Processing the Switchboard Dialogue Act Corpus

Utilities for processing the Switchboard Dialogue Act Corpus for the purpose of dialogue act (DA) classification. The data is split into the original training and test sets suggested by the authors (1115 training and 19 test). The remaining 21 dialogues have been used as a validation set.

Scripts

The swda_to_text.py script processes all dialogues into a plain text format. Individual dialogues are saved into directories corresponding to the set they belong to (train, test, etc). All utterances in a particular set are also saved to a text file.

The swda_utilities.py script contains various helper functions for loading/saving the data.

The process_transcript.py includes functions for processing each dialogue.

The swda_metadata.py generates various metadata from the processed dialogues and saves them as a dictionary to a pickle file. The words, labels and frequencies are also saved as plain text files in the /metadata directory.

Thanks to Christopher Potts for providing the raw data in .csv format and the swda.py script for processing the .csv data, both of which can be found here

Data Format

Utterance are tagged with the SWBD-DAMSL DA.

By default:

• Utterances are written one per line in the format Speaker | Utterance Text | Dialogue Act Tag.
• Setting the utterance_only_flag == True, will change the default output to only one utterance per line i.e. no speaker or DA tags.
• Utterances marked as Non-verbal ('x' tags) are removed i.e. 'Laughter' or 'Throat_clearing'.
• Utterances marked as Interrupted ('+' tags) and continued later are concatenated to make un-interrupted sentences.
• All disfluency annotations are removed i.e. '#', '<', '>', etc.

Example Utterances

A|What is the nature of your company's business?|qw

B|Well, it's actually, uh,|^h

B|we do oil well services.|sd

Dialogue Acts

Dialogue Act	Labels	Count	%	Train Count	Train %	Test Count	Test %	Val Count	Val %
Statement-non-opinion	sd	75136	37.62	72549	37.71	1317	32.30	1270	38.81
Acknowledge (Backchannel)	b	38281	19.17	36950	19.21	764	18.73	567	17.33
Statement-opinion	sv	26421	13.23	25087	13.04	718	17.61	616	18.83
Uninterpretable	%	15195	7.61	14597	7.59	349	8.56	249	7.61
Agree/Accept	aa	11123	5.57	10770	5.60	207	5.08	146	4.46
Appreciation	ba	4757	2.38	4619	2.40	76	1.86	62	1.89
Yes-No-Question	qy	4725	2.37	4594	2.39	84	2.06	47	1.44
Yes Answers	ny	3030	1.52	2918	1.52	73	1.79	39	1.19
Conventional-closing	fc	2581	1.29	2480	1.29	81	1.99	20	0.61
Wh-Question	qw	1976	0.99	1896	0.99	55	1.35	25	0.76
No Answers	nn	1374	0.69	1334	0.69	26	0.64	14	0.43
Response Acknowledgement	bk	1306	0.65	1271	0.66	28	0.69	7	0.21
Hedge	h	1226	0.61	1181	0.61	23	0.56	22	0.67
Declarative Yes-No-Question	qy^d	1218	0.61	1167	0.61	36	0.88	15	0.46
Backchannel in Question Form	bh	1053	0.53	1015	0.53	21	0.51	17	0.52
Quotation	^q	983	0.49	931	0.48	17	0.42	35	1.07
Summarize/Reformulate	bf	952	0.48	905	0.47	23	0.56	24	0.73
Other	fo_o_fw_"_by_bc	879	0.44	857	0.45	15	0.37	7	0.21
Affirmative Non-yes Answers	na	847	0.42	831	0.43	10	0.25	6	0.18
Action-directive	ad	745	0.37	712	0.37	27	0.66	6	0.18
Collaborative Completion	^2	723	0.36	690	0.36	19	0.47	14	0.43
Repeat-phrase	b^m	687	0.34	655	0.34	21	0.51	11	0.34
Open-Question	qo	656	0.33	631	0.33	16	0.39	9	0.28
Rhetorical-Question	qh	575	0.29	554	0.29	12	0.29	9	0.28
Hold Before Answer/Agreement	^h	556	0.28	539	0.28	7	0.17	10	0.31
Reject	ar	344	0.17	337	0.18	3	0.07	4	0.12
Negative Non-no Answers	ng	302	0.15	290	0.15	6	0.15	6	0.18
Signal-non-understanding	br	298	0.15	286	0.15	9	0.22	3	0.09
Other Answers	no	284	0.14	277	0.14	6	0.15	1	0.03
Conventional-opening	fp	225	0.11	220	0.11	5	0.12	0	0.00
Or-Clause	qrr	209	0.10	206	0.11	2	0.05	1	0.03
Dispreferred Answers	arp_nd	207	0.10	204	0.11	3	0.07	0	0.00
3rd-party-talk	t3	117	0.06	115	0.06	0	0.00	2	0.06
Offers, Options Commits	oo_co_cc	110	0.06	109	0.06	0	0.00	1	0.03
Maybe/Accept-part	aap_am	104	0.05	97	0.05	7	0.17	0	0.00
Downplayer	t1	103	0.05	102	0.05	1	0.02	0	0.00
Self-talk	bd	103	0.05	100	0.05	1	0.02	2	0.06
Tag-Question	^g	92	0.05	92	0.05	0	0.00	0	0.00
Declarative Wh-Question	qw^d	80	0.04	79	0.04	1	0.02	0	0.00
Apology	fa	79	0.04	76	0.04	2	0.05	1	0.03
Thanking	ft	78	0.04	67	0.03	7	0.17	4	0.12

Metadata

• Total number of utterances: 199740
• Maximum utterance length: 133
• Mean utterance length: 9.6
• Total number of dialogues: 1155
• Maximum dialogue length: 457
• Mean dialogue length: 172.9
• Vocabulary size: 22301
• Number of labels: 41
• Number of dialogue in train set: 1115
• Maximum length of dialogue in train set: 457
• Number of dialogue in test set: 19
• Maximum length of dialogue in test set: 330
• Number of dialogue in val set: 21
• Maximum length of dialogue in val set: 299

Keys and values for the metadata dictionary

• num_utterances = Total number of utterance in the full corpus.
• max_utterance_len = Number of words in the longest utterance in the corpus.
• mean_utterance_len = Average number of words in utterances.
• num_dialogues = Total number of dialogues in the corpus.
• max_dialogues_len = Number of utterances in the longest dialogue in the corpus.
• mean_dialogues_len = Average number of utterances in dialogues.
• word_freq = Dataframe with Word and Count columns.
• vocabulary = List of all words in vocabulary.
• vocabulary_size = Number of words in the vocabulary.
• label_freq = Dataframe containing all data in the Dialogue Acts table above.
• labels = List of all DA labels.
• num_labels = Number of labels used from the Switchboard data.

Each data set also has:

• <setname>_num_dialogues = Number of dialogues in the set.
• <setname>_max_dialogues_len = Length of the longest dialogue in the set.