Discourse In NLP
In Natural Language Processing (NLP), discourse analysis focuses on coherent sentence or word sequences and aims to comprehend language at a level higher than sentences or clauses. It entails developing ideas and models to explain how utterances combine to create cohesive speech.
Discourse analysis is frequently seen as a component of the pragmatic analysis stage of the NLP pipeline, which normally comes after syntactic and semantic analysis. Another phase in NLP pipelines is Discourse Integration, where a sentence’s meaning is determined by the words that come before it and also affects the meaning of the sentences that come after it. This process checks context for things like pronoun resolution.
Key concepts in discourse analysis include:
- Coherence: The characteristic that sets well-structured discourses apart from haphazard collections of sentences. If a speech exhibits coherence, with meaningful links between its words, it would be considered coherent. Coherence can be global, referring to genre-specific conventional patterns like assertions and supporting premises in essays or plotlines in novels, or local, referring to systematic relationships between adjacent words. The quality of human-generated text and the output of Natural Language Generation (NLG) systems are assessed using coherence. It also facilitates the selection of content for activities like summarization. It may also be possible to recognize signs of a language impairment by detecting incoherent text.
- Cohesion: This is the process of connecting textual units through the use of linguistic mechanisms. Lexical overlap/lexical chains, coreference chains, and cue words/discourse markers are the three primary categories of features for discourse coherence. The link between words in two units, such as the use of synonyms, indicates lexical coherence. One technique to identify structure in a discourse is through lexical repetition.
- Rhetorical relations, also known as discourse relations: These are links between statements or text spans, either explicit or implicit. Prompt words like “because” or “and then” can help to clarify these connections. These relationships are modelled by theories such as the Penn Discourse Treebank (PDTB) and Rhetorical Structure Theory (RST). RST separates major arguments (nuclei) from subsidiary ones (satellites) by employing relations across text spans to organize texts hierarchically into trees. Conversely, PDTB identifies flatter relations between span pairs, which are frequently brought about by discourse connectives (words that indicate relations). Temporal, contingency, comparison, and expansion are some of the categories in the hierarchical sense inventory for these relations, that is part of the PDTB. Comprehending these relationships is essential for tasks such as choosing content for summarization.
- Discourse structure: This has to do with how a conversation is structured. It can be built on relationships between discourse units or hierarchically, as in RST trees. The challenge of mechanically retrieving this structure is called discourse parsing.
- Discourse Units are textual segments that are connected and comprise the discourse structure. The atomic components in RST analysis are called Elementary Discourse Units (EDUs), and they are usually sentences. Relationships between text spans are maintained in PDTB and are frequently determined via connectives. The process of determining the limits of these discourse units is known as segmentation.
- Discourse Model: Similar to humans, natural language processing systems use a discourse model to interpret language. This is thought of as a mental model that is gradually constructed and includes representations of the entities mentioned in the text, as well as information about their attributes and relationships. A referent’s representation is evoked into the model upon initial mention, and it is accessed by subsequent mentions.
- Referring Expressions and Anaphora Resolution (Coreference Resolution): One of the main issues in discourse analysis is this. It entails connecting textual references to the same entity in the discourse model, such as names, pronouns, and definite noun phrases. Two well-known models that deal with reference and anaphora are Discourse Representation Theory (DRT) and Centring Theory. Even though some corpora are particularly labelled for coreference phenomena, work on creating coreference relationships in bigger corpora is still regarded as uncommon. A particularly complicated kind of reference is discourse deixis, in which an anaphor alludes to a speech segment.
- Dialogue Acts Statements in a discussion might be thought of as the speaker’s actions, such as declarations, enquiries, or requests. An essential first step in comprehending conversational discourse is identifying the dialogue activities that underlie utterances. One kind of corpus annotation is dialogue act labelling.
- The use of argumentation mining entails regaining the organization of arguments made in texts, especially when writing persuasively.
The leading theories and methods in discourse analysis are:
- Discourse Representation Theory (DRT): This theory was created to encapsulate the semantics of discourses, enabling quantifiers to handle phenomena like anaphora by utilizing discourse referents and conditions.
- The theory of segmented discourse representation (SDRT) is a DRT expansion that includes a theory of rhetorical relations and discourse structure.
- RST, or rhetorical structure theory, with a fundamental element (nucleus) and a dependent one (satellite) in several relationships, is a theory of text organization that is based on hierarchical relationships (relations) across text spans.
- PDTB, or the Penn Discourse Treebank a sizable corpus and framework that concentrates on discourse relations that are lexically grounded and labelled at the span-pair level, especially those related to discourse connectives.
- The Centering Theory is an entity-based paradigm that emphasizes how monitoring an entity’s prominence (or focus) helps a discourse’s local coherence.
- The entity grid Transitions between entity mentions are used to measure coherence in this formalization of entity-based coherence, which represents entities and their grammatical functions across sentences.
Many NLP applications depend on discourse analysis:
- Extraction of Information (IE) Discourse analysis is required to ascertain the number of events covered in a document and to link the material that is retrieved to the relevant event template. By identifying event-relevant sentences before extraction, discourse-oriented approaches to IE adopt a more global perspective.
- Summarization of Text Discourse structure theories offer techniques for choosing the most crucial information from a document to include in a summary, such as PDTB (using argument importance) or RST (using nuclearity or discourse depth).
- Question Answering (QA): Answering questions can be facilitated by having a thorough understanding of the discourse’s structure and context.
- Report Generation(RG): Choosing the general format and sequence of the content is part of the discourse planning stage of report creation.
- Chatbots and Conversational Systems: These systems focus on providing coherent responses, modelling language structure and conversation state, and identifying dialogue actions to comprehend user intentions.
- MT, or machine translation: Accurate translation requires resolving ambiguities, such as those of word sense and pronoun reference across sentence borders. Another characteristic of a high-quality translation product is discourse coherence.
- Sentiment Analysis: To improve document-level sentiment classification, alterations in subjectivity can be correlated with explicit speech markers.
- Bibliography Annotation Training and assessing NLP systems requires the creation of corpora containing discourse-level linguistic information, such as discourse relations, conversation acts, and annotated coreference relations.
The inherent complexity and ambiguity of natural language, the requirement to integrate multiple levels of linguistic information, the handling of domain-specific language, the computational cost of parsing complex structures, and the difficulty of obtaining large, consistently annotated corpora for many phenomena all pose challenges to discourse analysis, notwithstanding advancements. Classifying and annotating phenomena like implicit discourse relations is still quite challenging.
Several important ideas and occurrences are involved in comprehending and understanding discourse:
Other Phenomena: Understanding the meaning in context and needing practical knowledge are pragmatic features of conversation. Both pragmatic and discourse-interactional meanings can be communicated through prosody (rhythm and intonation). The discourse context can have an impact on word sense disambiguation (“one sense per discourse”). Collocations and vocabulary used in specialised discourse are domain-specific. Persuasive texts’ argumentation structure is often referred to as global discourse structure.
One important topic of NLP is the computational study of conversation processing. Discourse segmentation, coreference resolution, conversation act labelling, and discourse parsing (automatically identifying coherence relations) are among the tasks. Applications for discourse analysis include dialogue systems, information extraction, text summarization, and natural language production. For example, by choosing nuclei over satellites, the nuclearity principle from RST may be used for extractive summarization. Evaluation of produced or human writing is also done through coherence assessment.
Compared to fields like morphology and syntax, our comprehension of discourse and pragmatics is still far less advanced. Although interest in discourse-level phenomena was originally diminished by the shift in NLP towards statistical approaches, it is anticipated that this will change as syntax and semantics advance. Discourse comprehension relies significantly on real-world information, which is hard to fully represent, making it challenging to develop reliable computer models.
You can also read What Are The Basic Linguistic Concepts In NLP? Explained
Discourse relations
A key idea in natural language processing, especially in the fields of pragmatics and discourse processing, is discourse relations.
Discourse relations Definition
Explicit or implicit interactions between textual units, such as sentences, clauses, or phrases, that link them to generate coherent discourse are known as discourse relations (sometimes termed Coherence relations). By creating significant links between statements, they give a text coherence.
Discourse relations are broken down as follows:
- Result, Explanation, Parallel, Elaboration, Occasion, Reason, Contrast, Justification, Condition, Conjunction, Instantiation, Restatement, Alternative, Exception, and List are among the types of discourse connections that are classified according to different taxonomies. Various theories may employ various nomenclature; for example, “rhetorical relation” may refer to a less profound idea.
- Signalling: Discourse connectives or cue words (such as because, and then, nevertheless, moreover, meantime, if…then) can be used to clearly indicate these relationships. They may, however, also be implicit, in which case the relationship is not shown by any particular word or phrase. It is a specific task to distinguish explicit connectives from non-discourse uses.
- Function in Discourse Structure: In order to comprehend language structure beyond the sentence level, discourse interactions are essential. The hierarchical structure created by these relationships can be used to assess the coherence of a whole discourse.
- Discourse structure is represented as a tree by Rhetorical Structure Theory (RST), where relationships exist between textual spans, usually between two nuclei in symmetric relations or between a nucleus (more central) and a satellite (subsidiary). Elementary Discourse Units (EDUs), which might be phrases, clauses, or sentences, are the basic units in RST.
- Another dataset that focusses on discourse relations and is frequently anchored by connectives is the Penn Discourse TreeBank (PDTB). However, it does not assume a complete hierarchical tree structure for the discourse; instead, it identifies relations between pairs of spans.
- Computational Aspects and Significance: One of the main challenges in NLP is comprehending and recognizing discourse relations.
- The task of automatically identifying the coherence relations in a text is known as discourse parsing. This frequently entails categorizing the relationships between units and dividing content into EDUs. Graph-based techniques and transition-based neural models are two approaches. This task is frequently called shallow discourse parsing for PDTB.
- For Natural Language Generation (NLG) to generate text that makes sense, discourse interactions are essential. Determining the logical structure and connections between messages can be a part of strategic generation, also known as text planning.
- For many NLP applications outside of sentence processing, such as summarization, where nuclei or arguments judged more significant might be chosen, identifying discourse relations is crucial.
- Additionally, they can be applied to tasks like sentiment analysis or document classification, for example, by forecasting sentiment shifts based on relations like expansion or contrast or by recognizing subjective segments.
- The quality of generated or human-written texts is assessed using discourse coherence assessment, which can make use of discourse relations.
- In general, it is more difficult to categorize implicit discourse links than explicit ones. To aid in the training of models for implicit relations, researchers have looked into the use of explicit relations.
- It can be difficult to evaluate approaches to rhetorical relations, in part because it might be difficult to achieve consistent human annotation (interannotator agreement).
In conclusion, discourse relations give connected text its framework, enabling us to comprehend the logical and rhetorical relationships between various sections of a discussion or document. One important component of processing language beyond the single-sentence level is the computational resolution and use of these relations.