The website for Enhanced Rhetorical Structure Theory

Rhetorical Structure Theory

RST is a pragmatic theory of textual organization which aims to explain the structure of documents by imposing a labeled tree structure on natural language texts. In RST, each document consists of a sequence of Elementary Discourse Units (EDUs), which roughly correspond to the propositions (sentences/clauses) in the text. For example, the following fragment from a vlog contains five EDUs, each corresponding to a clause:

[We did n't do a whole lot of hiking here] [because there was not a whole lot of cloud coverage that day ,] [and it was so hot .] [So what we did was we drove around the entire perimeter of the lake] [and you get all these incredible views of the mountains and the lake ,] [and seeing it from all the different angles was so worth it .]

EDUs in an RST analysis are connected to each other via labeled relations which indicate the rhetorical purpose the EDU with respect to another unit, using a small inventory of pragmatically motivated relations, such as Cause (unit A is the cause of unit B) or List (A and B for a list of coordinated propositions). Groups of EDUs which are formed in this way will be connected to other units, until a tree spanning the entire document is formed.

If the two units being connected are of equal prominence, then the relation will be multinuclear, such as the List in the example below. Otherwise, if one unit is more prominent and the other could more easily be omitted, then the less prominent unit will be called a satellite, and its labeled relation will point to the more prominent unit, called the nucleus.

In this example, the speaker first concedes that they did not do much hiking (Concession), additionally specifying a Cause using a List of two EDUs. The causes are less important or prominent than the result (not hiking much), so they form a satellite. The concession itself it also not the main point: it serves as a satellite to the nucleus specifying what the speaker actually did, which is drive around a lake and see some views (the nucleus, itself a List). The second item in the nucleus list also has an Evaluation satellite, giving a positive evaluation that seeing the views was worth it.

What relations?

A range of relation inventories have been used for RST; here we'll be using the inventory from the largest English RST corpus, GUM. Relations are often classified as either presentational, meaning the Writer or speaker (W) intends to influence or persuade the Reader or hearer (R), or subject matter, meaning they represent more objective semantic relations. We also distinguish satellite-nucleaus relations, in which one proposition is more prominent (e.g. Concession), from multinuclear relations, in which participating propositions are equally prominents (e.g. symmetric Contrast).

RST relation labels in the GUM corpus - W is the writer or speaker and R is the reader or hearer
Subject matter(informational)
attribution (pos/neg)S is (or is not) source of information in N[we need to go now!] <- [Kim said]
causeS causes N[I'm full] <- [because I already ate]
conditionS is a condition for N to happened[If you rent a car ] -> [you can drive there]
elab.-additionalS provides more info about N[The hall was full] <- [100 guests were there]
    elab.-attributeS provides more info about phrase in N[The hall] <- [built in 2010] [was…]
evaluationS gives opinion about N (R needn't agree)[Madonna has a new song] <- [I like it]
mannerS gives manner: how N happened[It was shipped] <- [according to EU norms]
meansS indicates means by which N happened[I opened the door] <- [by kicking it with my boot]
purpose-goalN occurs in order for S to happen[I bought it] <- [so that I have a present too]
    purp.-attributeS provides purpose of phrase in N[a plan] <- [to win]
restatement-partialS reiterates part of N (else use multinuc)[It's big and heavy.] <- [Really huge.]
resultS is result of N (inverse of cause)[a bomb destroyed the house] <- [12 were injured]
solutionhoodN is answer to a problem in S[it's broken]->[so use a spare]
Presentational(influence R)
antithesisR finds N more credible than S[They're unemployed,] <- [they're not lazy!]
backgroundR needs to know S to understand N[He was offended] <- [in his culture that's an insult]
circumstanceS gives circumstances (time, place) of N[He got rich] <- [after the recession happened]
concessionW admits S, but still claims N[It's perfect] <- [even though it's scratched]
evidenceS gives evidence that N is true[Madonna's song is great] <- [it's in the top 10]
justifyJustifies why W can say this[Madonna's song is great] <- [the music is amazing]
motivationMotivates R to do something[Madonna's song is great] -> [you should buy it]
org.-preparationS prepares R for N[I'll tell you why:] -> [It hasn't changed since 1990]
org.-headingS is graphically arranged to prepare for N[Introduction] -> [No code is unbreakable. ]
org.-phaticS holds the floor for N, no semantic value[Um, I mean,] -> [so did they buy one?]
questionS requests the information in N[Why did you do it?] -> [I needed the money!]
Multinuclear relations(symmetric, all subject matter)
contrastW presents similar units with contrast[It makes things cheaper] [but it’s harder to do]
disjunctionW presents a set of alternatives[You can go by air] [or you can go by sea]
listW presents coordinate, like units[Last year all summer I read books] [and surfed]
sequenceW presents chronological sequence[Jack joined in 1990.] [Then I joined in 1991.]
otherW presents unlike units with no other rel.[The cliffs are worth seeing.][The beach is a sight too.]
repetitionW presents equivalent/redundant units[It's unbeatable.][You just can't beat it.]
same-unit(Technical device for interrupted EDUs)[Kim,] who …, [was also there]

Enhancements (or: why eRST?)

Although RST gives great insights into the intentional pragmatic structure of a text, it has at least two shortcomings:

  1. The tree constraint means that some relations cannot be expressed (between non-adjacent units, or multiple relations between two units)
  2. There is no indication how we know that a relation applies, or what the components of the relation are.

For example, it seems clear that the word because is an important signal (called a discourse marker or DM) for the Cause relation, and for the Evaluation we may want to know that the positive assessment relates to how there were different views, or that the activity was worth a lot to the speaker. Additionally, we can notice some other discourse markers with no corresponding relations, such as so, which indicates that the driving was also a Result of not hiking, or the and in the last EDT, which indicates that the evaluation is part of the List nucleus to which the first sentence is a Concession, and which is also part of the result of the first sentence.

eRST represents these observations using two enhancements:

  • Secondary edges (marked in blue arrows in the graph below), which are superimposed on the primary RST tree;
  • and signals, which are categorized spans of words specifying what kinds of devices can be used to identify the relations in the graph.

In this analysis, two additional relations are added in blue: Result, indicated by the So highlighted in blue, and List, indicated by the and highlighted in blue. Regular, tree-conforming RST relations, can also be marked by discourse markers, such as the red-highlighted because marking Cause and the two red instances of and marking List relations.

We can also see some non-discourse marker signals: the words different and worth form lexical signals (highlighted in yellow) which are indicative of the Evaluation relation. In total eRST currently recognizes over 40 different types of signals, outlined in the following table (see the Guidelines for more details).

Signal typeSubtypesExample
dmdiscourse markers[because they wanted to]<organization-preparation>
orphansecondary dm[but then they wanted to]<joint-sequence>
graphicalcolon, dash, semicolon[Let me tell you a story :]<organization-preparation>
 items in sequence1. wash [2. cut]<joint-list>
 parentheses, quotation marksit rained [(and snowed a bit)]<elaboration-additional>
 question mark[Did you?]<topic-question> No.
lexicalalternate expressionHe agreed. [That is he said yes]<restatement-repetition>
 indicative word/phraseThey planned a party! [That’s nice/Can’t wait!]<evaluation-comment>
morphologicalmoodGo with them [I think you should]<explanation-motivation>
 tense I started an hour ago, [now I’m resting]<joint-sequence>
numericalsame count[Two reasons.]<organization-preparation> First. . .
referencecomparative[I don’t want it]<adversative-antithesis> I want another one.
 demonstrative / personalThey met Kim. [This person / she was. . . ]<elaboration-additional>
 propositionalThey met Kim. [This encouner was. . . ]<elaboration-additional>
semanticantonymyBeer is cheap, [wine is expensive]<adversative-contrast>
 attribution source[Kim said]<attribution-positive> they would
 lexical chainit was funny [so they laughed]<causal-result>
 meronymyThe house was big, [the door two meters tall]<elaboration-additional>
 negationKim danced, [Yun didn’t dance]<adversative-contrast>
 repetition/synonymyThey met Dr. Kim. [Dr. Kim/The surgeon was. . . ]<elaboration-additional>
syntacticinfinitival/relative clausea plan [to win]<purpose-attribute>
 interrupted matrix clause[I meant –]<orgnization-phatic> I mean,
 modified heada plan [to win]<purpose-attribute>
 nominal modifierarticles [explaining chess]<elaboration-attribute>
 parallel syntactic constructionit’s all tasty [it’s all pretty]<joint-list>
 past/present participial clauseKim appeared [dressed in black]<elaboration-attribute>
 reported speech[Kim said]<attribution-positive> that they would
 subject auxiliary inversionI would have [had I known]<contingency-condition>

More examples

The following examples from the GUM corpus illustrate some more relation and signal types. See the guidelines for more.

In this example, a Question relation connects a question and its answer, and is signaled by the question mark (a graphical signal), the word why (a lexical signal) and the subject-auxiliary inversion for the auxiliary did (marked in cyan, a syntactic signal). DMs mark a multinuclear Contrast (signaled by but) and a secondary temporal Sequence (marked by then) within the Background to the question. An Attribution marks information provided by the semantic source (in gray) Kiara Perkins, the lexical speech verb admitted (in yellow) and the head of a reported speech clause willing (syntactic, in cyan). A referential signal for Background is provided by the pronoun She, which refers back to Kiara Perkins.

This fragment demonstrates a semantic meronymy signal for the Elaboration in unit 17, in addition to the DM and: the key is understood to belong to the glass cabinets in unit 16 (green highlight). Similarly, they referring to the parents is a reference signal indicating that 20 elaborates on 18, as does it, which is a reference to my interest. Units 21–23 are both the Cause for locking instruments, and the nucleus of a Concession regarding the parents' tolerating a hobby, signaled by the DM But (in blue).

This example shows multiple nested Attribution relations: the first is inside a temporal Sequence, indicated by the DM and and the lexical signals, now and again. The first attribution doubly marks its semantic source as I and the lexical attribution verb as say. The nested Attribution marked by the lexical signal glad indicates that the speaker, again I, is the one who is glad, but the word glad is also a lexical signal for the Evaluation relation along the same path.

Guidelines and papers

The eRST theory is discussed and evaluated in detail in the following paper – if you use or want to refer to the formalism in an academic paper, please cite the this article:

Zeldes, Amir, Aoyama, Tatsuya, Liu, Yang Janet, Peng, Siyao, Das, Debopam and Gessler, Luke (2024) "eRST: A Signaled Graph Theory of Discourse Relations and Organization". ArXiv preprint: arXiv:2403.13560.

      title={{eRST}: A Signaled Graph Theory of Discourse Relations and Organization}, 
      author={Amir Zeldes and Tatsuya Aoyama and Yang Janet Liu and Siyao Peng and 
          Debopam Das and Luke Gessler},

The formalism has been applied to the freely available English GUM corpus, a growing multilayer corpus currently covering 16 genres of spoken and written language. The guidelines for the project can be found in the GUM Wiki under the RST and Secondary Relations section.

Browse analyses

You can view analyses of entire documents from the GUM corpus below.

News (23)

Bio (20)

Fiction (19)

Interview (19)

Whow (19)

Academic (18)

Reddit (18)

Voyage (18)

Speech (15)

Textbook (15)

Vlog (15)

Conversation (14)

Court (6)

Letter (6)

Essay (5)

Podcast (5)