Minecraft Structured Dialogue Corpus

This repo contains the annotated corpus files and the code for the BertLine discourse parser. The data and the model are described in the 2024 LREC paper (forthcoming)

Note: Some minor corrections have been made to the data since LREC publication. The updated stats are summarized below.

Data

Corpus Stats

	Train+Val	Test	Total
No. Dialogues	407	133	540
No. EDUs	17135	5402	22537
No. EDUs (squished)	4687	1473	6160
No. Relation instances	26279	8250	34529
No. Multi-parent DUs	4789	1476	6265

Breakdown by relation type and endpoint type*.

*'Lin' is EDU and 'NL' is nonlinguistic DU, or EEU.

Forward Relations

	Lin-Lin	Lin-NL	NL-Lin	NL-NL	Total
Elaboration	4076	0	0	0	4076
Acknowledgement	1814	0	2736	0	4550
Continuation	2029	0	0	0	2029
Contrast	398	0	0	0	398
Correction	230	0	964	968	2162
Result	1944	6015	2369	0	10418
Comment	1512	0	164	0	1676
Question-answer_pair	1933	0	0	0	1933
Narration	4455	0	0	0	4455
Clarification_question	960	0	0	0	960
Confirmation_question	43	0	956	0	999
Q-Elab	229	0	0	0	229
Explanation	108	0	0	0	108
Alternation	173	0	0	0	173
Conditional	67	0	0	0	67
Sequence	0	38	0	0	38

Backwards Relations

	Lin-Lin	Lin-NL	NL-Lin	NL-NL	Total
Comment	242	0	0	0	242
Conditional	16	0	0	0	16

Builder Action Formats

For the BertLine parser, each Builder action move is compressed to a single code representing place/pick, color and xyz coordinates. The corpus data has with the natural language representation for each move is provided in the data/reformat directory. These files can also be generated from the BertLine data using the action_format.py script also provided in the data directory. This script can be modified if a different linguistic representation of builder moves is desired.

Parser

We use the BertLine parser as originally described in Bennis et al, 2023.

To reproduce our results, run the notebooks in this order:

1. bert_finetune.ipynb
2. bert_linear.ipynb
3. bert_multitaks.ipynb
4. bert_multitask_test.ipynb