wylupek/mrda-corpus

MRDA Corpus - Meeting Recorder Dialogue Act Dataset

⚠️ This is a reformatted version of the original ICSI MRDA corpus for easy use with HuggingFace Datasets. All credit goes to the original authors.

Original Work

This dataset is based on the Meeting Recorder Dialogue Act (MRDA) Corpus by Shriberg et al. (2004). The original corpus consists of approximately 75 hours of naturally occurring multi-party meetings transcribed and annotated for dialogue acts.

Original Sources:

• ICSI Meeting Recorder Dialogue Act Corpus
• Nathan Duran's MRDA-Corpus Processing Tools

Dataset Splits

Split	Examples
Train	75,067
Test	16,702
Validation	16,433

Data Format

Each line contains: speaker|utterance_text|basic_label|general_label|full_label

Example:

fe016|okay.|F|fg|fg
fe016|so um|F|fh|fh
fe016|i was going to try to get out of here like in half an hour.|S|s|rt

Labels Description

The MRDA corpus uses a hierarchical dialogue act annotation scheme with three levels of granularity:

Basic Labels (5 labels)

Dialogue Act	Labels	Count	%	Train Count	Train %	Test Count	Test %	Val Count	Val %
Statement	S	64233	59.36	45099	60.08	9571	57.30	9563	58.19
BackChannel	B	14620	13.51	10265	13.67	2152	12.88	2203	13.41
Disruption	D	14548	13.45	9739	12.97	2339	14.00	2470	15.03
FloorGrabber	F	7818	7.23	5324	7.09	1409	8.44	1085	6.60
Question	Q	6983	6.45	4640	6.18	1231	7.37	1112	6.77

General Labels (12 labels)

Dialogue Act	Labels	Count	%	Train Count	Train %	Test Count	Test %	Val Count	Val %
Statement	s	69873	64.58	48952	65.21	10472	62.70	10449	63.59
Continuer	b	15167	14.02	10606	14.13	2219	13.29	2342	14.25
Floor Holder	fh	8362	7.73	5617	7.48	1520	9.10	1225	7.45
Yes-No-question	qy	4986	4.61	3310	4.41	870	5.21	806	4.90
Interrupted/Abandoned/Uninterpretable	%	3103	2.87	2171	2.89	492	2.95	440	2.68
Floor Grabber	fg	3092	2.86	2076	2.77	489	2.93	527	3.21
Wh-Question	qw	1707	1.58	1110	1.48	310	1.86	287	1.75
Hold Before Answer/Agreement	h	792	0.73	474	0.63	134	0.80	184	1.12
Or-Clause	qrr	392	0.36	244	0.33	75	0.45	73	0.44
Rhetorical Question	qh	352	0.33	260	0.35	56	0.34	36	0.22
Or Question	qr	207	0.19	131	0.17	37	0.22	39	0.24
Open-ended Question	qo	169	0.16	116	0.15	28	0.17	25	0.15

Full Labels (52 labels)

Dialogue Act	Labels	Count	%	Train Count	Train %	Test Count	Test %	Val Count	Val %
Statement	s	33472	30.93	23238	30.96	4971	29.76	5263	32.03
Continuer	b	15013	13.87	10517	14.01	2175	13.02	2321	14.12
Floor Holder	fh	8362	7.73	5617	7.48	1520	9.10	1225	7.45
Acknowledge-answer	bk	7177	6.63	5117	6.82	1031	6.17	1029	6.26
Accept	aa	5898	5.45	4097	5.46	903	5.41	898	5.46
Defending/Explanation	df	3724	3.44	2790	3.72	531	3.18	403	2.45
Expansions of y/n Answers	e	3200	2.96	2360	3.14	540	3.23	300	1.83
Interrupted/Abandoned/Uninterpretable	%	3103	2.87	2171	2.89	492	2.95	440	2.68
Rising Tone	rt	3101	2.87	2015	2.68	516	3.09	570	3.47
Floor Grabber	fg	3092	2.86	2076	2.77	489	2.93	527	3.21
Offer	cs	2662	2.46	1878	2.50	402	2.41	382	2.32
Assessment/Appreciation	ba	2216	2.05	1605	2.14	354	2.12	257	1.56
Understanding Check	bu	2091	1.93	1405	1.87	371	2.22	315	1.92
Declarative-Question	d	1805	1.67	1153	1.54	350	2.10	302	1.84
Affirmative Non-yes Answers	na	1112	1.03	870	1.16	133	0.80	109	0.66
Wh-Question	qw	951	0.88	630	0.84	160	0.96	161	0.98
Reject	ar	908	0.84	594	0.79	152	0.91	162	0.99
Collaborative Completion	2	841	0.78	571	0.76	136	0.81	134	0.82
Other Answers	no	828	0.77	563	0.75	98	0.59	167	1.02
Hold Before Answer/Agreement	h	792	0.73	474	0.63	134	0.80	184	1.12
Action-directive	co	674	0.62	460	0.61	97	0.58	117	0.71
Yes-No-question	qy	669	0.62	476	0.63	90	0.54	103	0.63
Dispreferred Answers	nd	483	0.45	341	0.45	82	0.49	60	0.37
Humorous Material	j	463	0.43	326	0.43	67	0.40	70	0.43
Downplayer	bd	387	0.36	290	0.39	68	0.41	29	0.18
Commit	cc	371	0.34	258	0.34	51	0.31	62	0.38
Negative Non-no Answers	ng	351	0.32	236	0.31	56	0.34	59	0.36
Maybe	am	349	0.32	224	0.30	66	0.40	59	0.36
Or-Clause	qrr	345	0.32	216	0.29	66	0.40	63	0.38
Exclamation	fe	307	0.28	195	0.26	56	0.34	56	0.34
Mimic Other	m	293	0.27	200	0.27	48	0.29	45	0.27
Apology	fa	259	0.24	181	0.24	46	0.28	32	0.19
About-task	t	253	0.23	154	0.21	42	0.25	57	0.35
Signal-non-understanding	br	236	0.22	161	0.21	39	0.23	36	0.22
Accept-part	aap	219	0.20	158	0.21	27	0.16	34	0.21
Rhetorical-Question	qh	214	0.20	166	0.22	30	0.18	18	0.11
Topic Change	tc	212	0.20	127	0.17	35	0.21	50	0.30
Repeat	r	208	0.19	131	0.17	45	0.27	32	0.19
Self-talk	t1	198	0.18	120	0.16	38	0.23	40	0.24
3rd-party-talk	t3	165	0.15	105	0.14	36	0.22	24	0.15
Rhetorical-question Continue	bh	154	0.14	109	0.15	26	0.16	19	0.12
Reject-part	bsc	150	0.14	94	0.13	22	0.13	34	0.21
Misspeak Self-Correction	arp	150	0.14	89	0.12	18	0.11	43	0.26
Reformulate/Summarize	bs	141	0.13	89	0.12	17	0.10	35	0.21
"Follow Me"	f	128	0.12	98	0.13	12	0.07	18	0.11
Or-Question	qr	127	0.12	88	0.12	17	0.10	22	0.13
Thanking	ft	119	0.11	88	0.12	9	0.05	22	0.13
Tag-Question	g	87	0.08	58	0.08	9	0.05	20	0.12
Open-Question	qo	74	0.07	49	0.07	14	0.08	11	0.07
Correct-misspeaking	bc	51	0.05	29	0.04	13	0.08	9	0.05
Sympathy	by	11	0.01	5	0.01	2	0.01	4	0.02
Welcome	fw	6	0.01	5	0.01	0	0.00	1	0.01

About the "%" Label

The "%" label represents "Uninterpretable" dialogue acts (4.8% of corpus). These include:

• Incomplete utterances: "which", "but it", "i mean"
• False starts and repetitions: "you'd have you'd have"
• Unclear vocalizations: "huh?", "hhh."
• Single fragment words: "and", "the", "so"

Important: All "%" labels have basic_da: "D" (Disruption), indicating they represent authentic conversational phenomena like interruptions, false starts, and unclear speech.

Usage recommendation: Keep these labels as they represent real conversational patterns essential for robust dialogue understanding. They can be treated as a valid dialogue act category or filtered out depending on your specific use case.

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("wylupek/mrda-corpus")

# Access different splits
train_data = dataset["train"]
test_data = dataset["test"] 
val_data = dataset["validation"]

Citation

Please cite the original paper:

@inproceedings{shriberg2004icsi,
    title={The ICSI meeting recorder dialog act (MRDA) corpus},
    author={Shriberg, Elizabeth and Dhillon, Raj and Bhagat, Sonali and Ang, Jeremy and Carvey, Hannah},
    booktitle={Proceedings of the 5th SIGdial Workshop on Discourse and Dialogue},
    pages={97--100},
    year={2004}
}

License

This dataset follows the original licensing terms. See the LICENSE file for details.

Note: This is a convenience reformatting for HuggingFace. All rights belong to the original ICSI authors.