Abstract
This is an extended abstract of our LFG Winter school talk. We discuss Information Structure, the sentence internal partition of the information according to discourse functions, and Discourse Structure, the inter-clausal relations between successive utterances. We first present the basic issues, then how these have been addressed in LFG, and finally how they relate to applications.
The original slides used for this presentation are available from the LFG 2004 Winter school page as a ppt file.
Information Structure involves the syntax of discourse functions (DFs) and their link to pragmatics.
DFs encode and divide up the information structure of the sentence. Unfortunately, they are notoriously difficult to define. Notions that commonly appear in the linguistic literature include:
Some of these divisions classify all parts of the sentence as having a particular DF (e.g., the theme-rheme division is often used in this way). However, for many divisions, especially those most often used in formal linguistic analyses, much of the sentence does not have a DF. The question is then whether and how to represent non-DF information, e.g. background information.
Clefts provide a classic example for the syntactic encoding of discourse functions. For a sentence like:
It is [the box]Focus [that]Topic I opened.
the cleft construction indicates that the focus is the clefted constituent. However, the referent of that focused constituent is also the topic of the subordinate, clefted clause, whereas the material in the relative clause is presupposed.
Determining the DFs of the elements of a sentence can be difficult. In the linguistic literature, question-answer pairs are often used to determine DFs in that the part of the sentence that corresponds to the answer is focused. As seen in the example below, the focused information in a cleft can provide the answer to a question.
What did you open? The trunk?
It was [the box]Focus that I opened.
In this talk, we focus on how to encode DFs and what they can be used for, setting aside the issue of how to identify which elements of the clause have which DF. In our opinion, the choice of relevant DFs depends on what they are used for, i.e., some divisions may be needed for some phenomena and applications, while other divisions are needed for other ones.
DFs can be encoded in a number of ways. One is via a privileged structural position, often either clause initial or preverbal. A second is to use discourse markers or particles to mark the constituent with the DF. A third is to use a special intonation pattern. Often a combination of these methods is used, e.g., focus may appear in preverbal position with a specific intonation pattern.
In the structural encoding of DFs, a particular (phrase structure) position licenses a particular DF. Which positions encode which DFs is language specific. However, there are some general tendencies: topics are often initial; focus are often pre- or post-verbal; background information is often postverbal. In addition, languages may have particular constructions to encode DFs, such as English clefts. Finally, in many languages, the subject is a type of default topic. In LFG, the structurally encoded DFs are analyzed as occurring in designated c-structure positions whose functional annotations assign or license the DF in question. We first provide some examples of the structural encoding of DFs; in the next section, we sketch some proposed LFG analyses of this encoding.
Chichewa provides an example of initial topics (Bresnan & Mchombo 1987). In Chichewa, when there is an object marker on the verb, this object marker is the object (e.g., it has its own PRED that fulfills the OBJ role of the verb). The object marker can be anaphorically linked to an initial noun phrase that is interpreted as the topic.
[Alenje]Topic zi-ná-wá-lu-ma njuchi. hunters SM-past-OM-bite-indic bees `The bees bit them, the hunters.' (Chichewa)
Preverbal focus is common cross-linguistically, especially in verb-final languages. For example, Turkish is a verb-final language with preverbal foci (see inter alia, Enc 1991).
bu kitab-i Hasan [ban-a]Focus ver-dir this book-acc Hasan I-dat give `This book Hasan gave to ME.' (Turkish)
Some languages use special morphemes to mark DFs. One of the best-known such markers is the Japanese topic marker wa, although its exact function remains controversial. Here we show an example from Hindi which has several such discourse markers (Sharma 2003):
[rAdha=ne=hI]Focus baccho=kO kahAnI sunAyI Radha=erg=Foc children=ACC story hear `It was (only) Radha who told the children a story' (Hindi)
[mOmbattI=tO]Topic milI, kEkin abh mAchis gum gayE candle=Top found but now match lost go `The candle was found but now the matches are lost.' (Hindi)
Did you see Mary or John?
I saw [JOHN]Focus.
I noticed that you wore a blue hat to the party.
It was a [RED]Focus hat that I wore.
In many cases, more than one technique is used to mark DFs in a language and these are often used at the same time, i.e., there is multiple encoding of the DF element. In particular, most positionally and marker-signaled DFs also have intonation marking. It is also possible to combine position and marker, as is the case with ay inversion in Tagalog (Kroeger 1993). In ay inversion, ay is a DF marker which is the head of I; the constituent in SpecIP is a topic if it is the subject and a focus if it is a non-subject.
[Ni lapis ay]Focus hindi nagdala si=Rosa even pencil AY not bring nom=Rosa `Even a pencil Rosa didn't bring.' (Tagalog)
We now turn to the issue of how LFG analyzes DFs. This is an issue that has been explored in some detail but for which there is still much work to be done. We consider three issues here: the encoding of linguistic-DF interactions, whether to encode DFs at F-structure or I-structure, and OT-LFG approaches.
One interesting possibility is that predicates can subcategorize for DFs. Another more common approach is that DFs are not subcategorized but are assigned via functional annotations on special c-structure elements. These c-structure nodes are associated with DFs in a way similar to subcategorized grammatical function assignment in configurational languages. The DF units are then linked to subcategorized grammatical functions via anaphoric control.
Subcategorized DFs It has been proposed that topic in Malay is a subcategorized DF (Alsagoff 1992). In particular, the verbal affix identifies the topic and equates it with a grammatical function. There are three such affixes in Malay (one is unpronounced).
| (i) | (^ TOP)=(^ SUBJ) | |
| (ii) | < (^ SUBJ) | (^ OBL) > |
| log obj | log subj | |
A sentence like the one below with the meng- prefix on the verb will have a verbal predicate which subcategorizes for a topic, resulting in the f-structure below.
[Miriam]Topic MENG-cubit doktor itu Miriam MENG-pinch doctor the `Miriam pinched the doctor.' (Malay)
MENG-cubit (^PRED)='pinch< (^ SUBJ), (^ OBJ)> (^ TOP)'
| [ PRED | 'pinch<(^ SUBJ), (^ OBJ)> (^ TOP)' |
| SUBJ | [ PRED 'Miriam' ]1 |
| TOP | [ ]1 |
| OBJ | [ PRED 'doctor' ] ] |
Topic Assignment In both Chichewa (Bresnan & Mchombo 1987) and Tagalog (Kroeger 1993), topics have a designated c-structure position associated with them, although the exact c-structure position associated with topic varies. In Chichewa, the topic assigning position is the position before the VP:
S
/ \
NP VP
(^ TOPIC)=!
while in Tagalog the topic is assigned in SpecCP:
CP
/ \
NP C'
(^ TOPIC)=!
In both languages, the resulting f-structure is similar, with some grammatical function being anaphorically bound to the topic DF.
| [ TOPIC | [ ... ]1 |
| SUBJ | [ PRED 'pro' ]1 |
| PRED | 'X<SUBJ,...>' ] |
Preverbal Focus Position The LFG analysis of Urdu (Butt and King 1996) assigns focus in SpecVP. Unlike in the Chichewa case, the focus in Urdu also has a grammatical function assigned to it, e.g., it may be the subject or object of the verb. Languages differ as to whether these DF positions are also associated with grammatical functions.
VP
/ \
XP V'
(^ FOCUS)=!
(^ GF)=!
Mapping Constraints The question then arises as to which c-structure positions can be used as DF licensing positions. Bresnan 2001 provides a proposal as to how these c- to f-structure mappings can be constrained. One way is a restriction on the clause prominence of DFs (Bresnan 2001:192) which states that DF adjuncts (i.e., in adjoined positions) must be clause-prominent, occurring either at an edge of the clause or adjacent to the head of the clause. This is shown graphically below. If there are two XP adjoined positions, one with a YP in it and one with a ZP, then the outermost one (the YP in this example) must be the DF, while the innermost one (the ZP in this example) must be the Adjunct.
XP
/ \
YP XP
DF / \
ZP XP
Adjunct
Another restriction (Bresnan 2001:102) is that specifiers of functional categories are the grammatical DFs: Topic, Focus, and Subject. (Subject is often the default topic in languages.) That is, the specifiers of these positions are reserved for DFs, and hence the distribution of functional categories in the language will in part dictate the possible DF positions.
FP
/ \
SpecFP F'
DF
There has been much work done on the association of intonation to DFs (see especially Steedman (2000) on Categorial Grammar). However, there is relatively little work in LFG on how this encoding can be analyzed. Two LFG proposals for how intonation information can be incorporated into LFG analyses are Butt and King's 1998 analysis of Bengali and the syntax-prosody mapping and King's 1995 analysis of Russian clause-final focus, but the integration of prosody into the LFG projection architecture needs much more exploration. The work of O'Connor (2004) seeks to address this need.
Using discourse markers to encode DFs has been analyzed as involving constructive case/morphology (Sharma 2003). Under this analysis, the marker itself contains information about the DF that it is associated with. For the Hindi focus marker hI this would involve the inside out equation (FOCUS ^) in its lexical entry.
X(P)
/ \
X(P) Cl-disc
(FOCUS ^)
hI
One interesting question that arises given LFG's projection architecture is whether DF information should be represented at f-stucture or at I(nformation)-structure. Traditionally, DFs are often represented in the f-structure. This is necessary for languages like Malay that subcategorize for DFs, and for phenomena like Chichewa incorporated pronouns.
However, the scope of DFs may conflict with that of grammatical functions, and one solution to this problem is to project DFs into an I-structure. An example of such a mismatch can be seen in VP focus for English in which the focused elements do not form an f-structure constituent. The way around this is to have the I-structure constituents not be an exact match of the f-structure ones (King 1997).
Mary [ate the cake]Focus.
| [ PRED | 'eat<SUBJ,OBJ>' |
| SUBJ | [ PRED 'Mary' ] |
| OBJ | [ PRED 'cake' ] |
| TNS | past ] |
As seen above, there is no unit in the f-structure which contains the PRED of the verb eat and the object. In fact, any f-structure unit that contains the verb will contain the entire f-structure since the verb heads the f-structure. By putting the DF focus information in a separate structure, this problem can be solved, although more work on the exact representation of I-structure remains to be done.
| [ FOCUS | { eat |
| cake } ] |
OT-LFG provides a cross-linguistic way of analyzing the encoding of DFs. Under these approaches, there are Optimality Theory constraints for encoding DFs (Choi 1999). These constraints make statements about the encodings of DF information, e.g., new information (types of focus) must be in salient positions (e.g., specifiers of functional categories). Different languages rank these constraints differently and define possible instantiations of X in order to capture the cross-linguistic variation in encoding DFs.
In sum, DFs can be encoded by: structural position, morphological markers, intonation, or a combination thereof. Much work in LFG has been done on structural position and morphological markers, but much work remains to be done. Are focus and topic the only elements worth distinguishing? How should intonation be integrated into the LFG architecture? How do syntactic DF distinctions relate to semantic and pragmatic concepts?
Briefly consider the form and function relation. One radical proposal by Prince is that the relation between syntax and pragmatics is as arbitrary as that between sound and word meaning. Looking at cross language variation, for example, the functions of left-dislocation in Yiddish and English are different (Prince 1998); the functions of clefting and topicalization are different across Germanic languages; and the functions of left-dislocations (or Contrastive topicalization) and right dislocations in Romance languages and in Germanic are different (see e.g. Lambrecht 1981 on Spoken French). This shows that there is no universal one-to-one correspondence between form and function. However, this does not necessarily mean that the relation between form and function is arbitrary. OT approaches are one way of capturing general trends without forcing complete uniformity.
We now consider how DFs can be used in an application: anaphora resolution. The problem of anaphora resolution is that in a sentence with pronouns (him) or referring NPs (the president) one wants to know what they refer to. Some restrictions are purely syntactic, e.g., (most) reflexives refer to subjects. However, others are heuristic, e.g. to prefer closer referents and to prefer high saliency referents. DFs are relevant to anaphora resolution because the choiceof a particular DF indicates the saliency of a referent and salient elements are more often referred back to.
For example, topic, and topic shift, are relevant for anaphora resolution. This has been discussed in detail in Centering theory and its variants. Centering theory establishes an ordered list of salient elements for each sentence n. The claim is that if in sentence n+1, one or more elements of this list of salient elements are realized and there is at least one pronoun in sentence n+1, then the highest realized element of the salient element list must (also) be a pronoun in sentence n+1. Consider the example discourse below in which the DFs determine correct anaphora resolution for the pronouns she and her.
Brennan drives an Alpha Romeo. Brennan =Old, Alpha Romeo=New She drives too fast. She=Brennan=Old Friedman races her on weekends. Friedman=Old, Brennan=Old, Her=Brennan=Old She drives to Laguna Seca. She=Friedman=Old She often beats her. She=Friedman=Old Her=Brennan=Old
Another problem that anaphora resolution has to address is that of null pronominals, especially those found in pro-drop languages. Pro-drop is (partly) licensed by DFs in that already established topics are more likely to be pro-dropped. Centering theory provides an analysis for this in which Continue and Smooth-shift transition favor null subjects; this has been argued for Chinese (Song 2003) and Yiddish (Prince 1998).
DFs are essential for determining anaphora resolution, including pro-drop which is licensed in part by DFs. However, a lot remains to be worked out and the detailed formal framework provided by LFG should provide a solid platform from which to explore this.
In contrast to Information Structure which encodes discourse functions (note the confusing terminology), Discourse Structure encodes the relations between sentences. Here we discuss a relatively simple model and its relation to syntax. First consider an overly simple idea in which the Discourse Structure consists simply of the concatenation of the sentences of a discourse in the order in which they occur.
D
/ / | \ \
/ / | \ \
S S S S S
Such a simplistic model will not work because it does not encode the interaction of progression and elaboration as shown in the contrasting examples below.
Joan got up early. She showered. Then she made some tea. ...
Mary is a model professor. Last year she wrote ten papers. She also advised 20 doctoral students and she was a member of the Committee on Women in Science.
A slightly more elaborate, but still very simple idea which can capture elaboration and progression is that either the discourse progresses sentence by sentence or subparts elaborate on previous parts.
D
/ / | \ \
/ / | \ \
S S D S S
/ \
S S
We consider one type of discourse trees (Linguistic Discourse Model) in which the node labels indicate whether the events in the sentence are elaborations/subordinations (S) or progressions/coordinations (C). In the examples below a and b are BDUs (Basic Discourse Unit) where a BDU basically corresponds to a segment with an event variable in its semantics.
John fell. Bill pushed him.
S
/ \
a b
Bill pushed John. He fell.
C
/ \
a b
Not all types of relations can be classified as belonging to the subordinating or the coordinating type, but we will ignore the rest here. Some elements in a sentence can explicitly indicate what type of relation we have, e.g. because is a subordination relation. These elements will be called "operator segments." When building the Discourse Structure, we must ask how discourse trees relate to sentence syntax trees. Some textual elements guide the discourse tree construction. Importantly, a BDU is not necessarily a complete sentence or vice versa.
[The man dove into the pool.]a [It was warm and soothing]b and [he decided to remain for a little longer than usual.]c
C
/ \
S c
/ \
a b
[Joan left]a because [she was tired.]b
=Three segments: Two BDUs and 1 operator
S
/ \
a b
Textual elements can guide the construction of discourse trees. A number of hypotheses can be proposed as to how this is done. Hypothesis 1: Subordinating conjunctions indicate discourse subordination. This needs checking because although it is often true, it is not clear that it is always so. Hypothesis 2: Tense and aspect are important. For example, stative predicates do not push the discourse forward and often indicate subordination. In general, English is not very rich in this type of indicator, but perfective/imperfective distinctions are more explicit in other languages such as Romance and Slavic (e.g. Asher and Lascarides, 2003). Hypothesis 3: Pronominalization is important for building Discourse Structure. Often the "promotion" of (the referent of) an OBJ or a OBL to a SUBJ in the following sentence reflects a discourse subordination (Polanyi et al. 2004).
John Smith was wearing a long coat. It looked brand new. (elaboration)
Note however that the tense and aspect information takes precedence over that of pronominalization.
John hit Bill. He fell. (progression)
Given that syntactic information is relevant to the construction of discourse trees in Discourse Structure, we can then ask what the role of Information Structure is in the construction of discourse trees. That is, how do DFs like topic and focus affect the construction of the Discourse Structure, if at all? In the first of the following examples, the focus in the first sentence is the topic of the following one, but in the second example, the the topic of the first sentence is the topic of the second one (cf. centering theory "shifts"). However, in Discourse Structure both are subordinations.
[John Smith]T1 was wearing [a long coat]F1. [It]T2 looked brand new.
Focus-1 becomes Topic-2
[John]T1 likes [sweets]F1. [He]T2 eats [three dishes of ice cream]F2 and [five chocolate bars]F2 every day.
Topic-1 becomes Topic-2
Looking at the above data, one might then ask whether Information Structure and Discourse Structure are independent. This is not an impossible view: Information Structure is what the sentence/discourse is about, while Discourse Structure is how we talk about what we are talking about, e.g., narratives, explanations. However, there are clearly some links.
Consider the following quote from King 2003:140-141:
'The first Christian mission to New Zealand,..., was launched by Samuel Marsden on behalf of the Church of England's Church Missionary Society (CMS) in 1814... Marsden, a bluff Yorshireman with 'heavy shoulders and a face of a petulant ox', was both chaplain to the New South Wales penal settlement and a magistrate. He was severe in dealing with convicts... But he went out of his way to meet and greet Maori in Sidney, and often... He had even, in 1809, rescued the Maori sailor Ruatara, who was stranded in London, and taken him back with him to Sidney. It was this association in particular that led Marsden to set up the first CMS mission at Rangihoua in the Bay of Islands in 1814, on land that he would buy from Ruatara.'
In this narrative, the cleft, an Information Structure phenomenon, seems to indicate a "pop" from the subordinated material to the resumption of main narrative. Note also that the material in the clefted that-clause might be presupposed in the logical sense, but it is not old information (see Collins 1991 for ample examples). (Note that we do not claim that this is the only Discourse Structure function of it-clefts.)
Discourse Structure looks at how clauses and sentences are related to one another. Textual elements provide information on how to build up the structure, but they do not completely determine it.
Discourse Structure can be used in applications such as summarization and sentence condensation. In sentence condensation and summarization, a long text is "condensed" into a shorter version. To do this, one wants to retain the most salient features and maintain grammaticality. This can be done in part by choosing salient sentences via Discourse Structure and then condensing those sentences, partially based on Information Structure.
Consider the following example of Discourse-driven summarization. We have the text:
Our group is developing new techniques for helping manage information for enhanced collaboration. We explore solutions for seamlessly connecting people to their personal and shared resources. Our solutions include services for contextual and proactive information access, personalized and collaborative office applications, collaborative annotation and symbolic, statistical and hybrid processing of natural language. Our team includes researchers with diverse backgrounds including: ubiquitous computing, computer-supported collaboration, HCI, IR, and NLP.
The discourse tree is a coordination of two subtrees that have subordinated elements with again coordinated or subordinated elements.
C
/ \
S S
/ \ / \
S S e S
/ \ / \ / \
a b c d f g
[Our group is developing new techniques]a for [helping manage information for enhanced collaboration.]b [We explore solutions for seamlessly connecting people to their personal and shared resources.]c [Our solutions include services for contextual and proactive information access, personalized and collaborative office applications, collaborative annotation and symbolic, statistical and hybrid processing of natural language.]d [Our team includes researchers with diverse backgrounds]e [including:]f [ubiquitous computing, computer-supported collaboration, HCI, IR, and NLP.]g
Possible discourse based summarizations of this text include the following; note that the last one has been summarized to the point of being relatively uninformative.
Discourse Structure allows only "big chunks" to be deleted. However, we also need finer-grained structure for sentence condensation in which parts of a given sentence are deleted. In this way, Information Structure is relevant for sentence condensation. The idea is that salient information will be discourse prominent, implying that one wants to retain foci and topics, possibly in reduced form (e.g. as a pronoun), while deleting non-prominent background information. Salient information tends to correspond to the heads of arguments in main clauses (but there are a lot of special cases, e.g., adjuncts can be deleted but one should keep negative adjuncts).
[Our group is developing new techniques]a for [helping manage information for enhanced collaboration.]b [We explore solutions for seamlessly connecting people to their personal and shared resources.]c
Crouch et al. 2004 propose an LFG based system in which rewrite rules allow one to stipulate which part of an f-structure can be deleted in a well-formed sentence condensation. For example, the adjunct deletion rule might look like:
X $ (Y ADJUNCT) & (X ADJ-TYPE) =/ neg:
X ==> 0.
The basic sentence condensation system described by Crouch et al. works as follows. The sentence is parsed, producing an f-structure, rewrite rules manipulate this f-structure, removing unnecessary parts or simplifying the structure, e.g. via declefting or passivization. This new f-structure is then used to generate the condensed string.
In sum, Discourse Structure can guide summarization. LFG f-structures are easily manipulated for condensation and the distinctions in f-structures give broad guidance (e.g., adjuncts and modifiers vs. grammatical functions and DFs). However, there are distinctions that are important for condensation and other meaning related manipulations but that are very minor in the f-structure, e.g. difference between negative and other adjuncts.
In conclusion, Information Structure involves the sentence internal partition of the information according to discourse functions. In contrast, Discourse Structure involves the inter-clausal relations between successive utterances. Both of these are crucial in certain applications, including anaphora resolution and summarization/condensation. For such applications, it is necessary to get all the modules worked out, both with their content and how they interact. This crucially involves many aspects of linguistic theory, and we think that the projection architecture of LFG should be helpful in providing such analysis. However, a lot of theoretical and implementational work remains to be done in this area.