Discourse-annotated corpora | In the framework of RST, a coherent text can be represented as a discourse tree whose leaves are non-overlapping text spans called elementary discourse units ( EDUs ); these are the minimal text units of discourse trees. |
Discourse-annotated corpora | The example text fragment shown in Figure 1 consists of four EDUs (e1-e4), segmented by square brackets. |
Discourse-annotated corpora | The two EDUs e1 and eg are related by a mononuclear relation ATTRIBUTION, where e1 is the more salient span; the span (e1-e2) and the EDU e3 are related by a multi-nuclear relation SAME-UNIT, where they are equally salient. |
Method | Following the methodology of HILDA, an input text is first segmented into EDUs . |
Method | Then, from the EDUs , a bottom-up approach is applied to build a discourse tree for the full text. |
Method | Initially, a binary Structure classifier evaluates whether a discourse relation is likely to hold between consecutive EDUs . |
Text-level discourse parsing | The two EDUs associated with each sentence are coherent themselves, whereas the combination of the two sentences is not coherent at the sentence boundary. |
Introduction | Rhetorical Structure Theory (RST) (Mann and Thompson, 1988), one of the most influential theories of discourse, represents texts by labeled hierarchical structures, called Discourse Trees (DTs), as exemplified by a sample DT in Figure l. The leaves of a DT correspond to contiguous Elementary Discourse Units ( EDUs ) (six in the example). |
Introduction | Adjacent EDUs are connected by rhetorical relations (e.g., Elaboration, Contrast), forming larger discourse units (represented by internal |
Introduction | Discourse analysis in RST involves two subtasks: discourse segmentation is the task of identifying the EDUs , and discourse parsing is the task of linking the discourse units into a labeled tree. |
Our Discourse Parsing Framework | Given a document with sentences already segmented into EDUs, the discourse parsing problem is determining which discourse units ( EDUs or larger units) to relate (i.e., the structure), and how to relate them (i.e., the labels or the discourse relations) in the resulting DT. |
Our Discourse Parsing Framework | Note that the number of valid trees grows exponentially with the number of EDUs in a document.1 Therefore, an exhaustive search over the valid trees is often unfeasible, even for relatively small documents. |
Our Discourse Parsing Framework | 1For n —|— 1 EDUs , the number of valid discourse trees is actually the Catalan number Cn. |
Parsing Models and Parsing Algorithm | The observed nodes Uj in a sequence represent the discourse units ( EDUs or larger units). |
Related work | Given the EDUs in a doc- |
Introduction | In other words, for two adjacent EDUs not connected by any of the above three relations, the prior probability of staying at the same topic and sentiment level is higher than picking a new topic and sentiment level (i.e. |
Introduction | Drawing model parameters First, at the corpus level, we draw a distribution 95 over four discourse relations: three relations as defined in Table 1 and an additional dummy relation 4 to indicate that there is no relation between two adjacent EDUs (N oRel atz'on). |
Introduction | These parameters encode the intuition that most pairs of EDUs do not exhibit a discourse relation relevant for the task (i.e. |
Bottom-up tree-building | In particular, starting from the constituents on the bottom level ( EDUs for intra-sentential parsing and sentence-level discourse trees for multi-sentential parsing), at each step of the tree-building, we greedily merge a pair of adjacent discourse constituents such that the merged constituent has the highest probability as predicted by our structure model. |
Bottom-up tree-building | ,em}, which are the EDUs of the sentence; after merging el and 62 on the second level, we have E2 = {613,63, . |
Bottom-up tree-building | In contrast, J oty et al.’s computation of intra-sentential sequences depends on the particular pair of constituents: the sequence is composed of the pair in question, with other EDUs in the sentence, even if those EDUs have already been merged. |
Introduction | While research in discourse parsing can be partitioned into several directions according to different theories and frameworks, Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) is probably the most ambitious one, because it aims to identify not only the discourse relations in a small local context, but also the hierarchical tree structure for the full text: from the relations relating the smallest discourse units (called elementary discourse units, EDUs ), to the ones connecting paragraphs. |
Introduction | For example, Figure 1 shows a text fragment consisting of two sentences with four EDUs in total (el-e4). |
Introduction | shown below the text, following the notation convention of RST: the two EDUs el and e2 are related by a mononuclear relation CONSEQUENCE, where e2 is the more salient span (called nucleus, and e1 is called satellite); e3 and e4 are related by another mononuclear relation CIRCUMSTANCE, with e4 as the nucleus; the two spans em and e3;4 are further related by a multi-nuclear relation SEQUENCE, with both spans as the nucleus. |
Overall work flow | Each sentence 5,, after being segmented into EDUs (not shown in the figure), goes through an intra-sentential bottom-up tree-building model Minna, to form a sentence-level discourse tree Tgi, with the EDUs as leaf nodes. |
Related work | In particular, starting from EDUs , at each step of the tree-building, a binary SVM classifier is first applied to determine which pair of adjacent discourse constituents should be merged to form a larger span, and another multi-class SVM classifier is then applied to assign the type of discourse relation that holds between the chosen pair. |
Experiment | We also examined the ROUGE scores of two LEAD4 methods with different textual units: EDUs (LEADEDU) and sentences (LEADSNT). |
Experiment | Many studies that have utilized RST have simply adopted EDUs as textual units (Mann and Thompson, 1988; Daume III and Marcu, 2002; Hirao et al., 2013; Knight and Marcu, 2000). |
Generating summary from nested tree | A document in RST is segmented into EDUs and adjacent EDUs are linked with rhetorical relations to build an RST-Discourse Tree (RST-DT) that has a hierarchical structure of the relations. |
Generating summary from nested tree | RST-DT is a tree whose terminal nodes correspond to EDUs and whose nonterminal nodes indicate the relations. |
Generating summary from nested tree | converted RST-DTs into dependency-based discourse trees (DEP-DTs) whose nodes corresponded to EDUs and whose edges corresponded to the head modifier relationships of EDUs . |
Introduction | Elementary Discourse Units ( EDUs ) in RST are defined as the minimal building blocks of discourse. |
Introduction | EDUs roughly correspond to clauses. |
Introduction | Most methods of summarization based on RST use EDUs as extraction textual units. |
Abstract | In this paper, we present the limitations of constituency based discourse parsing and first propose to use dependency structure to directly represent the relations between elementary discourse units ( EDUs ). |
Discourse Dependency Structure and Tree Bank | Then, discourse dependency structure can be formalized as the labeled directed graph, Where nodes correspond to EDUs and labeled arcs correspond to labeled dependency relations. |
Discourse Dependency Structure and Tree Bank | We assume that the teth T is composed of n+1 EDUs including the artificial e0. |
Discourse Dependency Structure and Tree Bank | Let R={r1,r2, ,rm} denote a finite set of functional relations that hold between two EDUs . |
Introduction | Rhetorical Structure Theory (RST) (Mann and Thompson, 1988), one of the most influential discourse theories, posits a hierarchical generative tree representation, as illustrated in Figure l. The leaves of a tree correspond to contiguous text spans called Elementary Discourse Units ( EDUs )1. |
Introduction | The adjacent EDUs are combined into |
Introduction | We assume EDUs are already known. |
Conclusion | Using the vector-space representation of EDUs , our shift-reduce parsing system substantially outperforms existing systems on nuclearity detection and discourse relation identification. |
Experiments | Distance between EDUs |
Experiments | This suggests that using the projection matrix to model interrelationships between EDUs does not substantially improve performance, and the simpler concatenation construction may be preferred. |
Implementation | These templates are applied to individual EDUs, as well as pairs of EDUs: (1) the two EDUs on top of the stack, and (2) the EDU on top of the stack and the EDU in front of the queue. |
Model | 2After applying a reduce operation, the stack will include a span that contains multiple EDUs . |
Model | where A 6 1Rwa is projects the surface representation v of three EDUs into a latent space of size K < V. |
Model | In this form, we transform the representation of each EDU separately, but do not attempt to represent interrelationships between the EDUs in the latent space. |
Experiments | First, the basic units of their model are elementary discourse units ( EDUs ) from Rhetorical Structure Theory (RST) (Mann and Thompson, 1988). |
Experiments | Second, their model considers the forward relationship between EDUs , whereas ReNew captures both forward and backward relationship between segments. |
Experiments | Third, they use a generative model to capture the transition distributions over EDUs whereas ReNew uses a discriminative model to capture the transition sequences of segments. |
Building a Discourse Parser | In our work, we focused exclusively on the second step of the discourse parsing problem, i.e., constructing the RST tree from a sequence of edus that have been segmented beforehand. |
Building a Discourse Parser | At the core of our system is a set of classifiers, trained through supervised-learning, which, given two consecutive spans (atomic edus or RST sub-trees) in an input document, will score the likelihood of a direct structural relation as well as probabilities for such a relation’s label and nuclearity. |
Building a Discourse Parser | EDUS Syntax Trees I Syntax Parsing (Charniak's n/parse) ' v I Tokenization l I Lexicalization l . |
Features | Therefore, it seems useful to encode different measures of span size and positioning, using either tokens or edus as a distance unit: |
Introduction | Segmentation of the input text into elementary discourse units ( ‘edus’ ). |
Introduction | Generation of the rhetorical structure tree based on ‘rhetorical relations’ (or ‘coherence relations’) as labels of the tree, with the edus constituting its terminal nodes. |
Introduction* | Discourse segmentation is the process of decomposing discourse into elementary discourse units ( EDUs ), which may be simple sentences or clauses in a complex sentence, and from which discourse trees are constructed. |
Principles For Discourse Segmentation | Many of our differences with Carlson and Marcu (2001), who defined EDUs for the RST Discourse Treebank (Carlson et al., 2002), are due to the fact that we adhere closer to the original RST proposals (Mann and Thompson, 1988), which defined as ‘spans’ adjunct clauses, rather than complement (subject and object) clauses. |
Principles For Discourse Segmentation | In particular, we propose that complements of attributive and cognitive verbs (He said (that)..., I think (that)...) are not EDUs . |
CR + LS + DMM + DPM 39.32* +24% 47.86* +20% | To tease apart the relative contribution of discourse features that occur only within a single sentence versus features that span multiple sentences, we examined the performance of the full model when using only intra-sentence features, i.e., SRO features for DMM, and features based on discourse relations where both EDUs appear in the same sentence for DPM, versus the full intersen-tence models. |
Models and Features | Note that our marker arguments are akin to EDUs in RST, but, in this shallow representation, they are simply constructed around discourse markers and bound by an arbitrary sentence range. |
Models and Features | In RST, the text is segmented into a sequence of non-overlapping fragments called elementary discourse units ( EDUs ), and binary discourse relations recursively connect neighboring units. |