“Mama Always Had a Way of Explaining Things So I Could Understand”: A Dialogue Corpus for Learning to Construct Explanations

we present a first corpus for computational research on how to explain in dialogues (Section 3).

Where possible, we followed the literature, but the lack of research on human interaction in explaining (see Section 2) made us extend the state of the art in different respects.

In particular, we focus on turn-level category labels that capture the basic behavior of explainers and explainees in explaining dialogues. Our scheme models the three dimensions of dialogue turns that we agreed on to be needed for a computational understanding of the behavior:

• the relation of a turn’s topic to the main topic,

• the dialogue act performed in the turn, and

• the explanation move made through the turn.

In particular, a turn’s topic may be annotated as follows:

t1 Main topic. The main topic to be explained;

t2 Subtopic. A specific aspect of the main topic;

t3 Related topic. Another topic that is related to the main topic;

t4 No/Other topic. No topic, or another topic that is unrelated to the main topic.

Dialogue Act To model the communicative functions of turns in dialogues, we follow the literature (Bunt et al., 2010), starting from the latest version of the ISO standard taxonomy of dialogue acts.6 In explaining, specific dialogue acts are in the focus, though. In collaboration with the interdisciplinary team, we selected a subset of 10 acts that capture communication on a level of detail that is specific enough to distinguish key differences, but abstract enough to allow finding recurring patterns:

d1 Check question. Asking a check question; d2 What/How question. Asking a what question or a how question of any kind; d3 Other question. Asking any other question; d4 Confirming answer. Answering a question with confirmation; d5 Disconfirming answer. Answering a question with disconfirmation; d6 Other answer. Giving any other answer; d7 Agreeing statement. Conveying agreement on the last utterance of the listener; d8 Disagreeing statement. Conveying disagreement accordingly; d9 Informing statement. Providing information with respect to the topic stated in the turn; d10 Other. Performing any other dialogue act.

Based on a first inspection of a corpus sample, we established a set of 10 explanation moves that a speaker may make in the process, at a granularity similar to the dialogue acts:7

e1 Test understanding. Checking whether the listener understood what was being explained; e2 Test prior knowledge. Checking the listener’s prior knowledge of the turn’s topic; e3 Provide explanation. Explaining any concept or a topic to the listener; e4 Request explanation. Requesting any explanation from the listener; e5 Signal understanding. Informing the listener that their last utterance was understood; e6 Signal non-understanding. Informing the listener that the utterance was not understood; e7 Providing feedback. Responding qualitatively to an utterance by correcting errors or similar; e8 Providing assessment. Assessing the listener by rephrasing their utterance or giving a hint; e9 Providing extra info. Giving additional information to foster a complete understanding; e10 Other. Making any other explanation move.