Theoretical foundations for illocutionary structure parsing 1

2.1. Anchoring arguments in dialogue

Inference Anchoring Theory (IAT) [2] provides us with a framework for connecting dialogical structures with argument structures which can facilitate mining dialogical arguments. We will first sketch the general idea of IAT on a simplified example:2

In IAT, different relations between propositions distinguished by logic and argumentation theory can be expressed. For example, (1) contains a reasoning structure consisting of a conclusion q, a premise p, and a relation between p and q called a rule application. This relation may correspond to some logical rule like modus ponens or some presumptive argumentation scheme such as argument from consequences [32] (for the purposes of the example, it is marked simply as a numbered instance of an application of some rule). Propositions may also be connected through other types of relations such as conflict, when, for example, one is the negation or some opposition of another. Conflict corresponds to the relation of attack studied extensively in abstract argumentation framework [7].

OPEN IN VIEWER

Fig. 1.

Interaction between argument and dialogue.

At the same time, (1) contains dialogical structure composing three utterances whose performance is governed by dialogue rules which express how sequences of utterances can be composed (these compositional steps are called transitions in IAT). For example, the dialogue rules permitted Wilma to perform (1-b) (i.e., to ask Bob for grounds for believing p), by virtue of the fact that at the previous move (1-a) Bob committed himself to p (see transition instance #1 in Fig. 1). In other words, the disputants followed a dialogical rule stipulating that challenging is allowed after asserting. Such a normative view of dialogue structure is quite common – in philosophy, authors such as Mackenzie [16] have explored dialogue rationality in these terms in a thread of work rooted in Wittgenstein, whilst in linguistics, approaches founded upon discourse analysis such as Dialogue Macrogame Theory [17] and the HCRC dialogue coding scheme [5] aim to account for dialogue coherence.

Characterising the application of dialogue rules specifically as transitions between locutions is rather more unusual, but is a central part of IAT, because these transitions can act as “anchors”. Logical structures, like propositions, are thus “anchored” in dialogical structures via illocutionary connections related to different illocutionary forces [22,28,29]. The illocutionary force of an utterance can be of a number of types (asserting, challenging, etc.). What type of force should be associated with a given speech act is dependent on the act’s constitutive rules. For example, in order for the connection to be reconstructed as having an asserting force, the utterance has to be a felicitous assertion. In Searle’s model, it means that it has to satisfy, e.g., the constitutive rule of sincerity, i.e. the speaker has to believe that the proposition he utters is true.

In IAT, the dialogical acts (e.g. “Bob said, p”) are linked to their propositional contents (p), by illocutionary connections (asserting instance #1). Some of the illocutionary structures, such as arguing (see arguing instance #1), are anchored in transitions (transition instance #2 in this case). The idea is that for reconstructing the asserting force of (1-c), it is sufficient to inspect the linguistic surface of what has been said in (1-c). Thus, the asserting connection is anchored in the locution (1-c). This, however, does not allow us to identify that Bob was also arguing – if I entered the room exactly in the moment of him performing (1-c), I can only assume that Bob asserted q. In order to recognise that he also argued p because q, I would have to know that (1-c) was the response to (1-b). Thus, the arguing connection is anchored in the transition (relation between locutions (1-b) and (1-c)).

Inference Anchoring Theory provides a framework for mining dialogical arguments. In this simplified case, if a software tool is able to automatically recognise that in (1-b) Wilma performed challenging p and in (1-c) Bob performed asserting q (this is the step of illocutionary structure parsing); and we identified what this sequence typically anchors in a given type of dialogue (in this example a sequence of challenging followed by asserting anchored arguing; such rules are called transition schemes in IAT); then the tool will be able to automatically identify that an argument: p because q, was performed. In natural communication, people use more complex transition schemes which are dependent on the type of interaction they are engaged (the dialogue between friends, in the court, in science and so on). The example of transition schemes elaborated for the Moral Maze type of dialogue is presented in Section 3.3.

This approach offers a significantly more fine-grained account of the context of a dialogue in which a logical inference is established, than is available in, for example, Segmented Discourse Representation Theory, SDRT [1], whilst offering an illocutionary account of structure rather than a rhetorical account – although illocutionary approaches have been criticised by some authors [22], they provide a much better account of argument and debate [24]: what we are after, therefore, by analogy to work on rhetorical structure parsing [19,20] is a mechanism for illocutionary structure parsing.

One of the challenges of illocutionary structure parsing for the IAT method is the analysis and the manual identification of types of illocutions used in a given dialogue game. The Moral Maze programme can be described as an agreement-seeking type of dialogue where participants start with the difference of opinion and look for reaching an agreement during the course of the programme. Consider an example in which two people discuss whether the British Empire behaved in an uncivilised way during a war in Kenya in the 1950s:

Intuitively, what Longley claims in (2-b) is that uncivilised behaviour was typical for the Empire, and he supports this claim with a premise that such a behaviour was the product of an unsustainable policy. On the linguistic surface, however, we do have neither conclusion nor premise asserted explicitly qua assertions: instead, the act (2-b) is cast as a series of questions. So the problem can be phrased as: how to represent the dialogical acts that typically accompany argumentation in types of dialogues as our corpus, including such contexts in which argumentation is (at least partly) formulated as interrogative sentences?

According to, for example, Searle and Vanderveken’s [29] model, the utterances in (2-b) could be represented as having questioning force. But this would not allow us to express the relation of support between the propositional contents of those utterances, because pure questions do not (by definition as not belonging to assertive acts) convey opinions. One could claim that the utterances have in fact an (indirect) assertive force, for the interrogative form of question is here only the superficial grammatical surface.3

The annotation scheme used in switchboard dialog act corpus, SwDA [30], in contrast, introduces a category of rhetorical questions, e.g. “Who would steal a newspaper?”, which are grammatically formed as questions but are used to convey opinions. Yet they have only assertive force, because (by definition) they do not invite the hearer’s response. As a result, they cannot fulfil the agreement-seeking function, since if the hearer is not allowed to respond, he cannot express whether he agrees or disagrees with the propositional content of the (rhetorical) question. In particular, if we understood (2-b) as a series of two rhetorical questions, then James would not be invited to respond and express his agreement or disagreement with the thesis that the British behaviour was the product of an unsustainable policy (while in fact in the next move James disagrees by saying: “It is the product of a policy conceived in – by the Cabinet, in the context of the Cold War, when (…) the British Government is in a near state of funk about what will happen in their Empire”).

What we need is to be able to identify the assertive intentions behind such questions. Only then can we assemble the parts of the argument and attempt to model their composition into large structures. As far as we are aware, there is no model which would allow the representation of such double (asserting and questioning) function of utterances. Thus, as a general goal we need to identify what types of text units (unitary illocutionary structures) are characteristic for our corpus and then develop a comprehensive list of these types.

In the literature, there are several models (annotation schemes) that are useful for describing various dialogue, argument and illocutionary structures. These approaches can be divided into two subclasses:

Understanding the ways in which dialogical action can establish arguments and inferences stands at the intersection of many disciplines including (at least), discourse analysis, pragmatics and semantics. IAT is constructed to benefit from insights in all of these fields, whilst focusing, quite narrowly, on argumentation as a type (or rather, set of types) of discourse. In discourse analysis, Rhetorical Structure Theory has been enormously influential in facilitating computational models of discourse structure. It postulates a number of discourse relations that can account for the overall structure of any kind of text (a logical analysis of RST is given in [8]). Within our framework, RST can be used to represent the structure of arguments (a conclusion being a nucleus and a support its satellite), and, to a larger extent, to represent the structure of explanations, clarifications, reformulations, elaborations, illustrations, conditionals, causes, etc. From that point of view, RST offers a useful means to represent the linguistic structure of a dialogue. RST does not (not even in its dialogical extensions), however, capture inferential patterns, which lie at the very heart of IAT. It is insufficient for handling fine-grained argument structures for at least two reasons [23,24]: first, many argumentative relations seem to be quite orthogonal to rhetorical relations – whether two sentences are related through Elaboration or Justification says little or nothing about whether they form a part of a modus ponens or modus tollens, or a part of linked or convergent argument structure; and second, argument structure is often missed entirely by rhetorical structure, a problem which becomes manifest in the abundance of (vapid) joint relations in RST analyses of many arguments.

In pragmatics, one of the most significant approaches to dialogue meaning has come from DRT and particularly a more recent derivative, Segmented Discourse Representation Theory. Where Inference Anchoring Theory focuses specifically on argumentative discourse, SDRT has much broader goals than IAT and, as a result, is less well adapted at handling the structure of inference-establishing discourse units. In particular, the way in which the rules of dialogue influence the construction of argument lie beyond SDRT’s assumptions of context-independent semantics [2,31].

In semantics, Ginzburg [10] in particular has had a major impact on models of the meaning of dialogical actions. His KoS approach, like SDRT, is much more general than IAT, but, also like SDRT, suffers in its ability to handle argumentation as a result. Specifically, KoS does not allow the establishment of a relation between propositions (such as inference) to be the result of a pair of locutions which together establish an illocutionary force, such as arguing. Without such pairing, understanding how challenge-response sequences establish inferences becomes difficulty.

There are also a wide range of annotation schemes used specifically for corpus analysis of different dialogical structures. Probably the most comprehensive set of annotations originates from the Switchboard Dialog Act. This set of annotations was constructed from a large set of types of dialogues and is now widely used in the dialogue community. This rich tagset offers over 200 different tags (of which 60 are basic) that account for the various forms of exchanges found in dialogues, as found, e.g., in help desks. There are, however, no specific tags that can directly account for argumentation attitudes and it does not allow for expressing different propositional or psychological attitudes, and belief expressions important from a point of view of argument recognition.

Textual entailment (TE) [6] is a general approach that establishes a relation whenever the truth of a certain statement (or text fragment) follows from another statement or text fragment. The first statement is called the hypothesis. The type of inference developed in TE is quite different than those, much more specific, which are developed in Inference Anchoring Theory. IAT exploits the semantics and the dynamics of dialogical acts in order to identify inferences.

Presuppositional logic is a framework where a number of statements, called presuppositions, are generated by the preconditions that hold in particular on predicates and dialogue acts. Questions and assertions produced by the participants of a dialogue necessarily convey a number of presuppositions, be they historical, cultural or practical. In a statement such as “It was a fight for their independence”, presuppositions are e.g. that the population which is the focus of the dialogue was under the domination of another population or country and that it was not satisfied with this domination. These inferences may also be drawn to some extent by TE. Presuppositional logic involves general purpose reasoning as well as lexical inference (inferences which can be drawn from words taken in isolation) and offers a tool that can be used to draw some of the inferences introduced in the IAT model, for example to establish causal or thematic links between statements in different dialogue turns. IAT is more powerful in the sense that it includes many other forms of inferences related in particular to argumentation.

Speech Act Theory [22,28,29] considers interactions from the point of view of the intentions of the speakers (i.e. the force they give to their statements); however, it says little about whether and how, two illocutionary forces relate to each other. With IAT, illocutionary connections between locutions and propositional contents show the speaker’s intentions, and illocutionary connections derived from a sequence of locutions show their argumentative functions.

To summarise, the approaches cited above all have advantages over IAT, but have had influence on its coherence for handling argumentation in particular. For example, RST annotation schemes have influenced local structural analysis in IAT; Speech Act Theory offers an account of speakers’ communicative intentions that IAT renders through the analysis of illocutionary structures, etc.; SDRT reifies the relationship between an utterance and the dialogical history leading to that utterance, which is found in IAT in transition schemes. The essence of IAT therefore rests on its ability to combine linguistic, argumentation and pragmatic theories in order to deliver an analytical framework for dialogical argumentative discourse, which presents major challenges for all of the alternative approaches.

3.1. Annotation scheme

We assume that the utterance F ( p ) is a unitary (non-relational) illocutionary structure with a propositional content p and an illocutionary force (illocution) of F (cf. [28]). This paper focuses on five types of unitary illocutionary structures occurring in the context of dialogical arguments in the corpus of the Moral Maze programme. They are associated with the forces of: asserting, questioning, challenging, conceding and (dis-)affirming.

Asserting. The speaker S uses an act of assertion to directly communicate his opinion about p. Asserting p does not assume S truly believes p – it is rather a public declaration of belief to which the speaker can be held.

Depending on the strength of assertive force (as [29] discuss, the illocutionary point, i.e. the internal purpose of a type of illocution, can be achieved with different degrees of strength), we distinguish three subcategories: strong assertion +A, e.g. “I certainly think people who don’t work, who are unemployed, are depressed”; standard assertion A, e.g. “It’s important to bring to justice the actual perpetrators of atrocities or war crimes”; and weak assertion −A, e.g. “Well, I guess for my definition, I would talk about rich countries in the West occupying countries of other people and imposing their rule on them”.

The rationale behind this distinction is to allow and enhance the identification of argumentative behaviours in dialogue. When the speaker performs a strong assertion rather than a weak one, it is more likely that he thereby generates an argument as a response to the opponent’s challenge. In the case where there are no linguistic cues for recognising the strength of asserting force, we use the default category of standard assertion A, and no anticipation of the argumentative behaviour can be made.

For example, if Bob expresses p with strong asserting force +A in (1-a), then it is more likely that he will react as in (1-c) in response to Wilma’s challenge. On the other hand, if he utters p with a weak assertion −A, then the dialogue could proceed differently and in (1-c), Bob could retract his commitment to p rather than argue that p because q.

Asserting force can be applied to three types of propositional contents: a strong content A+; a weak content A−; and a “neutral” content A. Strong propositional content A+ is identified when the speaker explicitly expresses some type of necessity of the truth of propositional content, e.g. “(…) for sure, there should be no statute of limitations in prosecuting them”; and weak content A−, when he expresses only the possibility of the content being true, e.g. “There might be a right to compensation in this case”. Again, the neutral content is the default subcategory, when there is no cue to associate a strong or weak type of content.

As a result, different combinations of the strengths of assertive illocution are possible, e.g. the weak assertion of a weak propositional content −A− as in: “We reckon that something like 100 billion pounds has been transferred from savers to borrowers”. In case it is ambiguous whether a linguistic cue refers to a type of illocutionary force or a propositional content like in: “(…) certainly all debt is not immoral”, we assign the strength to both of them (in this case: +A+).

Questioning. The speaker S utters a question when he formulates p in the grammatical form of interrogative sentence: “Is/Isn’t p the case?”. Here, only yes-no questions are taken into account, i.e. questions for which the possible answers in the next move in a dialogue are: “Yes, p”, “No, not-p”, or “I don’t know”.

Three main subcategories of questions have been distinguished in our corpus: pure questions PQ, assertive questions AQ, and rhetorical questions RQ. When S utters a PQ, e.g. “(…) should it be the occasion for examining the colonial legacy?”, he expresses that he is asking for the hearer H’s opinion on p: whether H believes p (manifested in the next move by the utterance “Yes, it should”), or not (“No, it shouldn’t”), or he doesn’t have an opinion on p (“I don’t know”).

AQ and RQ, in contrast, convey some assertive communicative intention. In the literature, there is no common specification of rhetorical questions. For example, in semantics, RQs are defined as denoting a truth-conditional statement (cf. [12]), while in discourse analysis the focus is on investigating the precise linguistic form of RQs (cf. [9,14,26]). For our purposes, however, it is sufficient to use a basic feature of RQs, upon which all these approaches agree [15], i.e., when S performs RQ: “Is/Isn’t p the case?”, S is grammatically stating a question, but in fact communicating that he does (or does not) believe p.

Finally, when S uses AQ, e.g. “Is it not the case that you are stretching your definition in order to make the British Empire fit into that?”, he not only directly seeks H’s opinion on p, but also indirectly publicly declares his own opinion on p. These subcategories form the continuum of the speaker’s intentions, with PQ and RQ as the extremes at either end, and AQ in the middle.

Challenging. S uses a dialogical act of challenging p in a dialogue when he suggests that he is asking about the grounds for hearer H’s opinion on p. Challenges are a dialogical mechanism for triggering argumentation. For example, in (1) Wilma’s utterance “Why p?” is a challenge, since she asks Bob about the reasons for believing p. In the next move, Bob provides a reason q and, as a result, the argument “p since q” is constructed by means of the dialogue. In the same way as for questions, challenges form a continuum from pure challenges (PCh) through assertive challenges (ACh) to rhetorical challenges (RCh).

Conceding. A concession, Cn, of p is used to convey agreement with respect to p, e.g., “I understand your point and of course there are living victims”. Concessions are like assertions, but place weaker commitments on the speaker, and as a result, in many types of dialogues S might not be obliged to defend his concession as a response to the opponent’s challenge (i.e. does not have a burden of proof [33]). The situation for disagreement is not symmetrical, i.e. when S disagrees with the opponent, he will perform just a simple assertion, because in disagreeing S clearly puts forward his standpoint and therefore he can be challenged by the respondent in the next move.

Amongst concessions we distinguish the subcategory of popular concession PCn, in which the speaker communicates some sort of general knowledge which is taken to be obvious and as such does not require to be defended (i.e. does not place a burden of proof on the speaker): “We know Kenya was a brutal war, …”. Popular concessions are often introduced to the discussion in order to partly agree with the opponent and with generally accepted truths, but at the same time to prepare ground for expressing disagreement in the next statement (with the contrast typically signalled with “but”) and showing that the real disagreement in the discussion lies elsewhere: “we have known about these things for years, there is nothing new in this, in Kenya and the problems we had in India, end of Empire is always a difficult bloody thing, but to pin the blame on the British for this is simply absurd, it’s not true”.

Affirming and disaffirming. The illocutionary structures that do not have explicit propositional content and are performed as a reaction to the opponent’s question are called affirmations Af (for a positive reaction, e.g. “Yes”, “Indeed”, “Most definitely”) and disaffirmations Df (for a negative reaction, e.g. “No”, “I’m not saying that”, “Actually, that’s not correct”).

Observe that when the broader dialogical context is not known and an assertion is examined in isolation, such as e.g. “these crimes were committed”, it is impossible to know what other dialogical act may be being made, e.g. whether it constitutes disagreement with what the opponent previously stated, or argumentation supporting some previously claimed opinion. Similarly, when an affirmation is put forward, such as e.g. “Oh yes, of course”, it is neither possible to know what other act may be being performed, nor even what propositional content is being referred to when a speaker says, “yes”.

3.2. A corpus for illocutionary structures

The corpus describes the genre of the moral debate and comprises of three episodes of the Moral Maze BBC Radio 4 programme in 2012 in which the total number of 12 speakers participated. The episodes have been transcribed to support IAT annotations: the transcripts presented punctuation but no prosodic marks. In Moral Maze, participants discuss moral aspects of important social and political issues in Great Britain. The programme is chaired by Michael Buerk who leads the discussion between four panellists i.e. public personalities with a background in social activism (writers, journalists, lecturers, public commentators etc.), and (usually) four witnesses who are invited because of their expertise on a given topic. These well-structured debates present some very long sentences, several types of statements, discourse regulators and other forms of dialogue management. The language used by participants contains much explicit signalling and speakers’ positions are made clear in terms of e.g. beliefs, strength of position or of assumptions [3]. The annotation task is still in progress aiming to have analysed all episodes from the year 2012.4

The global characteristics for our corpus are described in Table 1. The corpus was divided in two sub-corpora: development and test. Analysing texts according to IAT is particularly time-consuming task because of the large variety of schemes and categories, hence the relatively small size of our corpus. It remains, however, the largest publicly available corpus of analysed dialogical argumentation currently available.

Table 2 shows the distribution of the different illocutionary structures which were manually annotated. It demonstrates the prominence of pure assertions (noted as A) over the other types: 124 questions have been identified as opposed to 317 assertions, which is relatively well balanced. Discourse regulators, meant to manage the overall discussion, occur on average every 15 assertion or question discourse units, which shows the vitality of the discussions and the diversity of the sub-topics addressed. These figures are also useful to manage priority identification in the implementation.

Using the taxonomy described in Section 3.1, together with the IAT framework described in Section 2.1, we are now able to adequately represent the dialogical context of arguments in example (2). The utterance (2-a) is an assertion (see asserting instance #1 in Fig. 2), in which James publicly declares his belief that the uncivilised behaviour of the Empire was a ghastly aberration. The first utterance in (2-b) is RQ (see rhetorical questioning instance #1), and as a result, we recognise that in saying “Or was it in fact typical?” Longley is not searching for James’ response, but is instead declaring his own belief that uncivilised behaviour was typical for the Empire.5

OPEN IN VIEWER

Fig. 2.

Asserting, rhetorical questioning and assertive questioning as the dialogical context for argument structures in example (2).

We can thus extract the contents of RQ and AQ, which allow us to identify a relation between them that constitutes an inference (inference instance #1) with the illocution of arguing (arguing instance #1) and thence to recognise the remaining components such as, e.g., the fact that Longley’s argument is a dialogical response in which he communicates his disagreement (disagreeing instance #1) and constructs a conflict (conflict instance #1) between the conclusion of his argumentation and the content of proponent’s utterance (2-a).

Inference Anchoring Theory is not a simple composition of a unit-based approach and a relation-based approach. It provides a framework of constraints on how particular unit types (unitary illocutionary structures) constrain the available relation types (relational illocutionary structures). As a result, IAT delivers a rich set of features for recognising arguments as dialogical sequences of, for example, assertion and rhetorical questioning, or rhetorical questioning and assertive questioning. The connections between unitary illocutionary structures and relational structures can be then operationalised in standard statistical machine learning techniques.

4.1. A grammar for text unit identification

Given a dialogue transcript, the first task is, for each dialogue turn, to identify the basic dialogue units (also called argumentative segments). These basic units are the minimal units, in terms of contents, which are autonomous and can be connected to other units via various types of dialogue relations. Our analysis posits that in a coherent argumentative text, these units are all linked to each other, and form a graph, as shown in Figs. 1–2. No unit is left pending: all units are connected to others. The structure is however not as hierarchical as discourse analysis in general, and there is no a priori notion of nuclearity.

For that purpose, we developed an elementary discourse unit (EDU) [27] analysis dedicated to dialogical situations. These specific EDUs turn out to have specific forms and linguistic marks proper to argumentative dialogue compared to those defined for discourse analysis in general. According to the observations made on our corpus, in most situations, units can be identified on the basis of discourse marks typical of dialogue, e.g. marks related to challenge, position or belief statements, or an aggregation of such marks, e.g. when there is an intention of persuasion. Marks are developed below. Since our corpus was rather small at that time, and due to the well-known linguistic sparsity of some constructions, besides corpus analysis, we added closely related terms, such as quasi synonyms. At the time of publication, the corpus has grown to a much larger size, making EDUs analysis much more reliable.

A number of psycholinguistic investigations summarized in [11] show that marks are used by human subjects both as cohesive links between adjacent clauses and as connectors between larger textual units. An important result is that discourse markers are used consistently with the semantics and pragmatics of the textual units they connect and they are relatively frequent and unambiguous. The semantics and pragmatics of a number of connectors are investigated in e.g. [34].

A grammar that allows for the identification and the delimitation of these text units can be elaborated on the basis of these marks and punctuation. The main categories of marks used to identify text units include:

verbs, which promote controversies, beliefs, position statement and argumentation. The following classes are the most prominent: propositional attitude verbs (think, believe, agree, deny), epistemic verbs (know, understand), communication and report verbs (claim, hold) and psychological verbs (dream, worry, be intrigued); we also noted a few metaphorical uses (it tends to);

To handle these marks and their associated linguistic elements (e.g. subjects, pronouns, negation, modals, etc.), potential synonyms, and morphological variations, we developed ‘local’ grammars dealing with e.g. propositional attitude expressions, position statements, psychological expressions, questions, etc. in their linguistic diversity. A total of 52 rules have been developed for that purpose. These rules implement what we call g-marks (grammaticalized marks), e.g. informally: g-mark(Type) –> pronoun, opt(negation), verb(Type), where the negation is optional. Type is the category of the verb, e.g. prop-attitude.

Rules of the same type have been developed for modals, opinion expression of question formulation, such as, with the same notations: g-mark(Type) –> pronoun, opt(Adverb), verb(Type), where Adverb is one of the terms called ‘opinion expression adverbial’. g-mark(Type) –> pronoun, modal( + negation), verb(Type). g-mark(Type) –> interrogative, opt(modal), verb(Type, + infinitive).

These rules were designed manually and emerged from corpus observations, leading to the above categorizations, and to generalizations (e.g. including all the relevant propositional attitude verbs in the lexicon even if only a few have been observed) in order to have an adequate linguistic coverage.

These local grammars are then integrated into larger grammatical forms which both identify and delimit the text units. The general form of these latter rules follows the Dislog syntax, which is roughly as follows: unit_identifier –> unit_delimiter, gap, (g-mark, gap)*, unit delimiter, where (g-mark,gap)* indicates a sequence of one or more g-marks, a g-mark being a call to a grammaticalized mark.

The symbol “gap” indicates a finite set of words which are of no present interest which can be skipped. For example, in: “Well, isn’t that a source of injustice?”, the first unit delimiter is the beginning if the sentence, the g-mark is isn’t that and the second unit delimiter is the question mark. The two gaps are respectively: well, and a source of injustice. In a number of cases where the unit delimiter is at the beginning of the unit, the gap is empty.

The grammar is implemented using the <TextCoop> platform and the Dislog language [25], specifically designed for discourse processing. Dislog extends and generalizes the expressive power of regular expressions in several ways, most notably via the introduction of typed feature structures, controls over skipped structures (gaps) and its ability to include reasoning schemas, e.g. to resolve ambiguities or to compute a result. In general, the result of an analysis is an XML tagged text; it can also be a dependency structure.

We have conducted an indicative evaluation (to identify improvement directions) on the previously manually annotated test text where 179 text unit occurrences have been identified. This is a small scale test but it turns out to be sufficient for a first analysis. Out of this set of 179 units, our system:

correctly annotated 153 units (85%) (identification and delimitation);

4.2. Illocutionary structure identification

The next step is to identify the illocutionary structures (depending on the type of an illocutionary force) presented in Section 3.1 for each text unit. A coherent dialogue requires that each text unit is assigned a type. Unlike at the previous step, it is not possible, however, to identify all illocutionary force types solely on the basis of linguistic marks. In some cases the illocutionary force cannot be assigned to a unit in isolation: it is necessary to take into account e.g. the illocutionary forces assigned to adjacent units, the position of the unit in the dialogue turn (starting, ending) and the role of the speaker in the debate (e.g. moderator).

Linguistic marks alone do, however, provide substantial information about illocutionary structure, and, in this preliminary analysis, we present the main results obtained on that basis. Our observations show that the linguistic criteria used to identify illocutionary forces are within each text unit, without any reference to other text units. Linguistic forms and aggregation of simple forms used to identify the types of dialogical acts are different from those used for text unit delimitation. These linguistic forms have been identified from our corpus, and then expanded to similar terms and represented as above for text units by means of g-marks. The technique and the current limitations (relatively small corpus) are the same as for the previous step. However, with the current development of a much larger and stable corpus, these results will most probably largely evolve and be improved.

A significant number of linguistic forms we have identified are very general and are shared by several illocutionary forces. Our strategy is not to have default illocutionary force assignment, but to explicitly represent the ambiguity, which may be resolved later by other means. For that purpose, we introduce polymorphic types that represent the ambiguity (or underspecification). For example, AQ–RQ is the polymorphic type assigned to a text unit which can be one of the two types AQ or RQ, when it was not possible to make a decision between these two types. Of interest is the study of the overlap situations, so that the pragmatic factors at stake can be better identified. Polymorphic types concern questions or assertions separately, since these two main categories can be relatively well identified via punctuation.

Linguistic forms are quite diverse (42 forms have been identified), with a large number of morphological variants. 12 types, simple or polymorphic, have been defined for questions, with an average of about 35 forms with morphological variants for each. Forms which are common to two or more basic types are then associated with the polymorphic types. These forms are implemented in the Dislog language, using g-marks as symbols in the rules.

From our analysis, it turns out that about 63% of the linguistic marks are typical of a unique illocutionary force, therefore proper to specific types. We have, for example the following types of marks (implemented in g-marks):

To identify illocutionary forces we have then written 18 rules in Dislog that implement the g-mark recognition and illocutionary force assignment.

Considering our corpus of three manually annotated texts, we have the following distributions for polymorphic types:

This means that about 62% of the text units are a priori unambiguous w.r.t. their illocutionary force, which is a rather low proportion. It is thus crucial to identify additional factors (pragmatic, typographic) that contribute to resolve the ambiguities.

To carry out a simple indicative evaluation, let us consider the 153 text units that were correctly identified, as indicated in the previous section, we obtain the results given in Table 3. In the first case, the assignment is made for the basic type of illocutionary forces, so there is no ambiguity. In the second case, it is important to note that 65% of the polymorphic types is the combination of different strengths of assertions where the distinction, even for humans, is difficult to make. The erroneous assignment can still be improved by some typographic and linguistic adjustments. These results – of 78% accuracy with polymorphic types – although still preliminary, are very encouraging considering the difficulty of the task.

The results of the parse is annotation over the original text, as shown for example (2):

<utterance speaker = “lj” illoc = “standard_assertion”> <textunit nb = “215”> it was a ghastly aberration </textunit> </utterance>.

<utterance speaker = “cl” illoc = “RQ”> <textunit nb = “216”> or was it in fact typical? </textunit> </utterance>.

<utterance speaker = “cl” illoc = “RQ–AQ”> <textunit nb = “217”> was it the product of a policy that was unsustainable that could only be pursued by increasing repression? </textunit> </utterance>.

An internal representation, in Prolog, has been developed in parallel with this XML annotation, so that (1) these two steps can be pipelined and connected to the following ones (transition analysis in particular) and (2) OVA+ diagrams can be generated.