A Unique Indexing Technique for Discourse Structures

2.1 Rhetorical Structure Theory

RST is a text representation technique proposed by Bill Mann, Sandy Thompson, and Christian Matthiessen at the University of Southern California as part of their studies on computer-based text generation in 1983 [8]. RST captures the coherence between the text using discourse relations and forms a discourse structure [8]. To relate the coherent texts using discourse relations, elementary discourse units (EDUs) are identified. EDUs are essentially at the clause level, whereas in certain circumstances, they represent sentences. Discourse units identified at the paragraph level are called as complex discourse units (CDUs) [1]. The discourse units are categorized as nucleus and satellite. The nucleus expresses the salient part of the text, whereas the satellite is the additional information supplied about the nucleus. Given a text document, RST builds a graph-like structure, where the nodes are represented by the EDUs or CDUs, and the edges are represented by the discourse relations.

Discourse relations can be categorized into three categories, namely, subject matter, presentational, and multinuclear [7]. Subject matter relations describe the parts of the subject or the main theme of the text. Presentational relations show the presentation aspects with which the author has written the text. Unlike presentational and subject matter relations, multinuclear relations have two nuclei and no satellite. “Elaboration,” “evaluation,” “interpretation,” “means,” “cause,” “result,” “otherwise,” “purpose,” “solutionhood,” “condition,” “unconditional,” and “unless” are subject matter relations. “Antithesis,” “background,” “concession,” “enablement,” “evidence,” “justify,” “motivation,” “preparation,” “restatement,” and “summary” are presentational relations. “Conjunction,” “disjunction,” “contrast,” “joint,” “list,” “multinuclear restatement,” and “sequence” are multinuclear relations.

Figure 1 shows an example of how the nucleus, satellite, and the discourse relation are identified for an English sentence. The triplet nucleus–discourse relation–satellite is denoted as NRS sequence in this article. When a discourse structure is constructed for a paragraph, it can be organized as a rhetorical structure (RS) tree to facilitate the easy access by the NLP applications [8].

Figure 1.

NRS Sequence for Example 1.

Example 1 If you walk daily, you will be healthy.

2.2 Sangatis

Shāstrās, which are expositions of Vedic texts, are organized in the form of sutras (statements), adhikaran¯a (subtopic), pāda (section), adhyāya (chapter), and Sashtra (whole content) [6]. A set of related sutras on a subtopic form adhikaran¯as. A set of adhikaran¯as form a pāda, and a set of pādas form an adhyāya. Sutras, being concise in nature, need to be explained. The explanation is normally organized using adhikaran¯a. An adhikaran¯a has five components, namely, subject of discussion, doubt/ambiguity in understanding the subject, sangati for this discussion, opponent’s view, and the proponent’s (proposed) view. Of these, sangati is used for explanations at various levels: at the sutra level, in terms of how a sutra is related to the previous sutra; at the adhikaran¯a level, as to how a sutra is relevant to the adhikaran¯a; at the pādas level, as to how it is relevant to that pāda. Similarly, sangatis have been used between adhikaran¯as and between pādas as well. Table 1 lists the sangatis that are considered in this article and their equivalent meaning in English.

Figure 2 illustrates the usage of upodghāta sangati in describing a text on cancer, where upodghāta links the introductory part of the text to its respective explanatory part.

Figure 2.

Usage of Upodghāta Sangati.

2.3 Sutras

Sutras have been used in the Tamil language grammar masterpiece Nannool and in the Sanskrit literature. In Nannool, the grammar rules are defined using sutras or noorpaas. As already mentioned, in Sanskrit literature, Shāstrās contain sutra-based texts. Per both the literatures, sutras are equivalent to formulae that explain a wide concept in a brief manner. The definition of the sutra as per Nannool is given in Tamil in Figure 3 followed by the English transliteration and its meaning in English.

Figure 3.

Definition of the Sutra as per Nannool in Tamil.

Per Sanskrit literature, the definition of sutra is transliterated in English as follows:

alpāksaram asandigdham sāravad viśvatomukham astobham anavadyam ca sūtram sūtravido viduh, which means “Of minimal syllabary, unambiguous, pithy, comprehensive, continuous, and without flaw: who knows the sutra knows it to be thus.”

Per Sanskrit literature, sutra is an aphorism (or line, rule, formula) or a collection of aphorisms in the form of a manual or, more broadly, a text in Hinduism or Buddhism. Literally, it means a thread or line that holds things together and is derived from the verbal root siv, meaning, “to sew.”

It can be seen that the sutras have been used by both the literatures to express the content of a coherent text in a concise manner. The first line in the definition of the sutras as per Nanool, “sil vagai ezhuththil palvagaip poruLai” (representing many things in few words), denotes the characteristic of a text index. An efficient text index should be a representative of the sentence, paragraph, document from which it is extracted. Furthermore, they express continuity between the texts they represent. Hence, in this article, this idea of sutras to index the coherent discourse structures built by RST and sangati has been used.

4.1 Pattern Identification

A pattern consists of connected NRS sequences present in a CDU. All such patterns are identified, and each pattern is given a weight depending upon various factors namely, node’s frequency within the pattern, node’s frequency across the patterns in the CDU, pattern’s frequency in the CDU, total number of discourse relations that links the pattern with the other CDUs present throughout the text, and influence of the discourse relations present in the pattern. Using these factors, a significant pattern is chosen as the representative pattern for that CDU.

Node’s frequency within the pattern is given importance because the higher number of nodes implies that the pattern contains more information.

Node’s frequency across the patterns in the CDU gives the cumulative frequency of each node in the pattern across all the patterns identified for a CDU. A high frequency indicates that the node has more semantic connectivity with the other nodes, and hence, it needs to be given importance.

Pattern’s frequency in the CDU gives the frequency of a pattern occurring as a subpattern in the CDU. A pattern occurring frequently as a subpattern in the other patterns indicates that it consists of nodes, which are semantically linked frequently with the other nodes in the CDU, and hence, such pattern needs to be given importance.

Total number of discourse relations that links the pattern with the other CDUs present throughout the text gives the total number links a pattern has with the other CDUs. If a pattern has nodes that have links with the other CDUs, then it indicates that the node is an important sentence of the entire text. When a pattern has many such nodes then the pattern becomes a significant pattern within the text and needs to be given more importance.

Influence of discourse relations indicates the sum of important discourse relations present in each pattern pertaining to the corpus. For instance, in the tourism domain-specific corpus, the discourse relations such as “elaboration,” “list,” “sequence,” and “preparation” predominantly occurs, and such discourse relations need to be given importance. The influence of a discourse relation “Ri” is given by Pi (Ri), where Pi (Ri) is the total number of occurrence of (Ri)/total number of discourse relations present in the corpus.

The pattern that gets the highest weight is chosen to form a sutra. Figure 7 shows the patterns identified for the CDU1, which is present in the discourse structure shown in Figure 6.

Figure 7.

Patterns Identified for CDU1.

Table 2 shows the weights for the patterns shown in Figure 7.

The influence of the discourse relations, “elaboration” and “prāsangika” are 0.6 and 0.33, respectively. It can be seen from Table 2 that the pattern “P1” is the top-weighted pattern and is chosen to form a sutra that represents the CDU1 as the semantic index.

4.2 Sutra Generation

The sutra for a CDU comprises a set of noun concepts and set of discourse relations chosen from the top-weighted pattern. Because nouns convey the essence of a text, the noun concepts are chosen to represent the CDU. Because a sutra needs to be concise, the number of noun concepts chosen to form the sutra is limited by choosing the frequent noun concepts and choosing the abstract noun concepts, leaving behind its instances. The abstract noun concepts can be found using any semantic knowledge base such as WordNet or ontology. The set of discourse relations chosen to form a sutra along with the noun concepts are nothing but the discourse relations that are present within the top-weighted pattern and also that links the top-weighted pattern with the rest of the patterns present in the CDU. Figure 8 shows the sutra generation for a single text document.

Figure 8.

Formation of Sutra for a Single Text Document.

With respect to our example, the noun concepts of the top-weighted pattern P1 are “Tirunelveli city,” “Tirunelveli district,” “India,” “city,” “Thaamirabarani,” and “river bank.” The noun concept, “Tirunelveli city” occurs thrice; thus, it is included in the sutra. The rest of the nouns occur only once and are abstract, and hence, they are included in the sutra. The discourse relations, namely, “elaboration,” “prāsangika,” and “evidence,” which are present in the pattern P1 and links it with the other patterns and other CDUs, are also included in the sutra of CDU1. Hence, the sutra of CDU1 is “Tirunelveli city,” “Tirunelveli district,” “India,” “city,” “Thaamirabarani,” and “river bank” – “elaboration,” “prāsangika,” and “evidence.” It can be seen that the sutra exhibits the complete essence of the CDU1 through the noun concepts and the discourse relations.

Each sutra is stored along with its respective CDU and the CDUs that are semantically linked to the sutra. Then, an inverted index is built for each sutra. Each sutra is tagged with its respective CDU identifiers and the document identifier. On the retrieval side, the discourse relations pertaining to the query are identified by mapping the cue words in the query. Also, the noun concepts in the query are identified, and the combinations of these noun concepts and the discourse relations are matched with the sutras to fetch the results. Figure 9 shows the inverted index data structure of sutras.

Figure 9.

Representation of the Sutra as an Inverted Index.

It can be observed that the proposed indexing technique is suitable for IR, QA, and summary generation systems. For instance, when used by a QA or summary generation system, the CDUs matching the query can be retrieved as an answer or a summary. When used by an IR system, the documents matching the query can be retrieved.

The next section discusses the evaluation of the proposed work and also facilitates the comparison of the proposed work with an existing RST-based indexing technique.

5.1 Evaluation of the Proposed Indexing Technique Using an IR system

The IR system has been tested using 15 queries. Precision for the first ten documents (P@10) and mean average precision (MAP) were used as the parameters for evaluation. MAP for a set of queries is the mean of the average precision scores for each query. A comparison has also been done with an existing RST-based indexing technique proposed by Sahib and Shah [11]. The P@10 values of both the IR systems are shown in Figure 10.

Figure 10.

Performance Comparison Through the IR System.

where N is the number of queries, the MAP score using the proposed approach is 0.7, and MAP score using the approach of Sahib and Shah [11] is 0.62.

The precision values and the MAP score achieved by the proposed indexing technique are higher than that of the existing RST-based indexing approach. The three main reasons behind such high performance are context coverage of the indices, sangati relations, and the presence of more than one discourse relation present in each index. Figure 11 shows a single Tamil query, its English transliteration, and its translation, which was used for evaluation.

Figure 11.

Query1.

The P@10 values obtained through the approach of Sahib and Shah [11] and the proposed approach for the above query are 40% and 70%, respectively. It can be observed from the query that the word sirappu (speciality) can be mapped with the discourse relation “elaboration” and the sangati, visesha. In the approach of Sahib and Shah [11], the documents indexed through the NRS sequence that contains the discourse relation, “elaboration” along with the terms present in the query are retrieved. Because corpus used is tourism domain specific, the queries also lean toward the same domain. It was observed from the corpus that the discourse relation, “elaboration” is predominantly present, and hence, despite the presence or absence of the cue term in the query, the “elaboration” discourse relation is used as a default discourse relation that is used along with the query terms for retrieval. Because the methodology of matching the query terms and the index terms is not stated by Sahib and Shah [11], to retrieve the documents, various combinations of all the query terms have been used along with the discourse relation “elaboration” for retrieving the results. By doing so, only four documents out of first ten retrieved documents were found to be semantically relevant to the query. This is mainly because of the context coverage of the indices in the approach of Sahib and Shah [11]. Because the indices represent a EDU, which is essentially a clause or a sentence, the documents retrieved through such indices may not be always semantically closer to the query. In the proposed approach, an index is a representative of a CDU, which is essentially a paragraph, and hence, the chance of retrieving semantically relevant documents is always higher through such indices. Also, in the proposed indexing technique, the search is done using various combinations of the noun concepts–discourse relations, which help in eliminating irrelevant documents, unlike the approach of Sahib and Shah [11] where all the query terms are used for retrieval. Furthermore, the usage of sangatis in addition to the RST-based discourse relations used for discourse structure construction is also a major reason for high precision. For instance, in the proposed approach, the word sirappu (speciality) present in the query is mapped with the discourse relation, “elaboration” and also with the sangati, visesha. Because sangatis are more capable of capturing the coherence between the complex sentences and paragraphs than RST-based discourse relations [13], it stands as one of the reasons for high-precision values. Using sangati, it can also be observed that the proposed indexer is scalable to handle other types of discourse relations as well. Because an index in the proposed approach is a set of noun concepts and a set of discourse relations, even if the query contains more than one cue term that could be mapped with the discourse relations or sangatis, the chance of retrieving semantically more relevant result is higher in the proposed approach than that of the existing RST-based indexing techniques. For instance, two discourse relations namely, “elaboration” and “joint” can be mapped from the Query2 shown in Figure 12. The “elaboration” is a default discourse relation, whereas the discourse relation “joint” is mapped using the cue term matrum (and).

Figure 12.

Query2.

Because an index in the approach of Sahib and Shah [11] is merely an NRS sequence, the terms and a single discourse relation can only be used for searching the documents. In the proposed approach, the noun concepts along with both the discourse relations are used to search the documents, which results in the retrieval of semantically closer document to the query.