Questions in the TED-Multilingual Discourse Bank and the development of an annotation scheme

1.1 The role of the DIT++ taxonomy of dialogue acts and the PDTB in the current work

In the DIT++ framework, a dialogue act is a “communicative activity of a dialogue participant, interpreted as having a certain communicative function and semantic content” (Bunt et al. 2017b: 110). The semantic content specifies the objects, propositions, events, and so on that the dialogue act pertains to, and the communicative function shows how an addressee should use the semantic content to update their information state. Communication involves performing several activities in parallel (e.g., pursuing a specific task or activity, providing and eliciting feedback, taking turns), each of which is referred to as a dimension. Two classes of communicative functions are distinguished: general-purpose information transfer and action discussion functions that can apply to any semantic content, and dialogue control functions that apply to acts in a specific dimension, such as Take Turn and Turn Release in the Turn Management dimension (Bunt et al. 2012). The DIT++ taxonomy has evolved into a comprehensive, multidimensional framework for capturing dialogue acts and their functional aspects through a series of releases. It currently includes a set of hierarchically organized general-purpose communicative functions, dimension-specific functions (release 5.1), and a core set of discourse relations (release 5.2). This framework serves as the foundation for ISO standards related to dialogue annotation.

The functions we find most suitable for analyzing questions in TED Talks include (a) general-purpose information-seeking and information-providing functions and (b) dialogue control functions involving dimension-specific allo-feedback and discourse-structuring functions. These functions, along with the PDTB’s attribution annotation, are at the core of the current analysis.

The PDTB framework offers a lexically grounded approach to discourse relations (DRels), which are semantic relations that hold between two adjacent text segments with an “abstract object” interpretation, such as eventualities, propositions, conditions, and facts (Asher 1993), referred to as Argument1 and Argument2. A lexically grounded approach means identifying the DRel as, for instance, a temporal, contingency, or comparison relation based on explicit markers, known as discourse connectives, such as then, because, and nonetheless. Even without an explicit cue, annotators identify the implicit DRel based on the adjacency of discourse segments and insert a connective that best represents the underlying semantic relation. It should be highlighted here that there is an important distinction between the DIT++ taxonomy and the PDTB framework. While the PDTB focuses on identifying the binary arguments of DRels, which serve as the basis for inferring DRels, the DIT++ scheme assigns functions to arguments rather than the relations themselves. This paper excludes DRel annotation from its scope, as any relations between a question and the adjacent text segment have already been annotated in the TED-MDB, and focuses on the functions of questions.

The PDTB annotates attribution by identifying the agent to whom a DRel or its arguments are attributed. In the PDTB framework, “attribution is a relation of ‘ownership’ between abstract objects and individuals or agents” (Prasad et al. 2007: 40). A DRel may hold between the attributions (along with the agents responsible for them) or between the abstract object arguments of the attribution. In (1), for example, the temporal relation instantiated by when holds between the eventuality of “Mr. Green winning the verdict” and “the judge giving him an additional award”, while in (2), the contrast relation conveyed by while holds between the agent arguments of the attribution relation. Thus, attribution is part of the contrast relation. Both examples are from the PDTB. In these examples, italic fonts indicate Argument1 and bold fonts indicate Argument2 of the DRel. Underlined words show discourse connectives.

In this study, we annotate attribution by applying the PDTB 2.0 guidelines to question-answer pairs as well as stand-alone questions in the corpus data.

1.2 Comparison of our approach with existing approaches

Our methodology focuses on explicitly posed questions within the local discourse context. It diverges from existing approaches, such as that of Westera et al. (2020), who annotate “evoked” questions after every sentence based on the TED-MDB within the Question Under Discussion (QUD) theories (Roberts 2012). While Westera et al. (2020) develop a methodology to elicit actual and potential QUDs (from nonexpert annotators), we leave implicit (pragmatically retrieved) questions out of scope.

Our approach also differs from that of Silvano et al. (2022) and Silvano et al. (2025), who annotate TED Talks in multiple languages according to ISO standards. While the annotation categories, such as question type and the communicative function of feedback elicitation, are similar to theirs, our focus differs. For instance, we concentrate on questions but also take into account the subsequent discourse segment to distinguish between stand-alone questions and those followed by an answer (given by the speaker or another agent). In addition, we address the questions’ general-purpose and dimension-specific functions, as outlined in the DIT++ framework.

The rest of the paper proceeds as follows: Section 2 describes the types of questions found within the data, examines their functions according to the DIT++ framework, and introduces how attribution is identified according to the PDTB framework. Section 3 outlines the methodology, while Section 4 presents the data analysis, along with illustrative examples and a brief comparison of English questions with their translations into Portuguese and Turkish in the TED-MDB, to understand whether the approach developed here can reveal similarities or differences in non-English languages. Finally, Section 5 offers the conclusion.

2.1 Questions with and without answers

This section details how we deal with questions with an answer and stand-alone questions in the data. In the rest of the paper, each question type is given initial capitals to denote that it is included in the TAQ-TED as an annotation category. In the examples, bold fonts represent the question, and italics indicate the answer. The file number in the TED-MBD of the TED Talk that each question comes from is shown at the end of each example (in parentheses), together with the DRel label assigned to the question and the adjacent segment; while the category of the question in the TAQ-TED is indicated in an introductory line.

2.2 Functions of questions

As mentioned, the DIT++ scheme identifies two essential communicative functions relevant to the analysis of questions: general-purpose information transfer functions; and dialogue control functions that apply to acts in a specific dimension. In the TED-MDB, we focused on the hypophora type of question-answer pairs and treated the two segments as part of a single communicative intention; here, we observe the functions of the questioning part of Hypophora and Stand-Alone Questions with respect to both functions in the DIT taxonomy.

Let us examine how the general-purpose information transfer functions and the dimension-specific dialogue control functions we are concerned with apply to questions in the data.

2.3 Attribution

We also identify attribution, which can apply to questions as well as answers. Attribution is not a new question type but rather an orthogonal feature that operates over both questions and answers.

Among the key properties of attribution identified by the PDTB framework, we chose to use source and type features. The source feature distinguishes between different types of agents:

1. Writer/Speaker (Speaker in our case)

2. some specific agents introduced in the text (Other agents)

3. some arbitrary individual expressed by a nonspecific reference in the text (Arbitrary) indicated by adverbs like reportedly and allegedly

In addition, we introduce a new source for questions that we observe in the dataset: the Virtual speaker (Grésillon and Lebrave 1984). The Virtual speaker differs from an Arbitrary individual; it represents a non-specified agent’s perspective without explicit clues and potentially involves the addressee’s perspective as well. It is implicit attribution, inferred from the discourse context. It typically applies to the questioning part of Hypophora but can also apply to Stand-Alone Questions.

The type feature describes the relationship between abstract objects and agents, leading to varying inferences about the factuality of the abstract object. The PDTB recognizes four types of abstract objects (assertion propositions, belief propositions, facts, and eventualities) that correspond to attribution relations and the types of verbs that convey them: (a) verbs of communication (e.g., tell, say), (b) propositional attitude verbs (e.g., think, believe), (c) factive predicates (e.g., know, regret) and (d) control verbs (e.g., persuade, order).

The source of questions and answers and the types of attribution relations are captured in the TAQ-TED by annotating verbs of attribution along with their subject – the results are known as attribution spans. As in the PDTB, attribution spans serve as surface cues for the features associated with attribution.

4.1 Results

Table 3 presents the distribution of Q/As in the English data according to their QTYPE and shows that while Choice Questions and Declarative Questions are not found, Stand-Alone Questions and Hypophora are equally prevalent.

The questions observed in the data include 14 Stand-Alone Questions comprised of five Check Questions, five Rhetorical Questions, four Informative Questions. There are also 14 instances of the question element of Hypophora. In addition, the data includes six Nonstandard cases. Among these, three involve questions and answers with attribution, two involve only answers with attribution, and one involves only a question attributed to another agent.

Table 3 highlights a predominance of wh-questions over yes/no questions. According to Levinson (1983: 184), “Yes/no questions will generally have vacuous presuppositions, being the disjunction of their possible answers.” For instance, the question Are investors, particularly institutional investors, engaged? (from TED-MDB Talk no. 1927) presupposes that investors are engaged, or that investors are not engaged. The predominance of wh-questions implies that the TED speakers in the corpus prefer to avoid questions with vacuous presuppositions. This might be related to the Discourse-Structuring and the Attention-Grabbing functions: the speaker starts from a proposition for which a truth-value is provided, and the question will focus on one variable the listener is asked to fill in. For instance, in “why does X?”, X is taken as true, and the listener is asked to reflect on the missing variable (why) and provide an answer (Attention-Grabbing). The listener is no longer passively listening and is actively thinking about the content of the utterance.

Table 4 lists the FUNCTIONS associated with questions, showing that the questions in the data serve more than one function simultaneously, as envisioned by the DIT++ framework. For example, dimension-specific dialogue functions can occur with both Information-Seeking and Information-Providing functions. The five Stand-Alone Questions shown in the first row of the table have Information-Providing and Allo-Feedback Elicitation functions. Check Questions are associated with both Information-Seeking and Attention-Grabbing functions, and the question element of Hypophora is associated with Information-Seeking and Discourse-Structuring functions.

4.2 Examples from the annotated corpus

In this section, we demonstrate the methodology with examples from the annotated corpus. Square brackets denote the functional aspects of the questions annotated in the current work. The table following each example presents the attribution features (Source, Type).

Example (6) illustrates an instance of Hypophora, where the question is annotated with the Discourse-Structuring function. The annotation also captures the fact that the question is attributed to a Virtual speaker.

Example (7) demonstrates a (Positive) Check Question asked by the speaker after recounting a success story.^[3] The Check Question is annotated as an Allo-Feedback Elicitation, and, simultaneously, it is attributed to the Speaker, the default scenario in most of the samples.^[4]

Example (8), yet another Stand-Alone Question, is a Rhetorical Question posed by the speaker to discuss the mastery of craft as a process rather than an end point. The Rhetorical Question states that “many inventors and untold entrepreneurs live out this phenomenon”, and it functions as an Attention-Grabbing element. Moreover, the question is attributed to a Virtual speaker.

We demonstrate how we annotate Nonstandard cases through examples (9) and (10). In the TED-MDB, example (9) is annotated with the implicit DRel Expansion:Conjunction that holds between the eventuality of the professor asking the speaker “if he knew how to solve this problem” (Argument1) and “the speaker saying ‘No, not yet, but I would love to figure it out’” (Argument2). This annotation has been revised by removing the attribution span from the DRel (indicated as boxed). The sources of both the question and the answer (Other and Speaker, respectively) are annotated, along with verbs of communication indicating the types of attribution relations. In addition, the function of the question is annotated (with the Discourse-Structuring function).

(9)

Nonstandard

Example (10) also involves a Nonstandard instance previously annotated in the TED-MDB as an implicit Contingency:Cause:Reason relation, linking the question in Argument1 to the answer in Argument2 along with its attribution. The updated annotation separates the attribution span and includes additional features, specifically, the source of the question and its answer (Virtual and Other, respectively) as well as the values indicating the types of attribution relations (Null and Propositional Attitude verb, respectively). Furthermore, the function of the question (Attention-Grabbing) is annotated.

(10)

Nonstandard

4.3 Crosslinguistic explorations

The annotations of the questions in English were compared with those in the Portuguese and Turkish translations. The comparisons indicated that the translations are similar in many respects: the functions of questions and their distribution across texts are the same; the sources of questions (or answers), types of attribution, and the verbs that convey them show close similarities. Of course, the syntactic properties of the questions may vary across languages due to the specific grammatical rules governing question formation in those languages, such as declarative questions in Portuguese (see (11) and its translation to Portuguese in (12a)). Declarative questions are not observed in the English talks or their Turkish translation in the dataset, although both languages can ask questions in the same way.

However, there are differences between English and translated languages in presenting the source of attribution. For example, Portuguese texts may present the source of the answer differently, as shown in (13) and (14). In this excerpt, the English expression one example is the subject of the text containing the answer, while the source is not mentioned. In Portuguese, the corresponding noun phrase is preceded by a verb in the first person plural form, temos um exemplo ‘we have an example’, as demonstrated in the back translation in (14b). This first person plural verb specifies the source as the speaker and the addressees, in effect reducing the distance between the TED speaker and the addressees.

This observation is supported by an analysis by Celle (2009), who discusses different strategies for expressing questions in academic texts in French and English, using a sample of translated examples. Her findings indicate differences concerning attribution. She shows that in French academic discourse, questions are very frequent and involve the addressee as a fictitious anchor point. In English, the distance between the speaker and the addressee tends to be closer (as seen in the use of we, for instance), and consequently, questions are translated as embedded or supplemented structures with a definite source of attribution.

A similar difference is observed in the Turkish translations of questions. In (15), the question and its answer are direct quotes attributed to others, expressed through communicative verbs meaning ‘ask’ and ‘say’. In the Turkish translation in (16), although the question and the answer are attributed to specific sources as in the original language, the answer is presented differently, by the expression şu cevabı ‘this answer’ meaning ‘the following answer’ (see the back translation (16b)).

Turkish has three determiners and three demonstrative pronouns derived from the determiners – bu ‘this’, şu ‘that’, and o ‘that, yonder’ – referring to objects at varying distances from the speaker. However, a major difference between şu and the others is that while bu and o usually refer to entities mentioned before, şu is almost always used to draw the addressee’s attention to an entity for the first time (Göksel and Kerslake 2004; Küntay and Özyürek 2006). It typically occurs when the speaker and the addressee are in a close, face-to-face relationship. In discourse, it can be used cataphorically to direct the listener’s attention to the forthcoming discourse segment. Therefore, it can be inferred that the choice of the determiner şu in the translation of (15) serves as a tactic to direct the listeners’ focus toward the answer. This usage is addressee-oriented and thus has the potential to reduce the distance between the TED speaker and the audience.

These preliminary observations indicate that TED speakers and translators use tactics to reduce the distance between themselves and the audience. The range of linguistic means by which this is achieved requires further research in multilingual corpora.