3.1.1.1. Discourse versus speech

Like the French linguist Gustave Guillaume, some consider that discourse, being the language implemented or the language used by the speaking subject, can be considered to be a synonym of speech (see [DUB 71]). Linguists in this current prefer the opposition language/discourse to the Saussurian dichotomy language/speech because the term speech refers exclusively to spoken language.

3.1.1.2. Discourse versus sentence

Between the sentence and discourse, the boundaries seem relatively clear because it is generally accepted that discourse is a supra-sentential entity. However, as noted in [BEN 66], the nature of relationships between sentences is completely different from relationships between sub-sentential units like phonemes, morphemes or syntactic units. This makes works about discourse a specific domain within linguistics.

3.1.1.3. Discourse versus narrative text

According to the French linguist Emile Benveniste, the situation of utterance production is the difference between discourse and narrative text. Normally constituted discourse refers to the space-time of the utterance production and the here and the now refer to the place and time of the utterance production, respectively (see [LAC 97] for a detailed discussion). The narrative text, generally a set of written utterances, is characterized by its detachment from the situation of utterance production with a total absence of the personal pronoun I. Thus, in a narrative text, the third person is most common and the past tenses are used, such as the simple past and past anterior.

3.1.1.4. Discourse versus text

The most delicate distinction concerns the differentiation between discourse and text. Indeed, several currents differ on this subject. In order to give an idea of these divergences, Table 3.1 summarizes the main perspectives.

All of these points of view have advantages and disadvantages. From an automatic processing perspective, it seems that the point of view presented in [GAR 03] is particularly pertinent. Like Kamp and Reyle [KAM 93], he considers that discourse is a sequence of connected sentences and that this sequence has its own semantic representation. Therefore, the interpretation of discourse is considered to be an incremental process, where the semantic representation of a sentence must be connected to that of preceding sentences. A text, considered as a sequence of discourses, is defined according to the nature of the connections that exist between the discourses that compose it. This led Joaquín Garrido [GAR 03] to consider that the nature of these connections is what determines the genre of the text.

Some, like Richard Party [PAT 93], believe that discourse can be defined as an entity that has its own characteristics like uniqueness, continuity, intentionality and topicality. Uniqueness allows discourse to be perceived as an independent whole. Continuity, guaranteed by the presence of particular elements like discursive connectors, emerges from extra-sentential semantic phenomena. Intentionality focuses on the fact that discourse is above all an act of communication through which the locutor interacts with one or several interlocutors. Topicality pertains to the fact that a discourse must concern a specific subject that is identifiable by the locutor.

Generally, the actualization of discourse involves several levels of structuring at the same time. Such complexity gives rise to individual variations that are not always possible to predict, hence the difficulty of formally modeling discourse. Similarly, this multi-level nature opens the door to fairly varied interdisciplinary works with the objective of clarifying its different facets: psychological, social, ethnological, etc.

3.1.2.1. Deictics (shifters)

The anchoring of utterances in a discourse is ensured by a particular category of words: deictics or shifters. This is a class of words that do not have a proper reference in language and only receive a meaning when they are included in a message. Linguists like Benveniste and Jakobson studied how facts of language relate to the context of utterance production. Deictics include a relatively large set of grammatical categories such as demonstratives, adverbs of time and space, personal pronouns and articles.

Deictics can refer to several aspects of the elocution context like:

• – The speaking subject or the modalization. For example, I told it to him. Other categories of words like adjectives and possessive pronouns like my, mine or Papa can also be associated with this category.
• – The time of the utterance production through words like now, today, tomorrow, in a week at the moment of the utterance. For example: The weather is nice today. Jakobson considered that because of their capacity to signal an anterior or posterior event at the time of the utterance production of the message, verbal tenses should also be included in this category.
• – The space in which the utterance is produced through words like here, beside and over there. For example, you can set that here.

3.1.2.2. Participants in a communicative event

In the classic models of communication, such as those presented by [DE 16, SHA 48], two participants are involved in the communication. One is charged with encoding or producing and the other is charged with decoding or understanding. In the context of utterance production, things are not so simple. Such analyses have identified four roles that can be filled by one or more people according to the case.

• – The locutor is the one that physiologically emits the (oral) message, or the writer or author of written messages. It is worth mentioning that the subject of the utterance (grammatical subject) does not necessarily correspond to the subject of the utterance production (locutor). For example, in the sentence Paul eats an apple, Paul is the subject of the utterance whereas the subject of the utterance production or the locutor is the person who pronounced this utterance.
• – The utterer is at the center of the communicative act because this is the participant who assumes the responsibility for the content of the message. The locutor and the utterer can be the same person or a different person.
• – The allocutor is the participant to whom the message is materially intended. This is sometimes also called the receiver or the recipient.
• – The addressee is the participant for whom the message is formulated. Note that, unlike the utterer, which is always singular because even in the case of we there is always an I behind it, the addressee can be plural, because a theoretically unlimited number of people can be addressed at the same time.

An extreme case of a communicative act is a monologue, when a person of sound mind talks to himself. In this situation, all of the participants are the same person. This observation is only valid in true monologues that can be observed in real life. In false monologues, which are relatively frequent in artistic works like operas, novels and plays, the monologue is used to reveal the feelings or thoughts of a character to the audience. In that case, things are quite different. First, the locutor and the utterer are no longer the same person: the locutor is the actor and the utterer is the character being played by the actor, and sometimes the author of the text performed. Second, even if the person is talking to himself, the true addressee of the message is the audience.

Another case that should be mentioned is quotes or indirect or reported speech. In this case, the locutor borrows the comment of another person, be this in a literal way or not. For example, in the case of [3.1], the utterance is pronounced by the professor, who is in fact the locutor, but the content is visibly not assumed by the professor. Rather, it is attributed to Michael, who is considered the utterer.

The history professor said to his students, “Your classmate Michael wrote, ‘World War I took place in the 19th century.’” [3.1]

Note that the distinction between the utterer and the locutor is not always clear. Indeed, between a total assumption and a complete detachment from the message’s content, there are nuances that can be configured by the locutor, notably through the use of the conditional or modal adverbs like most likely, maybe and probably.

There are also many cases where the allocutor and the addressee do not correspond to the same participant. This case is especially common in journalistic interviews on television or in the written press. The interviewee, in responding to questions, speaks to the journalist, the interviewer, who is the allocutor, but the real addressee of the message is the audience who watches or reads the interview.

When a message involves several participants (voices) during its production and/ or is intended for different audiences during its reception, this is called polyphony. Initially observed by Charles Bally and Mikhail Bakhtin, this phenomenon was taken up and developed by [DUC 87]. Note that Gérard Genette, the eminent French literary theorist, was also interested in similar cases in literature where it is possible to distinguish between the author, character and locutor. These literary studies were conducted under the label of narratology [GEN 07].

Finally, it is important to mention that the distinction presented here is not unanimously accepted by the community. For more details about the differences surrounding this issue, consult [RAB 10].

3.1.2.3. Time and space adverbials

In a discourse, a chronology makes it possible to situate events in relation to the moment of utterance production. It can be marked through a multitude of grammatical means such as articulators like first of all, in the first place, then, having said that and finally; conjunctions like when, once and after; time adverbs like today, yesterday and tomorrow; and nouns indicating time like Monday, night and evening. The use of verbal tenses like the present, future and past is the simplest way to express the chronological order of events.

The spatial situation of events is generally indicated by space adverbials like here, there, elsewhere and around. The adverbial here can be analyzed as the place where the I commits the act of utterance production.

3.1.4.1. Topic and comment

The topic is what the text/discourse is talking about. It is part of the information that is generally accepted or known by the interlocutors, either because it is provided by the text/discourse prior or because it is shared by the interlocutors. This information can concern generalities about the world like the sky is blue, gold is expensive and the moon rotates around the Earth, or individual information shared by the members of a given group such as members of the same family, work colleagues or members of a sports team.

The topic depends on the order of the sentence. Thus, two sentences that are considered semantically equivalent can have different topics if their elements are ordered differently. For example, in sentences [3.3], the topic in (a) is the agent the cat which in (b) becomes the patient the mouse:

In the case of sentences with an indefinite subject, the semantically empty subject has only a purely formal role in the sentence and consequently does not have a topical role (see sentences [3.4]):

The comment or rheme concerns what the text says about the topic. It consists of new information that the text provides about the topic (see [3.5]).

3.1.4.2. Topical progression

Because a coherent text does not necessarily concern only one subject or topic, it is essential to study the mechanisms of topical progression. Topical progression pertains to the relationship between the topic of a sentence with the topic and comment of previous sentences. Four forms of topical progression are often cited in the literature: constant topic progression, linear topic progression, derived topic progression and inserted topic progression.

In the case of constant topic progression, the topic is shared by several successive sentences. This type of progression is particularly useful when describing an object or a character by mentioning its different properties and actions. This form of succession is commonly observed in narrative texts. Very often, the grammatical topic is confused with the grammatical subject. The use of anaphora is the natural method for this kind of repetition. Schematically, this kind of progression can be presented as shown in Figure 3.1.

To make this type of progression clearer, examine text [3.6] where the topic Fabrice is repeated in the sentences either by the personal subject pronoun he or by the possessive adjective his:

Fabrice is a graduate of the Faculty of Pharmacy in Boston.
Two years ago, he went to do a six-month internship in
Copenhagen at a famous Danish company. There, he learned
many things about the pharmaceutical industry. Upon his
return to the US, he managed to get a good job in New York. [3.6]

In linear topic progression, each sentence activates another one. This occurs through a comment/topic exchange: the comment of one sentence becomes the topic of the following one and so on. A diagram of this kind of progression is provided in Figure 3.2.

Ashley listened attentively to the words of Marilyn, her best friend. These words conjured up the bitter memory of her breakup with her ex-fiancé Mark. It was this break-up that pushed her to attempt suicide and then to shun everything to do with men. [3.7]

With derived topic progression, there is a main topic presented first and then derived (less central) topics follow. Often, the relationship between the main topic and the derived topics is meronymic (whole/part) or hyponymic (type/sub-type). See Figure 3.3 for the diagram of this progression.

In passage [3.8], the relationship between the topic of the first utterance, Venice, and the topics of the subsequent utterances is meronymic. Canals, castles and gondolas are all components of the city of Venice:

Venice is by far the most beautiful city I have ever visited. Its canals made me forget about the stress of everyday life. Its marble castles made me think of the adventures in One Thousand and One Nights. The gondolas reminded me of the romantic period of my engagement. [3.8]

Finally, with inserted topic progression, it is sometimes possible to insert a local topic related to the main topic of the text within a constant topic progression. This results in the diagram provided in Figure 3.4.

Consider example [3.9] to illustrate this kind of progression. In this text, the discursive topic is Frank’s car. A new local topic is added in the following sentence, which is Frank repeated by the anaphor this man. This does not prevent new comments on the discursive topic from being added to the third utterance:

Frank’s car is really superb. This man, who is passionate about mechanics and everything related to the automobile industry, takes great care with what he owns. Even though it was manufactured in the 1960s, the car is in excellent shape. [3.9]

It should be noted that, naturally, authors do not hesitate to combine the types of topical progression to avoid monotony in their texts even if a particular form of progression generally prevails depending on the genre of the text. This wealth of types of topical progression makes the automatic detection of topics addressed by a text particularly delicate work.

3.1.4.3. The implicit, the presupposed and the assumed

The interpretation of an utterance does not only occur through the elements that are explicitly present. Often, it requires inferences about new elements (not given in the sentence) from the given elements. This information is called the implicit. The study and formalization of these interpretive mechanisms of this kind of utterance attracted the interest of several researchers (see, for example, [KEE 71, KEM 75, DUC 69]).

For example, from an utterance as simple as Close the door, we can infer the following information about this utterance:

1. a) The locutor has a certain authority over the allocutor, as indicated by the use of the imperative.
2. b) There is a door in a space nearby and it is open or at least that is what the locutor thinks.
3. c) The locutor thinks that the allocutor is capable of closing the door.
4. d) The locutor has a specific goal behind this request: he is cold, he wants to have a private conversation with someone in the room, he wants to start his break, etc.

In this example, the inferences a, c and d were made based on the utterance context whereas b can be inferred in all possible contexts because it depends on the verb to close. Some researchers prefer to talk about the implicit of the utterance production (the information in a, c and d) and the implicit of the utterance (the information in b). A classification on similar criteria consists of distinguishing two sub-types of the implicit: presupposition and the assumed.

A presupposition is information deduced from a lexical unit in the utterance. Such a unit could be an adverb like again, already or always that suggests an event prior to the moment of utterance production. It can also consist of an adjective, a verb, a superlative, etc. Consider the examples in Table 3.2.

The main difference between the assertion and the presupposition is that only the assertion is the focus, which is why it alone is affected by negation. For example, in The volcano in Paris is erupting, there are two distinct items: 1) there is a volcano in Paris (presupposition) and 2) this volcano is erupting (assertion). In the negative form the volcano in Paris is not erupting, it is the assertion that is refuted, the fact that the volcano is erupting. The presupposition remains valid, although this does not necessarily mean that it is true.

On the other hand, the assumed is information that the locutor implies without giving it and without the presence of a linguistic cue in the utterance that would make it possible to deduce it. Consider utterance [3.10] that can receive one or more of these interpretations: let’s go to another location, close the window, turn down the volume on the television, I cannot hear you clearly.

It is very noisy here. [3.10]

Thus, the addressee of the message with an implicit is obliged to infer the information because, taken literally, it does not have any significance for him or at least it seems marginal. In other words, in this case, the allocutor is entirely responsible for the interpretation constructed. Consequently, it can be said that, unlike the presupposition, the locutor has the possibility of negating the assumed. In practice, because it does not directly engage the locutor, the implicit is a very good way to make an observation to someone or to make a potentially embarrassing request.

3.1.4.4. Given versus new

The dichotomy given/new concerns what is known in opposition to what is unknown. Traditionally, two types are distinguished by the use of syntactic criteria: indefinite articles are used to mark nominal groups that correspond to unknown entities and after their introduction into the discourse, these entities are referred to by definite articles or anaphora ([HAR 51] cited in [BRO 88]).

Today, there are several diverging points of view regarding the specification of given or unknown information [CHA 76, CHA 87, HAL 70]. This presentation will adopt that of [PRI 81] who considers the different aspects of the problem. Starting from the observation of the simple dichotomy of new/old to consider the complexity of the organization of the information in the discourse, Prince proposes a taxonomy that includes three classes: new, inferrable and evoked. She proposes a tree diagram of this taxonomy, which is shown in Figure 3.5.

The term new concerns new information introduced to the discourse. There are two sub-types: new information for the interlocutor and new information for the discourse. In the first case, a completely new entity is introduced for the interlocutor for which he must entirely create the mental representation. In the second case, where an entity that is already known to the locutor is introduced into the discourse, the interlocutor must update his representation of the discourse to include this new entity. Consider the utterances in [3.11]. Assuming that these utterances are found at the start of a discourse, the entities the sea and Paul Broca, newly introduced in the discourse, are assumed to be known by the interlocutor. On the other hand, the entities a person and a store are new to both the interlocutor and the discourse. It should also be noted that Prince distinguishes between two types of these entities: anchored entities and unanchored entities. Anchored entities including the ones such as a person are considered as such because they are syntactically related to a nominal group that is not new, I:

1. a) The sea is very rough today. [3.11]
2. b) Paul Broca was born in Gironde.
3. c) A person with whom I traveled told me she read your book.
4. d) I entered a store where the manager was very angry.

Evoked entities are both known by both interlocutors (shared or supposedly known knowledge between the two interlocutors) and already introduced into the discourse. Prince distinguishes two sub-types of evoked entities: textually evoked entities and situationally evoked entities. Textually evoked entities are entities that have already been mentioned in the text, often by means of third person pronouns or a nominal group with a definite article. Situationally evoked entities are entities that have been indicated by a person, time or space deictic and whose interpretation is anchored by the situation.

Contrary to evoked entities, inferred entities do not have an explicit antecedent. Their presence can be guessed using an element that serves as an indirect antecedent. As emphasized in [CHA 76], these entities, which are not exactly new, are not entirely given either, because they cannot be replaced by an anaphor. Thus, inferred entities can be considered to be an intermediary category between given and new. For example, in [3.12], the use of a definite article before cotton wool indicates that even if this entity is being introduced for the first time in the discourse, its presence in the first aid kit can be inferred logically. This means that first aid kit can be considered to be an indirect antecedent of cotton wool:

We checked the first aid kit. Everything was there but the cotton wool was missing. [3.12]

3.1.6.1. Lexical cohesion

To repeat a word from a previous sentence in a new sentence, anaphora seems the most natural means. However, two other important processes can be cited: repetition and the use of a lexical chain.

The simple repetition of a lexical unit several times is shown in [3.19]. This kind of repetition is often used in literature in order to emphasize something or to provide a sense of monotony or weight.

Sam is looking at the flowers. Sam is walking in the garden. Sam is going to school. [3.19]

A lexical chain consists of one word with lexical substitutes that are distributed throughout a given text. The relationship between the lexical units in a chain can be synonymy, antonymy or belonging to the same lexical field. Consider fragments [3.20]:

The dog is an incredible animal. Man’s best friend does not hesitate to risk its life to save its master. (Synonym)

The dog is an incredible animal. Sometimes, this canine/mammal/carnivore does not hesitate to take risks to save its master.
(Hyponym/Hyperonym) [3.20]

The dog is an incredible animal. Like the cat, it is a top choice for a pet for many people. (Semantic field: domestic animals)

3.1.6.2. Anaphora

Anaphora is a very important way to ensure a link between different discursive units in speech and in writing. However, its role is more central in speech than in writing due to the dialogic structure that implies an exchange between two interlocutors and therefore requires referencing previously mentioned fragments of discourse.

The definition of anaphora that is generally given is that anaphora is a mechanism that relates two linguistic units. The first unit is often pronominal (a personal or demonstrative pronoun) and is called the anaphor. The second unit is an anterior segment, typically a nominal group (see [DUB 71] for example). Consider utterance [3.21a]. In this utterance, the nominal group Peter is repeated by the anaphor him. Note that sometimes the order can change, which means that the anaphor can come before or after the nominal group [3.21b]. This is called cataphora:

1. a) Peter, I see him often.
2. b) I often see him, Peter. [3.21]

A more specific definition of anaphora was proposed in [KRA 00]. It is based on several criteria to interpret anaphora like contextual dependency, the type of antecedent, the type of relation between the anaphor and the antecedent, and the interval of interpretations authorized by the anaphora. In a dialogue, two types of antecedents can be distinguished. The immediate antecedent concerns cases where the anaphor and the antecedent occur in the same conversational turn, as in utterance [3.21a]. In the case of a distant antecedent, the anaphor and the antecedent occur in two conversational turns that are potentially spoken by two different locutors. As an example, consider the segment in [3.22] extracted from a dialogue in a hotel reservation corpus collected at the CLIPS-IMAG laboratory [HOL 97]. In this extract, the anaphor it and its antecedent occur in two conversational turns of two different locutors, the customer C and the hotel receptionist H, respectively. Note the formal ambiguity of the attachment of the anaphor because in the utterance H, there are several nominal groups that could be candidates: a room, a person, a shower:

H = So, I have a room for one person with a shower and WC on the fourth floor looking out on the garden for 380 dollars, breakfast included. [3.22]

C = Very good, I’ll take it.

There are even more complex cases where the anaphor returns throughout several segments of the discourse. In these cases, the anaphora is very difficult to detect automatically because it requires a very large contextual window that often contains several ambiguities to be considered.

Several linguistic factors make it possible to find the referent of an anaphor. Some are decisive and therefore impossible to violate while others are facultative but they allow the search field to be limited.

• – The number, person and gender: obligatory, these distinctive features can direct the search (see examples [3.23]):

• – Novelty: from a cognitive point of view, in case of a multitude of possibilities for the resolution, the most recent element, being the most prominent in the memory, has the advantage. Consider example [3.24]:

• – The topical role: an agent in a sentence has more chance of remaining in a second sentence. Examine examples [3.25]:

• – Repetition: the repetition of an anaphor throughout a text makes it possible to identify the referent in all occurrences of this anaphor. Examine text fragment [3.26] where the anaphor she is repeated several times, making it possible to identify that the referent is Cynthia and making it unlikely that Martha is the referent of the utterance: she went...

Cynthia is a nurse. She likes her job a lot because she loves helping people. She works with her friend Martha in the same hospital. During the holidays last year, she went to Morocco with her husband. [3.26]

• – Reciprocity or parallelism: sometimes the agent of a sentence becomes the patient or the recipient in another sentence and its patient or recipient becomes the agent in the other sentence. This logical reciprocity allows the referent of the anaphor to be deduced. For example, in passage [3.27], the agent in the first utterance, John, becomes the recipient in the second utterance, while the recipient of the first utterance becomes the agent in the second utterance:

John gave a bouquet of flowers to Mary, Maurice’s sister. Mary gave him a pen (him = John). [3.27]

• – The semantics of the verb: sometimes the semantic relations between the arguments of certain verbs imply relations between the actants of the verbs, as in [3.28] where the agent of the verb to send is the recipient of the verb to receive and vice versa. This helps deduce that the referent of the anaphor he, the agent of the verb to receive, is Max:

Fred sent a postcard to his work colleague Max. He was very happy when he received it. (He = Max) [3.28]

3.1.9.1. Locutionary acts

These are acts that are accomplished as soon as a meaningful utterance is uttered orally or in writing, independent of the communicative role of this utterance. Consider utterance [3.33] whose phonetic, morphosyntactic and semantic formulation is a locutionary act.

3.1.9.2. Illocutionary acts

An illocutionary act is carried out on the world by producing the utterance. For example, a judge saying I declare that court is in session will effectively open the session. Austin describes five major types of illocutionary acts:

• – Verdictive: acts pronounced in a judicial and legal context. For example, the following verbs express a verdictive act: to declare guilty/innocent, to condemn, to decree, to acquit.
• – Exercitive: acts related to the exercise of authority generally in contexts as varied as military, administrative or familial contexts. The verbs to command, to order, to rule, to dictate, to prescribe, to pardon are examples of these acts.
• – Permissive: acts that commit the locutor to carry out actions in the future. For example, verbs like: to promise, to commit, to bet, to gamble, to guarantee.
• – Behabitive: acts related to behavior expressed by verbs like: to thank, to bless, to praise, to criticize.
• – Expositive: acts concerning the relationship between the locutor and the information that he transmits. For example, to affirm, to certify, to assure, to deny, to note, to notice, to observe, to discover.

It should be noted that John Searle conducted a critical study of Austin’s taxonomy and concluded by proposing his own. Like Austin’s taxonomy, it has five groups: assertives (assertion, affirmation), directives (order, command, demand), permissives (promise, offer, invitation), expressives (thanks, praise) and declaratives (declaration of marriage, nomination for a prize).

3.1.9.3. Perlocutionary acts

These are acts that either have a voluntary effect or not on the allocutor. They are the consequences of an illocutionary act:

1. a) I will take you to the park next week. [3.34]
2. b) I declare that court is in session.

Beyond the locutory act related to the formulation of the utterances in [3.34] (both a and b) and the illocutionary act related to the commitment, utterance [3.34a] has a possible perlocutionary effect, which is to persuade the allocutor of the kindness of the locutor. If we assume that utterance [3.34b] was produced by a judge in a court, it has the perlocutionary effect of attracting the attention of the listeners to the judge, putting an end to the lateral conversations of the people present, etc.

Finally, it is useful to mention that the theory of speech acts saw important developments in the 1990s, notably in the works of Sperber and Wilson on Relevance Theory in the cognitive pragmatics movement [SPE 89] and the works of Albert Assaraf on Link Theory [ASS 93].

3.2.4.1. The naive syntactic approach of Hobbs

Syntactic approaches are used to filter a pronoun’s antecedent candidates. As noted, syntactic factors are important but not sufficient to determine an antecedent. The basic idea of Hobbs’ algorithm [HOB 76, HOB 78] is to move through the syntactic parsing tree of a sentence searching for a nominal group that satisfies predefined constraints like number and gender. More specifically, the search is carried out according to the following steps [HOB 78]:

1. 1) Start with the node of the Nominal Group NP that immediately dominates the pronoun.
2. 2) Move higher in the tree until the first node NP is encountered or until the node of the sentence S. Call this node X and the route taken to find it R.
3. 3) Go through all of the branches above X and to the left of the route R, following a breadth-first left-to-right manner. Propose as an antecedent any node NP encountered that has a node NP or S between it and X.
4. 4) If X is the highest node S in the parsing tree, then go through the parsing trees of preceding sentences in chronological order, with the most recent sentence first, always from left to right and breadth first. When a node NP is encountered, propose it as antecedent. If X is not the highest node S in the sentence, then move on to phase 5.
5. 5) Go from X until the first node NP or S is encountered. Call this new node X and the route to it R.
6. 6) If X is an NP and if R does not cross the node N’ directly dominated by X then propose X as the antecedent.
7. 7) Go through the branches above X to the left of the route R following a breadth-first left-to-right manner. Propose any NP encountered as an antecedent.
8. 8) If X is a node S, then go through all the branches of X to the right of R following a breadth-first, right-to-left manner, but without going above any NP or S encountered. Propose any node NP encountered as an antecedent.
9. 9) Go to step 4.

Although this algorithm is rather basic, a comparison with systems such as those described in [BAL 97, MIT 97, TET 99] show that it is not entirely obsolete. In addition, as noted by Hobbs, its low cost computationally and its simplicity of implementation make it appealing even today, especially in the context of applications that require processing in real time. This explains the adaptation of this algorithm to other languages (see for example [DUT 08]).

3.2.4.2. The discursive approach: centering theory (CT)

The theory of Grosz and Sidner distinguishes between three main components of discourse [GRO 86]: the linguistic structure, the intentional structure, and an attentional state. Regarding the linguistic aspects, discourse is seen as a set of segments and relations relating the pairs of segments. The intentional structure in turn includes the communicative intentions and relations between them. The attentional states concern the focus of the attention of the participants in the discourse at a given moment. Two types of attentional states are distinguished here: local and global. The first concerns relations between utterances in a discourse segment that contribute to the creation of a local coherence (intrasegment). The second concerns the relations between the segments within the discourse (intersegment). Changes in the attentional state depend on both the intentional structure and the linguistic structure.

The foundational ideas of CT were discussed in the works of [GRO 77, SID 79, ARA 81] and concern the semantic links between an utterance and the preceding and subsequent utterances in the context of a discourse segment for local relations. Among others, these ideas pertained to the levels of focus in a discourse, the possible centers of focus in a discourse and using them to choose the referent of an anaphora, and finally the relations between the change of focus and the complexity of inferences necessary for the integration of the current utterance with the rest of the discourse. The working hypothesis is that at any moment, there is a single unit (among the units evoked in the discourse) that takes center stage. This is the centered unit. To clarify this principle, consider passages [3.40] and [3.41]:

1. a) John helped the truck driver Michael to find the street where the factory was. [3.40]
2. b) He looked at a map of the area while Michael drove.
3. c) He finally managed to find the location.

1. a) John helped the truck driver Michael to find the street where the factory was.
2. b) He looked at a map of the area while Michael drove. [3.41]
3. c) Suddenly, he braked when he saw a small cat crossing the street.

In passage [3.40], the pronoun he in utterance c refers to John because it is logical to infer that the person looking at the map will manage to find the location. Similarly, the action of driving the truck normally excludes looking at a map. The anaphor he in passage [3.41] logically corresponds to Michael because the realization of the action of the verb to brake presupposes the action of driving.

According to Hobbs’ syntactic approach, passages [3.40] and [3.41] are perfectly equivalent in terms of coherence. On the other hand, according to CT, passage [3.40] is more coherent than passage [3.41]. Indeed, as Michael is the center of the utterances a and b, the change of the center in passage [3.41] makes inferring the referent of the anaphora more challenging and consequently decreases the coherence of the discourse.

More formally, according to the model proposed by CT, discourse is composed of a set of segments, each of which represents a part of the discourse model. The centers are semantically prominent entities that are part of the discursive model of each utterance E_i in a discourse segment (DS) that consists of a series of utterances: E₁, E₂, …. E_m. Each utterance is associated with two centers: forward-looking centers and backward-looking centers.

Forward-looking centers C_v (E_i, DS) represent the set of entities evoked by the utterance E_i in a DS. These entities can be repeated by subsequent entities in the DS and are ordered according to their prominence. In other words, the more highly ranked an entity is, the more chance it has of being repeated. The order of elements is determined by a number of factors such as the grammatical role of the entity, the order of the entity and the informational novelty that it conveys according to Prince’s criteria [PRI 81]. For example, in passage [3.40]: C_v(E_a)={John, Michael, street, factory}. The highest ranked entity is called the Preferred Center C_p. Thus, the preferred center is John: C_p(E_a) =John. The C_p serves to predict the backward-looking center C_r of the next utterance.

The backward-looking center C_r(E_i,DS) represents the most highly ranked discourse segment for U_i, among those evoked by the previous utterances in the DS. It has the role of connecting an utterance to a previous utterance, which is in turn connected to a previous one and so on. Based on the results of several psycholinguistic studies, CT stipulates that a given utterance U_i cannot have several C_r at one time [HUD 88, GOR 93]. For example, in passage [3.40], the backward-looking center of the utterance a is undefined: Cr(Ea)=[?]⁴, because this utterance is the first in the DS. However, the backward-looking center for utterance b is John: Cr(Eb)= John (the highest ranked entity in utterance a cp(Ea)=John or Cv(Ea)).

Aside from the centers just presented, CT includes the following two rules concerning the center realization and movement: the pronominal rule and the transition specification rule.

The pronominal rule stipulates that pronominalization is a possible way to mark the prominence of an element and that the Cr are often removed or pronominalized. Consequently, if there are a multitude of pronouns in an utterance and if they correspond to several entities in the previous utterance, then the Cr must be one of these pronouns. Naturally, if there is only one pronoun in an utterance, this pronoun must correspond to the Cr. More formally, this rule asserts that if an entity among the ones listed in Cv(Ei-1, SD) is realized in the form of a pronoun in Ei, then Cr(Ei, SD) must also be a pronoun.

The transition specification rule concerns relations between adjacent utterances. It stipulates that the different types of transition require variable quantities of inferences, thus causing variations in levels of coherence. The transitions that require fewer inferences correspond to a more coherent discourse, where there is minimal change to the center. Four types of transitions are described in [BRE 87]. The order of preference of the transitions is: continuation > retention > smooth shift > rough shift (see Table 3.6 for a summary).

Continuation occurs when a locutor indicates that he intends to discuss a particular topic. This implies that the center of an utterance has a good chance of spreading not only to the next utterance but also to the rest of the utterances in the discourse segment. In other words, in the case of a continuation: C_r(E_n+1)=C_r(E_n) or C_r(E_n)=[?] for E_n at the start of a discourse and C_r(E_n+1) = C_p(E_n+1). It is also very likely that C_r(E_n+2)=C_r(E_n).

In the case of retention, the locutor repeats the center of an utterance from one utterance to another but seems to want to change the center in the ensuing utterances. This translates into the fact that the current backward-looking center and the next utterances are the same, but the center is not the most highly ranked entity among the ones listed. More formally, this gives: C_r(E_n+1) = C_r(E_n) or C_r(E_n) = [?] for E_n at the start of the discourse and C_r(E_n) != C_p(E_n+1). This means that it is very likely that C_r(E_n+2) != C_r(E_n).

With a smooth shift, the locutor turns his attention toward a new entity about which he intends to continue speaking. In this case, C_r(E_n+1) != C_r(E_n) and C_r(E_n+1) = C_p(E_n+1).

In the case of a rough shift, like a smooth shift, the locutor turns his attention toward a new entity. However, it is very likely that this new entity will not be the focus of the utterances that follow. Therefore, C_r(E_n+1) != C_r(E_n) and C_r(E_n+1) != C_p(E_n+1).

To illustrate these transitions, examine passage [3.42]:

E1 The genie has lived in the magic lamp for thousands of years. [3.42]

E2 He currently serves Aladdin.

E3 He promised him he would find the lost treasure.

E4 He does not believe him.

An analysis of passage [3.42] utterance by utterance according to CT gives the steps presented in Figure 3.18.

The algorithm in [BRE 87] proposes processing pronominal references according to the principles just described (see Figure 3.19).

A generalization of CT called Veins Theory (VT) was proposed by [CRI 98]. The basic idea of this theory consists of identifying the veins in discursive trees. The role of these veins is to define the domains of the referential accessibility of each unit in the discourse. It should be noted that CT has been applied in the context of automatic scoring systems for written texts based on the distinction between the types of transitions, like in the e-rater system [MIL 00].

From a psycholinguistic point of view, several works have confirmed the cognitive relevance of CT principles. The studies in [CLA 79] found that the reading time for passages where the referent is found in the preceding utterance are shorter than in passages where the referent is found two or three utterances back. Similarly, [CRA 90] found that subjects managed to identify the referent more quickly when it was a grammatical subject than when it was an object.

3.2.4.3. Learning-based approaches

Learning-based approaches use a different philosophy to solve the problem. Rather than manually coding linguistic rules that describe syntactic, semantic and discursive patterns, they annotate representative corpora. The linguistic knowledge is therefore reduced to a minimum and the parameters of the final model are automatically induced. Typically, anaphora and/or their possible referents are annotated with features such as gender, number, grammatical role (like subject or object) or linguistic form (like proper noun or definite or indefinite nominal group). In order to coordinate efforts in the domain of the annotation of corpora intended to train algorithms to process anaphora and coreferences, several standards were created, including MATE⁵ [POE 04, POE 99].

Bayesian classifiers and recurring neuronal networks have been used with the same objective [HUA 13, FRA 08, WIS 16], but the basic classifiers of Decision Trees (DTs) were the most commonly used framework for these tasks [MCC 95, KWA 08, SOO 01, PAU 99] until the recent return to neuronal network-based methods. Simplicity and performance are the reasons cited to justify the use of Decision Trees for classification. The fact that these trees generate rules makes them more appealing for linguistic applications. Historically, several versions of DT algorithms were proposed, like ID3, C4.5 and C5.1.3 [QUI 79, QUI 93, KOH 99]. In the DTs employed for anaphora resolution, the leaves represent the classes and the branches correspond to the conjunction of features that lead to the leaves and therefore the classes. [MCC 95] described a comparison of DTs with a handwritten rule-based system. The work was realized on the English Joint-Venture (EJV) corpus. This is a corpus that is rich in conferences from projects realized by several entities: government, private companies, individuals, etc. The list of features used in the corpus annotation includes (see Figure 3.20 for an overview):

• – NAME i: does the reference i contain a name? The possible values are Yes and No.
• – JV-CHILD i: does the reference i refer to a Joint Venture child? For example, a company or a project resulting from a partnership between two entities. The possible values are Yes, No and Unknown.
• – ALIAS: does one reference contain an alias for another? In other words, is the name of a reference a substring of the name of another? The possible values are Yes and No.
• – BOTH-JV-CHILD: do both references refer to a JV child? Three values are possible: Yes, No, Unknown.
• – COMMON-NP: do the two references share a common Noun Phrase (NP)? These are generally references to a complex nominal group, as in the case of relative propositions and appositions. This feature compares the simple constituent NPs to each reference. The possible values are Yes and No.
• – SAME-Sentence: do the references appear in the same sentence? Yes and No are the possible values for this feature.

The features in the example can have three possible values at maximum: Yes, No, Unknown, while only two classes are possible in this tree: coreference (+) and absence of coreference (-). Note that the interest of decision trees obtained by automatic processing, compared to rules written by linguists, resides in the fact that these rules are automatically induced according to criteria related to the gain of information and are therefore more optimal compared to the content of a corpus used for learning. Moreover, these trees have a much lower production cost. To understand these differences, examine the tree construction algorithm presented in Figure 3.21.

The skeleton of the decision tree to be constructed will have the form given in Figure 3.22.

The choice of the most informative attribute is the most interesting aspect in the DT construction algorithm. To do this, criteria based on information theory are used, starting with the concept of entropy that indicates the degree of impurity or the disorder of the data in the database in the case formalized by the equation [3.43] [SHA 48]. In this equation, we assume that an attribute A takes n values while p(i) is the probability that s belongs to the category i. The attributes will make it possible to share the data in the corpus in subsets according to the values of A:

[3.43]

In this algorithm, what we are looking for is not exactly entropy, but the attribute that reduces this entropy. To do this, we use another measure that depends directly on the entropy, the information gain, whose formula is given in [3.44].

[3.44]

In equation [3.42], v corresponds to the values of A, p(v) is the probability of the value v in the set S and p(v, c) is the probability that an object will be in class C and that it will have the value v. The probabilities p(v) and p(v, c) are statistically calculated according to the occurrences in the corpus.⁶

The comparison of the F-measures (see section 4.1.9.2 for an introduction) by McCarthy between the handwritten system and the computer system shows that the system that uses DTs gives better results than rule-based systems, at 86.5% and 78.9% respectively. Despite their advantages in terms of results, the major inconvenience of learning-based systems is the necessity of having good-quality corpora with appropriate annotation. Unfortunately, such corpora are not always available, especially for languages with few linguistic resources.