim 
and evidence relations between the propositions. The 
notion of evidence is discussed in more detail in section 
4. A useful starting definition is: "A proposition E is 
evidence for a proposition C if there is some rule of 
inference such that E is premise to C's conclusion - in 
other words, there is some logical connection between E 
and C". 
In order to design an argument understanding system, 
what is then required is a computational model for the 
analysis of arguments. This in turn necessitates a theory 
of argument understanding, as a basis for the model. We 
suggest the following three components for the model: 
• a theory of expected coherent structure, used to drive a 
restricted processing strategy. Analysis is kept to a 
computationally reasonable task, by limiting the input 
to be recognized to a characterization of expected 
coherent forms of transmission. 
Copyright1987 by the Association for Computational Linguistics. Permission to copy without fee all or part of this material is granted provided that 
the copies are not made for direct commercial advantage and the CL reference and this copyright notice are included on the first page. To copy 
otherwise, or to republish, requires a fee and/or specific permission. 
0362-613X/87/010011-24503.00 
Computational Linguistics, Volume 13, Numbers 1-2, January-June 1987 11 
Robin Cohen Analyzing the Structure of Argumentative Discourse 
• a theory of linguistic clue interpretation, including 
insight into both the occurrence of clue words and their 
possible function in overall discourse. Clue words are 
those words and phrases used by the speaker to direct- 
ly indicate the structure of the argument to the hearer 
(e.g. connectives). Detecting clues can thus also serve 
to constrain the processing for the hearer. Moreover, it 
is important to have some facility for recognizing and 
interpreting clue words, to build a model that is robust 
enough to process a wide variety of input. 
• a theory of evidence relationships. The most important 
observation is that pragmatic analysis is mandatory for 
an analysis model in order to recognize beliefs of the 
speaker, not currently held by the hearer. Evidence 
connections between propositions often appeal to 
unstated information not currently in the hearer's set of 
beliefs, but recognizable as an intended support 
relation on the part of the speaker. 
2 RESTRICTED PROCESSING STRATEGY 
Consider the following framework for the model. An 
argument is considered to be a set of propositions. The 
model is then designed to analyze the argument a propo- 
sition at a time, incrementally building a representation 
for the underlying structure. The representation devel- 
oped is a tree of claim and evidence relations comprising 
the argument, where a claim node is father to its evidence 
sons. In order to assign an interpretation for a given 
proposition, one must thus simply assign it a place in the 
tree. In this way, one can tell to which propositions it 
serves as evidence and from which other propositions it 
receives support. 
A key design decision is to separate the two main 
operations of determining for each proposition (i) where 
it fits with respect to the argument so far, and (ii) howit 
relates to some prior proposition. The question of how 
two propositions relate in this framework is one of veri- 
fying that an evidence relation holds between the prop- 
ositions. This task is extracted and relegated to an 
evidence oracle, which, passed two propositions A and B, 
will respond "yes" or "no", as to whether A is evidence 
for B. 
With the problem of evidence determination factored, 
the model must still cope with the question of where a 
proposition may fit. This is handled by characterizing 
possible coherent transmissions (ordering of prop- 
ositions) on the part of the speaker and then limiting 
analysis to reception algorithms designed to recognize 
these coherent transmissions. In the section below we 
illustrate possible coherent strategies from the speaker, 
and present the associated reception algorithm required 
to recognize the input. 
Note that the computational model for the analysis of 
arguments is designed with certain aims and limitations. 
In particular, the model is to provide for analysis of 
"spontaneous discourse", demanding construction of a 
representation for the argument as each new statement is 
processed. The following restriction to the processing 
model is thus applied: the processor does not weigh all 
possible interpretations for a proposition; if the oracle 
sends back a yes answer to the question Is P evidence for 
Q?, then P is attached as a son to Q in the tree. In other 
words, the model simulates a hearer who does not have 
the luxury of "looking back" and re-interpreting previous 
statements. Moreover, the model aims to provide an 
interpretation for the current proposition, so that once an 
interpretation has been found (e.g. that P supplies 
evidence for Q) that proposition has been processed. The 
evidence relation is also taken to be transitive - i.e., if P 
is evidence for Q and Q is evidence for R, then P is 
evidence for R. (See Section 4 for more detail on the 
evidence relation). 
2.1 PRE-ORDER 
Pre-order transmissions are those where the speaker 
consistently presents a claim and then states evidence. 
The example below illustrates this form: 
EXI: 
1) 
2) 
3) 
4) 
5) 
With the tree representation: 
Jones would make a good president 
He has lots of experience 
He's been on the city board 10 years 
And he's very honest 
He refused bribes while on the force 
4 
2.2 POST-ORDER 
Another coherent strategy is post-order, where the 
speaker consistently presents evidence and then states 
the claim. Consider the comparable example below: 
EX2: 
1) 
2) 
3) 
4) 
5) 
With the representation: 
Js 4 J 
1 
Jones has been on the board 10 years 
He has lots of experience 
And he's refused bribes 
So he's honest 
He would really make a good president 
3 
In this example, the claim is the same as in EX1, that 
Jones would make a good president, but the evidence 
precedes the claim in each case - i.e., 1 is evidence for 2; 
3 is evidence for 4; and 2 and 4 together are evidence for 
the final claim 5. 
12 Computational Linguistics, Volume 13, Numbers I-2, January-June 1987 
Robin Cohen Analyzing the Structure of Argumentative Discourse 
The associated reception algorithms for pre- and post- 
order are described in detail in Cohen (1981). Both can 
be shown to operate in linear time (linear in the number 
nodes of the tree) and so are quite efficient. The pre-ord- 
er reception essentially finds an interpretation for the 
current proposition of an argument by searching for a 
father, up the right border of the tree. The post-order 
algorithm employs a stack; the current proposition tests 
to be father to the top of the stack; then all sons are 
popped off the stack, and the resulting tree pushed on 
the stack. 
We trace the construction of the trees of EX1 and EX2 
below, to provide details of the pre-order and post-order 
reception algorithms. 
EXI: PRE-ORDER 
Algorithm: 
- The current proposition is NEW. 
- The proposition immediately prior is LAST. 
1) Try NEW as evidence for LAST. 
2) If that fails, try NEW as evidence for each of LAST's 
ancestors, in turn, up to the root of the tree. 
3) If the test in 1 succeeds, stop. 
Consider as well a dummy root (D) for the tree for which 
all nodes are evidence (to place the first proposition). 
evidence 
candidates oracle 
in order test for NEW response 
1,D 2 ev. for 1? yes 
2,1,D 3 ev. for 2? yes 
3,2,1,D 4 ev. for 3? no 
4 ev. for 2? no 
4 ev. for 1? yes 
4,1,D 5 ev. for 4? yes 
EX2: POST-ORDER 
Algorithm: 
- Keep a stack of elements eligible to be evidence for 
a current proposition, with the latest one as TOP. 
- To interpret the current proposition NEW: 
1) Test if TOP is evidence for NEW 
2) If yes, then pop TOP (TOP:= TOP - 1) and make it 
a son of NEW (build a tree under NEW); then repeat 
step 1 with NEW value of TOP. 
3) If no, make the tree with NEW as root the new TOP 
of stack (push NEW onto stack). In essence, all sons 
for a proposition are picked up at once. 
stack 
elements 
1 
tree:l under 2 
3, 1 under 2 
3 under 4, 1 under 2 
evidence 
test for oracle 
evidence response 
1 ev. for 2? yes 
2 ev. for 3? no 
3 ev. for 4? yes 
2 ev. for 4? no 
4 ev. for 5? yes 
2 ev. for 5? yes 
1) 
2) 
3) 
4) 
5) 
With the representation: 
2.3 HYBRID 
As a first approximation to a general processing strategy, 
consider designing a reception algorithm to accept hybrid 
pre- and post-order arguments (i.e., any given sub-argu- 
ment may be transmitted in pre- or post-order). An 
example of hybrid transmission is EX3 below. 
EX3: 
Jones would make a good president 
He has lots of experience 
He's been on the board 10 years 
And he's refused bribes 
So he's honest 
1 
2 j 3J 
Here, the first sub-argument is in pre-order, the second 
one in post-order, and the argument overall is still coher- 
ent. 
The reception algorithm for accepting hybrid trans- 
missions is basically a combination of techniques from 
the pre- and post-order receptions. Now, to process a 
current proposition both a father and possible sons must 
be searched for. But the search is still restricted - certain 
propositions get closed off as possible "relatives" to the 
current one (e.g., earlier brothers of an ancestor). Thus, 
the complexity of the algorithm is still reasonable; it can 
also be shown to be linear. Once more, see Cohen 
(1981) for further examples. 
A full description of the hybrid algorithm is included 
below. 
L is kept as a pointer to the lowest node of the tree 
still eligible to receive evidence. It is initially set to a 
dummy root node to which all other nodes succeed as 
evidence. Consider as well a labelling where the last 
proposition in the stream that succeeds as evidence for a 
node is stored as the rightmost son. 
For each node NEW in the input stream do: 
/* find father */ 
do while (NEW is not evidence for L 
and L is not dummy root) 
set L to father of L 
end; 
Computational Linguistics, Volume 13, Numbers 1-2, January-June 1987 13 
Robin Cohen Analyzing the Structure of Argumentative Discourse 
/* see if there are any sons to re-attach */ 
if (rightmost son of L is not evidence for NEW ) 
/* no sons to re-attach */ 
then do; 
attach NEW to L 
set L to NEW 
end; 
else 
do; 
/* some son will re-attach; attach all sons of L 
which are evidence for NEW below NEW */ 
do while (rightmost son of L evidence for NEW) 
attach rightmost son of L to NEW 
remove rightmost son of L from L 
end; 
attach NEW to L 
end; 
Note that the hybrid algorithm, in searching for both 
sons and father to the current node, must contend with 
cases where a proposition is attached to a higher ancestor 
and must be re-attached to its immediate father. This 
occurs in a framework where the evidence relation is 
considered transitive. The kinds of orderings that involve 
this "re-location" are of the form: A, C, B - where C is 
evidence for B, B evidence for A (hence C also succeeds 
as evidence for A when tested). 
For EX3: 
candidates 
D 
1,D 
2,1,D 
3,2,1,D 
4,1 ,D 
(3&2 
"closed off") 
test for evidence oracle 
evidence response 
1 ev. for D? yes 
2 ev. for 1? yes 
3 ev. for 2? yes 
4 ev. for 3? no 
4 ev. for 2? no 
4 ev. for 1? yes 
5 ev. for 4? no 
(4 is a case of 5, so can't 
assume 5 ev. for 4) 
5 ev. for 1? yes 
4 ev. for 5? yes 
(testing re- 
attach of 
sons) 
2.4 SUMMARY OF PROCESSING STRATEGY 
The processing strategy proposed for our model of argu- 
ment analysis is designed to produce a selective interpre- 
tation. The particular restrictions to processing chosen 
for the model are, moreover, both 
• useful for measuring the efficiency of the model, since 
expressed in a framework of an algorithm operating on 
a tree structure, where complexity of the algorithm can 
be studied and 
• well-motivated, since based on a characterization of 
coherent input. 
In fact, it is an expression of a theory of argument coher- 
ence that serves to produce a model that is both efficient 
and robust (can handle a wide variety of input). 
Note that the particular restrictions suggested are 
drawn from analysis of a body of examples from rhetoric 
texts, together with some "naturally occurring" argu- 
ments (letters to the editor in newspapers). The aim is to 
characterize coherent transmission forms (ordering of 
propositions) that can be understood, without additional 
"clue words" to the underlying structure. We feel that 
the hybrid model is a good first approximation because 
examples of various forms of hybrid were encountered, 
and exceptions to the form all involved additional clue 
information. This leaves the study of clues and possible 
extended transmission forms to the next section. (In 
Cohen (1983) a few longer examples are run through the 
model to illustrate the forms it can accept and to moti- 
vate provision for the recognition of a wide variety of 
argument forms. Note that an actual implementation was 
not produced. There is now a scaled-down first imple- 
mentation (Smedley 1986), which will be discussed 
further in section 6.) 
3 LINGUISTIC CLUES 
The second main component of the argument under- 
standing model is a theory of linguistic clues. These are 
the words and phrases often used by the speaker to 
directly indicate the structure of the argument to the 
hearer. It is important to specify 
• what kinds of clues exist 
• the function of these clues in analysis - i.e., what inter- 
pretation can be assigned to a proposition that contains 
a clue word, and 
• (a more difficult question) when clues are necessary to 
ensure comprehension of the argument structure by the 
hearer. 
This approach to the study of clue words is much more 
detailed than the initial suggestions of Hobbs (1976) on 
how to interpret a few connectives such as and in his 
framework. It is also distinct from the investigations of 
Reichman (1981) and (recently) Grosz and Sidner 
(1986). Grosz and Sidner acknowledge the existence of 
clues and discuss various discourse structures that can be 
formed in the presence of clues. Reichman also gives a 
longer list of clue words and the particular conversational 
moves they signal. But there is no systematic proposal for 
interpreting a clue word that may occur (classification), 
and there is little discussion of how to process some of 
these more complex discourse structures without clues 
(suggesting when clues are necessary). In this section we 
clarify more deeply some of the discussion of clues in 
Cohen (1983, 1984b). 
3.1 CLUES OF RE-DIRECTION 
One type. of clue is expressions that specifically re-direct 
the hearer to an earlier part of the argument. Consider: 
14 Computational Linguistics, Volume 13, Numbers 1-2, January-June 1987 
Robin Cohen Analyzing the Structure of Argumentative Discourse 
EX4: 
1) The city is a mess 
2) The parks are a mess 
3) The park benches are broken 
4) The grassy areas are parched 
5) Returning to city problems, the highways are bad too 
With the representation: 
3 / ~4 5 
Here, the clue in proposition 5, returning to city problems 
signals additional support for proposition 1. In the 
absence of a clue, the specifications for the hybrid algo- 
rithm would have 5 test to be a son for 4, 2, and 1 (up 
the right border of the tree). So, adding clues should 
reduce the processing effort of the hearer. (In fact, if 
clues are consistently used by the speaker after long 
chains of evidence, the processing complexity of the 
reception algorithm can reduce from linear to real-time). 
3.2 CONNECTIVES 
Another popular type of clue word is the connective. We 
present a classification or taxonomy of connectives, and 
then associate a common interpretation rule within the 
claim and evidence framework for each category of the 
taxonomy. In this way, the interpretation of any proposi- 
tion containing a connective can be determined on the 
basis of the clue word's classification. For example: 
EX5: 
1) This city is a disaster area 
2) Houses have been demolished 
3) Trees have been uprooted 
4) As a result, we need national aid 
With the representation: 
Here,the connective as a result in proposition 4 belongs 
to a category known as inference, indicating that there 
should be some prior proposition that connects to 4 and 
serves as son to 4 - i.e., supplies evidence for "the 
result". In this example, in fact, 1 acts as the son. The 
interpretation rules are necessarily default suggestions for 
translating the semantic relations between propositions 
into our claim and evidence classification. (See further 
discussion on the evidence relation (section 4) for possi- 
ble exceptions). 
The taxonomy and its associated interpretation rules 
are presented below. 
Each category is a classification of connectives, made 
on the basis of semantic meaning of the connective - 
e.g., the parallel category would include all words that 
extended a list, including next, then, first, secondly, third- 
ly, etc. The set of classes was produced by considering 
the categories proposed in Quirk et al. (1972), and merg- 
ing some classes that had similar semantics and suggested 
the same interpretation rule for claim and evidence. The 
inference category covers phrases that suggest one 
proposition can be inferred from another - e.g., as a 
result, because of this, etc. The detail category moves in 
the other direction, and includes connectives that specify 
further - e.g., in particular, specifically, etc. The summary 
category is used for phrases that conclude a list. Refor- 
mulation captures clue words that repeat an earlier idea - 
e.g., in other words, once more, etc. Finally, contrast 
covers the phrases that introduce comparisons, like on the 
other hand or but. 
In conjunction with the semantic classes, evidence 
interpretation rules were then assigned as default inter- 
pretations. (See Section 4.3 for pragmatic "overrides" to 
the defaults). Note that some words may fall into more 
than one category, based on the meaning used - e.g., 
then meaning 'next' (in a list of actions) compared with 
then meaning 'as a result'. Clue interpretation thus 
requires a classification process as well. 
P is prior proposition; S is the proposition with the 
clue 
CATEGORY RELATION: P to S EXAMPLE 
parallel brother First, Second 
inference son As a result 
detail father In particular 
summary multiple sons In sum 
reformulation son (& father) In other words 
contrast brother or father On the other hand 
The taxonomy is described in detail in Cohen 
(1984b). We include discussion of one category here, as 
an example. The detail class is one case where connec- 
tives with a range of meanings were merged into one 
category. The title "detail" suggests that the connective 
will further specify some prior proposition. Included 
cases are: for example, in particular, and as another 
instance. The interpretation rule assigned to this category 
is that the proposition with the connective provides 
evidence for the earlier connecting statement. The moti- 
vation is that an accumulation of specific cases leads to a 
conclusion of a general statement (a form of reasoning 
used very often in naturally occurring arguments). 
EX6: 
1) The people in this town deserve a city-wide holiday 
2) For example, Old Man Jones has worked non-stop 
since Christmas 
3) And Mayor Flood is still recovering from all his 
efforts for the tornado relief 
4) In short, all of us are tired 
Computational Linguistics, Volume 13, Numbers 1-2, January-June 1987 15 
Robin Cohen Analyzing the Structure of Argumentative Discourse 
1 / 
4 2J "--3 
2 is son to 1 (detail class); 3 is also son to 1 (brother to 
2) (parallel class); 4 is father to 2 and 3 (summary class). 
3.3 THE FUNCTION OF CLUE WORDS 
So far we have discussed two types of clue words that 
can occur in conjunction with arguments transmitted 
according to the specifications of the hybrid algorithm 
presented in section 2 (our characterization for coherent 
discourse). We point out that re-direction clues provide 
additional information concerning which of the eligible 
propositions is related to the current one, and that 
connective clues specify the kind of relation that must be 
found between the current proposition and one of the 
eligible priors. 
In certain cases these restrictions will prevent some of 
the tests for evidence that would Otherwise have been 
conducted, thus saving some processing effort. For exam- 
ple, consider the following: 
EX7: 
1) The city is in serious trouble 
2) There are some fires going 
3) Three separate blazes have broken out 
4) In addition, a tornado is passing through 
with the representation: 
1 
The clue in 4 prescribes an interpretation for 4 as brother 
to some prior proposition. This means that 4 must act as 
evidence for some different proposition. Thus, even 
though 3 is technically the first proposition to test out 
when interpreting a new proposition (NEW evidence for 
LAST is the test), in this particular case this option is not 
possible. Thus, one round of work for the evidence oracle 
is avoided. In fact, for this example, the test "4 evidence 
for 2" fails, and the final test of "4 evidence for 1" 
succeeds. (Note that the brother relation is tested by way 
of son relation to a (common) father. This is because the 
model only processes evidence relationships). 
3.3.1 THE NECESSITY FOR CLUES 
We now examine the use of clue words in conjunction 
with transmissions that violate the specifications of the 
hybrid base case. The hypothesis is that more complex 
transmissions can be accommodated by the argument 
analysis model provided there exist clues to assist the 
hearer in recognition. In these cases, the clues are there 
by necessity. Their function in the discourse is not to 
merely add detail on the interpretation of the contained 
proposition, but to allow that proposition an interpreta- 
tion that would otherwise be denied. 
There is an advantage to adhering to the base ease of 
acceptable argument structures and treating the use of 
clues with other transmission forms as exceptional. In the 
first place, this provides a framework for detecting one 
interpretation for an argument when another possible 
interpretation could be generated if further testing 
occurred. In other words, in this model a representation 
drawn using the rules of the hybrid reception will always 
be accepted as the intention of the speaker, unless clues 
specifically override possible tests. 
To explain, consider the following example: 
EX8: 
1) The park benches are rotting 
2) The parks are a mess 
3) The highways are run down 
4) (And another problem with the parks is that) the 
grass is dying 
5) This city is in sad shape 
Without the clue phrase in 4, re-directing to proposition 
2, to add more evidence out of turn, a coherent represen- 
tation could be built just the same, as below: 
,J 
If the speaker intends 4 to add detail to the parks prob- 
lem, he cannot expect the hearer to make this connection 
without more information, simply because a more effort- 
less interpretation of 4 is possible (and the speaker 
should realize that it is this representation that the hearer 
will draw). 
Another important reason for separating the base case 
of acceptable transmissions is to allow for input that can 
be characterized as somehow a coherently generated plan 
of the speaker. Recall that the proposition analyzer will 
continuously call on an evidence oracle to determine if 
some proposition A acts as evidence for some other 
proposition B. Suppose there were no restrictions in the 
testing of evidence relations. Then, tests for evidence 
that would be interpreted as positive would return this 
response, regardless of when asked. In other words, a 
totally random display of propositions would result in the 
same representation for the argument as the reception of 
a coherently ordered presentation. Consider the follow- 
ing example: 
EX9: 
1) Yogi's been a shrewd manager 
2) He hired a hitter who now bats .400 
3) He traded in a pitcher who is now 0 and 12 
4) But he got involved in drug deals to the players 
5) And that's inexcusable for a manager 
6) He really needs to be axed 
16 Computational Linguistics, Volume 13, Numbers 1-2, January-June 1987 
Robin Cohen Analyzing the Structure of Argumentative Discourse 
with the representation: 
4 j 2 j ~3 
(where 5 and 1 are contrasting evidence to 6). 
EX9B: The argument as above, presented in the order:: 
2,5,3,6,4,1 
1) Yogi hired a .400 hitter 
2) (And) that's inexcusable for a manager 
3) He traded in a pitcher who is now 0 and 12 
4) Yogi really needs to be axed 
5) He got involved in drug deals to the players 
6) He is a shrewd manager 
This argument now appears incoherent. One reason is 
that there is contrast overall that must be clearly sepa- 
rated. In any case, the ordering does not conform to the 
specifications of the hybrid and as such is a candidate for 
an unacceptable t/'ansmission. 
This example suggests that the use of clue words, 
though helpful to signal exceptional transmissions, should 
still be studied as a systematic process of the speaker to 
assist the hearer in comprehension. In EX9B, could any 
number of clue words be added to still make this recogni- 
zable? Consider the following attempted repair to EX9B: 
EX9C: 
1) Yogi hired a .400 hitter 
2) But he's done some things inexcusable for a manag- 
er 
3) Although he also did other smart things like trading 
a pitcher who is now 0 and 12 
4) No, Yogi really has to be axed 
5) He got involved in drug deals to the players 
6) Though he still is a shrewd manager 
The question is whether an argument of this form would 
still be judged coherent. Our preference will be to specify 
particular types of exceptional transmissions that may be 
judged coherent. To this end, we have tried to isolate a 
few specific cases where clues can be used in conjunction 
with coherent orderings beyond the specifications of the 
hybrid algorithm. These are highlighted in the next sub- 
section below. 
One more point about the last example above is that it 
emphasizes our concern that the analysis of arguments be 
done efficiently. In this case, we want to avoid making 
tests for evidence relations that could not exist between 
certain propositions as part of a coherent input. In other 
words, we don't want the model to simply test out all 
possible combinations of evidence relations (even though 
this would only be n*n vs. k*n number of operations), 
because an input that is not coherent would then be 
accepted. 
We would also not want to derive computationally a 
representation for an argument that involved more 
computational effort, if a simpler interpretation of the 
same argument could be derived. (Again, the speaker is 
to assume that the hearer will not be searching unneces- 
sarily). This is illustrated in EX8 above. 
3.3.2 CASES OF NECESSITY 
Three kinds of acceptable extended transmission strate- 
gies are studied in Cohen (1983): parallel evidence, multi- 
ple evidence (a proposition acting as evidence for several 
claims, in restricted conditions) and mixed-mode sub-ar- 
guments (with evidence both preceding and following a 
claim). We present an example of the parallel 
construction below. See Cohen (1984b, 1983) for further 
examples. 
EXI0: 
1) The city has problems 
2) The parks are a mess 
3) The highways are a mess 
4) The buildings are a mess 
5) Here's some evidence for the fact that the parks are 
a problem: the benches are broken 
6) As for the highways, they're full of potholes 
With the representation: 
1 
5/2 6~~ ~4 
Here, the argument breaks the rules of hybrid trans- 
mission by adding evidence for an earlier brother (propo- 
sition 2); however, this shift is signalled with a phrase of 
intention in proposition 5, and the hearer may then 
expect a parallel expansion of additional support for each 
of the earlier brothers, in turn. (Note that without the 
clue, the argument structure derived would simply record 
all of 2 to 6 as evidence for 1, in the same vein as EX8). 
This example illustrates another interesting feature of 
clues - the variety of possible surface forms that can 
signal the same evidence relation between propositions. 
In EX10, the clue in 5 could also have been The problem 
with the parks is... or I will now explain why the parks are 
such a problem - .... A range of explicitness is thus possi- 
ble. In cases other than this parallel construction, in fact, 
a signal to a relation between propositions may be advo- 
cated by the simple use of an anaphor. For example: 
EXI 1: 
1) The mayor hasn't'done much for this city 
2) He doesn't seem to want to do much 
3) That man is a complete loser 
Here, the phrase that man signals a link to the mayor. It is 
difficult to decide whether the phrase qualifies as a clue 
word. The problem is determining a "bottom line" - i.e., 
"can't every sentence be seen to have some clue, from 
semantics, to its interpretation within the argument?". 
For now, we do not consider the cases of anaphora as 
above. 
Computational Linguistics, Volume 13, Numbers 1-2, January-June 1987 17 
Robin Cohen Analyzing the Structure of Argumentative Discourse 
3.4 SUMMARY AND FUTURE DIRECTIONS OF CLUE THEORY 
Our theory of clue interpretation so far has outlined the 
following principles: 
• Clues may occur with expected coherent transmissions 
or to signal exceptional cases. 
• Connective clues can be assigned common interpreta- 
tion rules according to the semantics of the clue. 
• To distinguish helpful versus necessary clues, the pref- 
erence will always be to recognize the hybrid trans- 
mission; if clue rules or semantics force an exceptional 
reading, only certain exceptional structures should be 
accepted. 
• In all cases, a reading that can be derived with less 
computational effort will always be taken as the 
intended reading. 
• The cases where clues are necessary to force a certain 
interpretation provide insight into the function of clues; 
their use in conjunction with acceptable transmissions 
suggests a function of additional processing reductions. 
In order to recognize clues and incorporate their inter- 
pretation into a larger model that derives argument struc- 
ture, we propose a separate clue interpretation module, 
interacting with the basic reception algorithm and the 
calls to the evidence oracle. Exactly how this module 
would function is left as future work. We do have a few 
initial insights, to suggest that clue interpretation is not 
only quite useful (argued previously in this section) but 
feasible. 
For connectives, the clue can be recognized from a 
classification. Then, determining whether a related 
proposition available from the list of eligibles is in fact 
related can be tested, according to the restrictions of the 
interpretation rule (as suggested in EX7). Further, the 
required semantic relation to the prior proposition can be 
passed as additional information to the oracle. The prob- 
lem is that this oracle must do some kind of search for 
connections between facts and axioms of a knowledge 
base. How this semantic analysis is achieved depends on 
the underlying representation, but additional semantic 
constraints should help to restrict operations. 
For re-direction clues, a processor would first have to 
recognize the appropriate phrase used. Some standard list 
(e.g., returning to, on the topic of) may be specified as a 
start. Then, the phrase should suggest some particular 
semantic content to the prior proposition (e.g., returning 
to the parks mentions parks as central). Now it is the 
form of representation of the propositions which may 
influence what is acceptable on a list of eligibles. If this 
semantic processing can be done very broadly, some calls 
to the oracle may be avoided, and this would be an 
improvement, assuming the oracle's operations encom- 
pass a more extensive search. 
4 EVIDENCE DETERMINATION 
The third main component of the argument analysis 
model is a theory of evidence, to govern the verification 
of evidence relations between propositions. 
An initial definition for evidence offered in Section 1 
is: "A proposition P is evidence for a proposition Q if 
there is some logical connection from P to Q - i.e., some 
rule of inference such that P is premise to Q's 
conclusion". The main problem in establishing evidence 
relations is that not all the premises are stated. For exam- 
ple, one common rule of inference used in arguments is 
Modus Ponens, of the form: P ~ Q, P therefore Q. The 
way this rule is most often used, the speaker will simply 
state P and Q and leave out the major premise "P ~ Q", 
expecting the hearer to be able to fill in the unstated 
connection to recognize the evidence relation from P to 
Q. Omitting certain premises is referred to as Modus 
Brevis and studied in Sadock (1977). 
We list below the rule of inference frames included for 
our model. Each rule has a slot for major premise, minor 
premise, and conclusion, to be filled by stated or unstated 
propositions, in recognizing an evidence relationship. 
RULE OF INFERENCE FRAMES: 
Major Minor Conclusion 
Modus Ponens P ~ Q P Q 
Modus Tollens P ~ Q ~Q ~P 
Modus Tollendo Ponens P v ~Q Q P 
Modus Ponendo Tollens P v Q Q ~P 
The form most often used is Modus Ponens. When the 
major premise is missing, this is the rule of inference to 
consider as the intended link from P to Q. 
EXI2: 
The Jays had a fantastic team this year 
All their players had averages over .250 
1) 
2) 
fill: 
3) If a team has all players with averages over .250, 
then that team is fantastic 
The common form for arguments, then, is one where the 
hearer must supply missing statements in order to estab- 
lish the connections for the representation of the argu- 
ment. 
There are several possible Modus Brevis forms for 
each frame above. The possible missing parts are classi- 
fied below for the case of Modus Ponens. 
MODUS BREVIS FORMS (MODUS PONENS): 
Given 
Premises Conclusion 
Normal P ~ Q, P Q 
Missing Minor P--,- Q Q 
Missing Major P Q (most popular form) 
Only Major P ~ Q (assume rest) 
Only Minor P (assume rest) 
4.1. OVERVIEW OF ORACLE'S PROCESSING 
It is important to demonstrate that the part of processing 
relegated to the evidence oracle within the overall model 
is not insurmountable, to defend the model as useful. In 
this section we examine more closely the operation of the 
18 Computational Linguistics, Volume 13, Numbers 1-2, January-June 1987 
Robin Cohen Analyzing the Structure of Argumentative Discourse 
oracle, opening up the "black box" just enough to 
suggest how it operates. 
In general, the oracle is asked to test two propositions 
to be in an evidence relationship, responding "yes" or 
"no" to a question of the form: "is P evidence for Q?". 
We sketch the operation of the oracle for the example 
below: 
EXI3: 
1) Joey is a dangerous 
2) He is a shark 
Assuming some resolution of anaphora, etc., a crude 
representation of the example in terms of predicates and 
arguments is: 
P: Shark(j); Q: Dangerous(j) (j: Joey). 
A Modus Ponens template would be of the form: S(j), 
for-all (x) (S(x) ~ D(x) ), therefore D(j) (S:is-shark, 
D:is-dangerous). (Note that we are not addressing 
certain questions here such as the significance that is a 
shark is definitional, while is dangerous is really assertion- 
al). 
Recall that the argument analysis model seeks to 
recognize intended argument structures. So, in this exam- 
ple the hearer can at least recognize that "P is evidence 
for Q" would follow through if All sharks are dangerous 
were believed (i.e., for-all(x) (S(x) ~ D(x) ). 
It is thus proposed that the oracle: 
1. identify missing premises (the Modus Brevis form of 
the argument being presented); 
2. verify plausibility of these missing premises (that 
they could be intended by the speaker to be believed 
by the hearer); and 
3. conclude that an evidence relation holds if the miss- 
ing premises are plausible. 
For step 2, we try to specify more precisely in the 
remainder of this section the kind of tests the hearer can 
apply. A summary is provided below: 
a) Identify the missing premise within a knowledge 
base of shared knowledge. 
b) Identify a "relaxed version" of the missing premise 
within own private knowledge. 
c) Identify the missing premise within a model of the 
speaker's beliefs. 
d) Judge the beliefs of a hypothetical third party 
(which could be simplified if the bottom line is 
"believe it, unless there's reason to strongly doubt, 
from within one's own beliefs"). 
We have found in simulations of the model on a varie- 
ty of examples that most of the tests for evidence can be 
answered through (a) and (b). We hypothesize that, 
given a specification of a knowledge base, the search for 
connections between propositions can be controlled. This 
hypothesis would be best verified with an implementation 
of the oracle and extensive analysis of examples, and 
could be the focus of the next stage of our implementa- 
tion (beyond Smedley (1986)). (The two large examples 
dissected in Cohen (1983) do have this property.) 
In addition to the problem that the major premise may 
be unstated is the problem that this premise should really 
be tempered by the beliefs of the speaker. In other 
words, the missing major premise that the hearer must fill 
in is really of the form: H believes that S wants H to 
believe (P ~ Q). In other words, this premise is not 
necessarily one of the hearer's beliefs. It is important to 
emphasize the importance of pragmatic processing in 
establishing evidence relations. The tree of claim and 
evidence relations built as a representation for the argu- 
ment is really an indication of the plan of the speaker, in 
the sense that each evidence relation recorded is the 
bearer's conception of a support connection intended by 
the speaker. 
Note that it is difficult to specify how a plan of a 
speaker is determined during analysis. What we are advo- 
cating is to interpret the intention of each proposition of 
the argument, the other propositions for which it 
provides evidence. The result of processing the entire 
discourse is not a complete plan of the speaker, in the 
sense that each of the "steps" could be executed and the 
top level goal (convince the hearer of some overall point) 
would then follow. It is more an indication of the moti- 
vation behind each utterance towards the ultimate goal of 
convincing the hearer. The difficulties in plan inference 
for discourse are discussed in more detail in Grosz and 
Sidner (1986), and are in fact a topic of our current 
concern (see discussion of future work in Section 6). 
There is in fact a whole spectrum of problems the 
hearer must face in recognizing evidence relationships 
between propositions. The four main tests for the hearer 
can be described as: 
• use logic, 
• relax the logic, 
• stereotype the speaker, 
• judge plausibility (reason as a "hypothetical person"). 
We illustrate these possible operations with examples 
below. 
4.2 LOGIC AND RELAXED LOGIC 
In example EX14, all the premises of the Modus Ponens 
argument are present. The hearer should realize that 1 
and 2 are evidence for the claim in 3. Then EXI5 illus- 
trates the more typical case of missing major premise. If 
the hearer fills in All machine candidates win, the 
connection from 1 to 2 can be seen. The problem is that 
the speaker probably believes something more along the 
lines of: Most machine candidates win. And yet, one 
couldn't record a Modus Ponens relation in the argument 
with a quantifier such as most instead of for all. Thus, the 
hearer must use some relaxation to the rules of logic to 
recognize the evidence relation. The detection of 
evidence through "relaxed logic" can be accomplished by 
having the hearer judge the unstated connection as a 
plausible generalization, based on a few known cases. 
For example, if the hearer tries to recognize an evidence 
relation from 1 to 2 in EX16 below, by filling in All 
Computational Linguistics, Volume 13, Numbersl-2, January-June 1987 19 
Robin Cohen Analyzing the Structure of Argumenlative Discourse 
sharks are dangerous, and the hearer doesn't believe this 
"axiom" but knows of a few sharks that are dangerous, 
he may reason that the missing major premise is reason- 
able. 
EX14: 
1) Aristotle is a man 
2) All men are mortal 
3) So Aristotle is mortal 
EXI5: 
1) Bilandic will win 
2) He's the machine candidate 
The point is that the hearer is still able to recognize 
connections he does not believe. In EXI7, the hearer 
should be able to understand an evidence relation from 1 
to 2, upon filling in a missing premise of the form "If a 
person stands for apple pie and Morn then he is great". If 
the hearer does not believe this statement himself, he 
may still consider it to be a reasonable belief of the 
speaker; having stereotyped knowledge of the speaker's 
views may thus be use'ful. 
EXI7: 
1) Reagan is great 
2) He stands for apple pie and Morn 
EXI6: 
l) Joey is a shark 
2) So, he is dangerous 
The idea of recognizing an intended connection from 
some other conversant based on one's own beliefs is not 
necessarily simple to implement. There has been some 
recent work by Pollack (1986) that suggests a more 
concrete foundation for this operation. Pollack discusses 
the problem of inferring a questioner Q's plan from his 
discourse. A first process has the responder R ascribing 
to Q a belief about some connection ("conditional gener- 
ation relation") that she herself believes true. Occa- 
sionally, R will need to recognize a connection that is not 
one of her beliefs. Then Pollack suggests there is a rule 
where "R ascribes to Q a belief about a relation between 
act-types that is a slight variation of one she herself has". 
In particular, one slight difference possible has Q believ- 
ing a stronger conditional generation relation, which is 
missing one of the required conditions. 
This related research is quoted here, not to argue that 
this problem is solved, but to acknowledge that it is 
important to specify this "relaxed" connection between 
one's own beliefs and those attributed to another. Some 
groundwork is being laid with more formal descriptions 
of plans such as Pollack's. 
4.3 DIFFERENCE IN BELIEFS 
The other type of problem faced by a hearer in recogniz- 
ing evidence relations arises because of difference in 
beliefs between speaker and hearer. As mentioned, the 
hearer is actually discerning intended relations on the 
part of the speaker, and must often reason outside his 
personal framework of knowledge. Again, an issue is 
raised of how to discern intentions from discourse. Some 
work has been done at the level of one utterance (e.g., 
Allen 1979). We are mostly concerned with advocating 
the inclusion of reasoning beyond one's own beliefs, 
without a full theory of how to infer another person's 
beliefs. Instead, we advocate a simplified framework, 
discerning evidence relations and allowing a connection 
to be drawn as intended if it is plausible to the hearer. 
For future work, we are studying how to specify this 
process more precisely. (See also Grosz and Sidner 
1986). 
Finally, if the hearer is testing a possible evidence 
relation between two propositions, does not believe the 
missing premise, and has no prior knowledge of the 
speaker, the best option available is to try to judge the 
plausibility of the unstated information. Essentially, the 
hearer must postulate new facts (which he may not be 
sure he wants to also believe) and consider relationships 
between these facts as plausible or not. It is in this sense 
that he adopts a "hypothetical person's" beliefs. Note 
that it is often the case that one will accept new facts 
unless something from one's own beliefs suggests a 
contradiction. In this sense, the judgement of plausibility 
does relate back to the hearer's own beliefs. 
An example with an implausible missing connection is 
EX18 below. If the hearer tests 2 as possible evidence for 
1, a major premise of the form "All sharks like tap 
dancing" would establish the relation. But the hearer 
would not regard this as a plausible belief of the speaker, 
and so would fail to recognize an evidence relation 
between the two propositions. 
EXId: 
1) Joey likes to tap dance 
2) He is a shark 
The problem of judging plausibility is difficult. To 
make this process more computationally tractable, one 
suggestion is to incorporate into the model some tracking 
of mutual belief between speaker and hearer. (See Cohen 
(1978) for further discussion on the use of mutual belief 
in natural language processing.) Then, certain tests for 
evidence relations can be blocked in the oracle, based on 
mutual belief. 
For instance, rules can be postulated regarding the use 
of claims and evidence that are mutually believed. Two 
sample rules are: 
(i) "If P is mutually believed, it can't be used as 
claim". 
(ii) "If ~P is mutually believed, P can't be used as 
evidence". 
In addition, some of the default interpretation rules 
associated with the taxonomy of linguistic clues can be 
overruled by pragmatic considerations. The idea is to 
possibly override the default semantic interpretations of 
evidence relations otherwise specified, by testing whether 
20 Computational Linguistics, Volume 13, Numbers 1-2, January-June ! 987 
Robin Cohcn Analyzing ihc Siructurc of Argumcntalivc Discourse 
propositions are already mutually believed. Note that the 
idea of a "pragmatic override" is also employed in 
Gazdar (1979) for the problem of determining presuppo- 
sitions. The importance of pragmatic processing for argu- 
ment analysis is once more emphasized, as it is a critical 
component to the difficult procedure of judging plausibil- 
ity. 
While considering mutual belief will help to eliminate 
some potentially difficult tests for the oracle, the model 
would require a more detailed specification of the main- 
tenance and use of mutual belief. This is left as a topic 
for future work. Some current ideas are explored in more 
detail in Cohen (1985). 
4.4 SUMMARY OF EVIDENCE THEORY 
The "theory" of evidence relationships, developed to 
specify the operation of the evidence oracle component 
of the argument analysis model, really presents insight 
into the problems relevant to evidence relations, rather 
than offering solutions. Still, the fundamental question of 
how connections between propositions can be verified 
has not been studied to any extent by other researchers. 
It is extremely worthwhile to acknowledge that it is not 
sufficient to indicate what relations do occur, without 
also suggesting how these relations could be established 
during analysis. 
In addition, we have provided some insight into the 
more general question of how to accommodate a possible 
difference in beliefs between conversants in a natural 
language dialogue processing system. We also suggest 
that a less sophisticated oracle can be constructed that 
merely searches known facts and axioms, possibly includ- 
ing relaxations, to handle a large amount of naturally 
occurring arguments. 
5 RELATED WORK 
5.1 ARGUMENT UNDERSTANDING 
Other researchers in natural language have studied argu- 
ments, in particular. However, the focus of the research 
in each case has been different. Birnbaum and the group 
at Yale (Birnbaum et al. 1980, McGuire et al. 1981) 
study two-way communication, developing an argument 
graph to display the points raised by both conversants. 
This graph is then used to determine the best moves on 
the part of an adversary, to challenge the position of the 
other conversant. Thus, the question of what responses 
to generate is investigated. On the other hand, there is 
little insight into how a hearer can detect the points being 
raised by the speaker, to construct this argument graph. 
Our focus has thus been on this preliminary problem to 
argument understanding. 
Archbold (Archbold 1976, Archbold and Hobbs 
1980) is most concerned with evaluative arguments, 
those with strong underlying ideologies. For example, 
Lenin's speeches are appropriate sample input. Thus, the 
difficult question of recognizing differing opinions is a 
focus to Archbold's investigations. In addition, he studies 
text rather than discourse, allowing for a deeper review 
(re-reading) of the input in order to derive an analysis. 
Weiner (1979) describes a representation for argu- 
ments that is also a tree structure, with a variety of links 
possible. His main concern is to characterize types of 
argument structures, for use in the generation of explana- 
tions. There is thus little attention on the problems 
encountered in deriving argument structures during anal- 
ysis. 
Weiner's (1979) model for the structure of explana- 
tion bears some resemblance to the representation 
described for arguments here. Weiner claims that natural 
explanation can be regarded as a series of transforma- 
tions of an underlying tree structure that represents the 
abstract form of the argument being developed. The 
ways in which support can be supplied are listed more 
extensively, including examples, alternatives, etc. How 
the tree can be built up relies on tracking a node that is 
"in focus". The fact that determining the relations 
between statements may make use of clue words is 
mentioned briefly as well. Basically, some of the features 
we advocate appear in this research. But we are trying to 
provide more insight into operational questions such as: 
• How do you determine the (best) relation between 
propositions? 
• How is the focus set? and 
• When are clue words likely to occur? 
By contrast, Weiner concentrates on how to generate 
explanations using his precisely specified characteriza- 
tion. 
Reichman (1981) is concerned with a larger problem 
of producing a model of discourse (not just arguments), 
but her approach should handle arguments as well. The 
core of the model is an ATN grammar for parsing and 
generation, coupled with a representation of "context 
spaces" containing conversational moves. The conversa- 
tional moves provide a classification of larger compo- 
nents of discourse (not just single propositions). For 
example, there is an extensive study of a "challenge" 
operation. Since Reichman's aims are broader than ours, 
the lower level issues we address of verifying evidence 
and studying the necessity for clue words are not consid- 
ered for the model. Moreover, there is an intentional 
lack of concern with pragmatic processing, another 
crucial feature our model. Instead, Reichman presents a 
model for the analysis of a variety of two-person inter- 
actions. 
In sum, our efforts in argument understanding are 
worthwhile because we focus on the necessary first step 
in argument analysis - determining the intended struc- 
ture, or "what the argument is about". We study the 
possible structural relations between propositions, and 
investigate the difficult issue of verifying evidence 
relationships. The importance of pragmatic analysis to 
recognize whole classes of arguments that involve differ- 
ing beliefs is stressed in our work. And finally, the use 
and interpretation of clue words is addressed. It is worth 
Computational Linguistics, Volume 13, Numbers 1-2, January-June 1987 21 
Robin Cohen Analyzing the Structure of Argumentative Discourse 
noting that the differences in our existing studies can 
possibly be exploited by pooling efforts and suggesting a 
powerful general model for argument analysis. 
5.2 REFERENCE RESOLUTION AND FOCUS 
Some of the work on using focus for reference resolution 
contains similarities to the model presented here for 
analyzing argument structure. Sidner (1979) maintains a 
focus stack of possible items in focus and an alternate 
focus list to support shifts of focus. The candidates for 
resolving references are thus restricted and ordered. The 
point is that these restrictions are drawn from a charac- 
terization of coherent discourse, the same approach 
taken for our control of processing. Grosz (1977) 
presents a model of focus spaces which may be used for 
several purposes, including the resolution of definite 
noun phrase resolution. The spaces are organized into a 
hierarchy, thus similar to our tree representation for 
argument structure. Both active and 9Pen spaces are 
tracked, similar to our tracking candidates eligible to be 
relatives to the current proposition. 
Because of similarities in the representations and tech- 
niques for controlling search for interpretations, it is 
worth investigating as future work the precise relation- 
ship among coherence, reference resolution and focus 
determination for dialogues (some of this is being done 
(Grosz and Sidner 1986)). 
5.3 PSYCHOLOGICAL RESEARCH 
Although our model is not designed according to psycho- 
logical studies of discourse comprehension, there are 
some interesting parallels with existing psychological 
research. Labov and Fanshel (1977) investigate thera- 
peutic discourse, dialogue between a psychologist and his 
patient. The research describes several properties of the 
arguments advanced by the patients including: the use of 
poor logic, the tendency to omit premises in arguments, a 
variety of transmission forms (claims before and after), 
and the existence of statements about the structure of the 
argument (clues). Since our characterization of input 
provides for all these forms, it strengthens our case for 
having a robust model. 
Geva (1981) investigates the usefulness of flowchart- 
ing text structure to assist students in comprehending the 
underlying structure. The top-down influence of building 
and using a representation is mentioned as important. 
The fact that many texts do not follow a strict linear 
ordering of connections between statements again 
confirms our concern with a variety of possible coherent 
transmissions. 
In brief, discovering psychological experiments that 
agree with the constraints of our model serve to defend 
its design. Further, some suggestions we make about 
computational measures of discourse processing may 
serve to inspire new experiments into the nature of 
human processing. So, the relationship with psychology 
should be a two-way exchange, and suggests future work. 
6 USEFULNESS OFTHE THEORY 
The computational model for the analysis of arguments, 
as described in the previous sections, is built on a theory 
of argument understanding; it, in turn, can be used as the 
basis for an implementation of an argument understand- 
ing system. One suggested real-life application area is a 
complaint bureau for department stores. Future work 
could include a full implementation of the mbdel, and 
fine-tuning the design by selecting a particular applica- 
tion area for arguments. 
Although there is no complete implementation of the 
model to date, an overview of a possible design is 
presented here, to indicate how the various components 
of the model could come together into one integrated 
"system". (Note that an initial implementation of the 
model does exist now, written in Prolog (Smedley 1986). 
But this program merely tests the various reception algo- 
rithms described in section 2. The evidence oracle is 
replaced by a "query the user" facility. Nonetheless, the 
groundwork is in place for a future implementation that 
tests the other components of the model). 
In Figure 1, there are three main modules: the Propo- 
sition Analyzer, Clue Interpreter, and Evidence Oracle. 
The Proposition Analyzer takes as input the argument 
itself and produces a representation of its underlying 
structure. For each proposition of the argument the 
Proposition Analyzer attempts to assign it a location in 
the representation tree, indicating to which other propo- 
sition it relates (provides evidence for or receives 
evidence from). The Proposition Analyzer may call on 
the Clue Interpreter in the presence of clues, to assist in 
the interpretation. In addition, once an eligible relative to 
the current proposition is selected, the decision of wheth- 
er an evidence relation exists is made by the Evidence 
Oracle, which is passed the two propositions and 
responds with a yes or no answer. The Evidence Oracle 
has available, a knowledge base of shared facts and, if 
,possible a model of the speaker. Moreover, if certain 
beliefs of the speaker can be extracted during the tests 
for unstated premises, the model of the speaker may even 
be updated by the Evidence Oracle, to aid in the process- 
ing of later propositions. 
In the absence of an implementation, the model can 
still be defended as a useful prescription of analysis of 
arguments. This is accomplished in Cohen (1983) by 
hand simulations of a variety of examples, to demon- 
strate robustness, and analysis of the complexity of the 
processing algorithms, to demonstrate efficiency. 
Another argument for the usefulness of the argument 
analysis model is that the theories developed for the 
model may be applied to the solution of other language 
understanding problems. As a result, the study of argu- 
ments may be viewed as a worthwhile exercise in the 
study of language. Some examples of the wider applica- 
bility of the model are: 
• It has been shown that extracting the underlying struc- 
ture of discourse is useful to study the complexity of 
22 Computational Linguistics, Volume 13, Numbers 1-2, January-June 1987 
Robin Cohen Analyzing the Structure of Argumentative Discourse 
PROPOSITION ANALYZER 0 
EVIDENCE ORACLE 
model of speaker 
\ 
knowledge base 
argument representation 
~ argument input 
CLUEINTERPRETER 
- --> data flow 
--> control flow 
Figure 1. System design. 
analysis. In the modal, the separation of where and 
how propositions relate has provided a means of moni- 
toring the number of calls to an inference engine, apart 
from the more difficult measurement of the process of 
actually filling in missing inferences. Hopefully, some 
characterization of structures for language problems 
other than arguments would be extremely beneficial. 
• It has been shown that certain linguistic constructions 
serve a function in facilitating the analysis process for 
the hearer. Developing common interpretation rules 
for various "linguistic clues" should continue for 
several language understanding tasks. 
• Some insight has been offered into how communication 
can proceed despite differing beliefs of speaker and 
hearer. The ideas outlined for recognizing beliefs simi= 
lar to one's own, for judging plausible generalizations, 
should extend to other problems where reasoning 
beyond one's current beliefs is required. 
In addition, very few researchers seem concerned with 
truly "low-level" operations, determining not just 
"what's a good representation for discourse" but also 
how this representation can be derived, the specification 
of some algorithm for processing. It is in this domain that 
we feel our research is making a contribution. 
For future work, we are currently developing a model 
for discourse analysis in general, based on the principles 
of this model's design. A hypothesis worth investigating 
from the existing model is that the resulting represen- 
tation serves to outline both the linguistic structure of the 
discourse and the intentional structure (the speaker's 
intentions behind utterances). The idea is that determin- 
ing evidence relations in an argumentative discourse may 
best be described as uncovering the intended uses of 
utterances (e.g., speaker utters P in order to get hearer to 
believe Q), hence reflecting the plan of the speaker. But 
this main function of deriving intentional structure must 
be performed in conjunction with testing "logical" 
connections between propositions, and recognizing 
clues, thus isolating linguistic structure (or grouping into 
segments). We are interested in specifying a processing 
model for discourse understanding that operates at the 
level of individual utterances, in the manner of the argu- 
ment model, to gain insight into how to derive linguistic 
and intentional structure simultaneously. This research is 
of significance to the current work of Grosz and Sidner 
(1986). 
ACKNOWLEDGMENTS 
This work was supported in part by the Natural Sciences 
and Engineering Research Council of Canada. I am 
grateful to the anonymous referees for their valuable 
comments and to Ray Perrault for his initial supervision 
of this research. 

REFERENCES 
Archbold, A. 1976 A Study of Some Argument Forms in a Persuasion 
Dialogue. ISI internal document, Marina del Rey, California. 
Archbold, A. and Hobbs, J. 1980 Notes on "The Analysis of Evalua- 
tive Argumentation in Text": A Report on On-Going Work. SRI 
unpublished draft, Menlo Park, California. 
Birnbaum, L.; Flowers, M.; and McGuire, R. 1980 Towards an AI 
Model of Argumentation. Proceedings of American Association .for 
Artificial Intelligence (AAAI) Conference: 313-315. 
Cohen, P. 1978 On Knowing What to Say: Planning Speech Acts. 
Technical Report No. 118, Computer Science Department, Univer- 
sity of Toronto, Toronto, Ontario, Canada. 
Cohen, R. 1980 Understanding Arguments. Proceedings of Canadian 
Society for Computational Studies of Intelligence (CSCSI) Conference: 
272-279. 
Cohen, R. 1981 Investigation of Processing Strategies for the Struc- 
tural Analysis of Arguments. Proceedings of the Association for 
Computational Linguistics (ACL) Conference: 71-75. 
Cohen, R. 1983 A Computational Model for the Analysis of Argu- 
ments. Ph.D. thesis, Technical Report No. 151, Computer Systems 
Research Group, University of Toronto, Toronto, Ontario, Canada. 
Cohen, R. 1984a A Theory of Discourse Coherence for Argument 
Understanding. Proceedings of Canadian Society for Computational 
Studies of lntelligence(CSCSl) Conference: 6-10. 
Cohen, R. 1984b A Computational Theory of the Function of Clue 
Words in Argument Understanding. Proceedings of lOth Interna- 
tional Conference on Computational Linguistics (COLING 84): 
251-258. 
Cohen, R. 1985 The Need for Pragmatics in Natural Language Under- 
standing. Proceedings of CSCSI-sponsored Theoretical Advances in 
Natural Language Understanding Conference. 
Gazdar, G. 1979 Pragmatics: lmplicature. Presupposition and Logical 
Form. Academic Press, New York, New York. 
Geva, E. 1981 Flowcharting Expository Texts and Reading Compre- 
hension. Paper presented al, Annual Meeting of American Educa- 
tional Research Association. 
Grosz, B. 1977 The Representation and Use of Focus in Dialogue 
Understanding. SRI Technical Note No. 151, Menlo Park, Califor- 
nia. 
Grosz, B. and Sidner, C. 1986 The Structures of Discourse Structure. 
Report No. 6097. Bolt, Beranek and Newman (BBN) Cambridge. 
Computational Linguistics, Volume 13, Numbers 1-2, January-June 1987 23 
Robin Cohen Analyzing the Structure of Argumentative Discourse 
Massachusetts. (also Report No. CSLI-85-39; Attention, Intentions, 
and the Structure of Discourse, Computational Linguistics 12(3): 
175-204. 
Hobbs, J. 1976 A Computational Approach to Discourse Analysis. 
Rescarch Report No. 76-2. City University of New York Depart- 
ment of Computer, New York, New York. 
Labov, W. and Fanshel, D. 1977 Therapeutic Discourse. Academic 
Press, New York, New York. 
McCarty, L. and Sridharan, N.S. 1981 A Computational Theory of 
Legal Argumentation. Research Report LRP-TR-13, Rutgers 
Univcrsity Lab for Computer Science, New Brunswick, New Jersey. 
McGuirc, R., Birnbaum, L. and Flowers, M. 1981 Opportunistic Proc- 
cssing in Arguments. Proceedings of the International Joint Conference 
on Artificial lntelligence.bf one;(lJCAl) Conference: 58-60. 
Pollack, M. 1986 A Model of Plan Inference that Distinguishes 
Between the Beliefs of Actors and Observers. Proceedings of the 
Association for Computational Linguistics (ACL) Conference: 207-214. 
Quirk, R. et. al. 1972 A Grammar of Contemporary English. Longmans 
Co., London, England. 
Reichman, R. 1981 Plain Speaking: A Theory and Grammar of Spon- 
taneous Discourse. Report No. 4681, Bolt, Beranek and Newman 
(BBN), Cambridge, Massachusetts.. 
Sadock, J. 1977 Modus Brevis: The Truncated Argument. Papers from 
the 13th Regional Meeting, Chicago Linguistics Society: 545-554. 
Sidner, C. 1979 Towards a Computational Theory of Definite 
Anaphora Comprehension in English Discourse. AI Lab Report 
TR-537, Massacusetts Institute of Technology (MIT), Cambridge, 
Massachusetts. 
Smedley, T. 1986 An Implementation of a Computational Model for 
the Analysis of Arguments: An Introduction to the First Attempt. 
Research Report CS-86-26, University of Waterloo Department of 
Computer Science, Waterloo, Ontario, Canada. 
Weiner, J. 1979 BLAH, A System which Explains Its Reasoning. Tech- 
nical Report TR 79-14, University of New Hampshire Computer 
Science Department, Durham, New Hampshire. 
