Processing the BT Oasis Corpus

Utilities for Processing the BT Oasis Corpus for the purpose of dialogue act (DA) classification. The data has been randomly split, with the training set comprising 80% of the dialogues (508), and test and validation sets 10% each (64).

Scripts

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

The process_transcript.py includes functions for processing each dialogue.

The oasis_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.

Data Format

Utterance are tagged with the SPAAC Annotation Scheme for 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 uninterpretable, unclassifiable, frag, decl and thirdParty are removed.

Example Utterances

b|i think i'm a bit behind on my payments|expressOpinion

a|right|ackn

a|um can you just bear with me|hold

Dialogue Acts

Dialogue Act	Labels	Count	%	Train Count	Train %	Test Count	Test %	Val Count	Val %
Inform	inform	3066	20.35	2422	20.06	307	20.77	337	22.27
Acknowledge	ackn	2015	13.37	1621	13.42	192	12.99	202	13.35
Request Inform	reqInfo	1404	9.32	1141	9.45	139	9.40	124	8.20
Backchannel	backch	1131	7.51	902	7.47	120	8.12	109	7.20
Answer	answ	844	5.60	677	5.61	92	6.22	75	4.96
Initialise	init	797	5.29	626	5.18	82	5.55	89	5.88
Thank	thank	770	5.11	621	5.14	74	5.01	75	4.96
Greet	greet	529	3.51	432	3.58	51	3.45	46	3.04
Accept	accept	464	3.08	373	3.09	43	2.91	48	3.17
Answer Elaborate	answElab	457	3.03	376	3.11	42	2.84	39	2.58
Inform Intention	informIntent	428	2.84	341	2.82	47	3.18	40	2.64
Bye	bye	412	2.73	335	2.77	36	2.44	41	2.71
Direct	direct	401	2.66	323	2.67	39	2.64	39	2.58
Confirm	confirm	349	2.32	284	2.35	21	1.42	44	2.91
Express Regret	expressRegret	255	1.69	213	1.76	18	1.22	24	1.59
Hold	hold	219	1.45	178	1.47	21	1.42	20	1.32
Express Opinion	expressOpinion	197	1.31	144	1.19	26	1.76	27	1.78
Offer	offer	157	1.04	129	1.07	12	0.81	16	1.06
Echo	echo	128	0.85	107	0.89	7	0.47	14	0.93
Appreciate	appreciate	111	0.74	90	0.75	11	0.74	10	0.66
Refer	refer	109	0.72	81	0.67	19	1.29	9	0.59
Suggest	suggest	103	0.68	81	0.67	10	0.68	12	0.79
Request Direct	reqDirect	94	0.62	74	0.61	12	0.81	8	0.53
Negate	negate	91	0.60	73	0.60	3	0.20	15	0.99
Exclaim	exclaim	83	0.55	68	0.56	10	0.68	5	0.33
Pardon	pardon	83	0.55	68	0.56	6	0.41	9	0.59
Identify Self	identifySelf	73	0.48	60	0.50	5	0.34	8	0.53
Express Possibility	expressPossibility	71	0.47	56	0.46	8	0.54	7	0.46
Raise Issue	raiseIssue	35	0.23	30	0.25	4	0.27	1	0.07
Express Wish	expressWish	34	0.23	28	0.23	2	0.14	4	0.26
Request Modal	reqModal	26	0.17	20	0.17	2	0.14	4	0.26
Complete	complete	20	0.13	13	0.11	6	0.41	1	0.07
Direct Elaborate	directElab	20	0.13	14	0.12	3	0.20	3	0.20
Correct	correct	19	0.13	15	0.12	2	0.14	2	0.13
Refuse	refuse	16	0.11	12	0.10	1	0.07	3	0.20
Inform Intent Hold	informIntent-hold	13	0.09	11	0.09	0	0.00	2	0.13
Inform Continue	informDisc	12	0.08	11	0.09	1	0.07	0	0.00
Inform Discontinue	informCont	12	0.08	10	0.08	2	0.14	0	0.00
Self Talk	selfTalk	10	0.07	7	0.06	2	0.14	1	0.07
Correct Self	correctSelf	7	0.05	7	0.06	0	0.00	0	0.00
Express Regret Inform	expressRegret-inform	1	0.01	1	0.01	0	0.00	0	0.00
Thank Identify Self	thank-identifySelf	1	0.01	1	0.01	0	0.00	0	0.00

Metadata

• Total number of utterances: 15067
• Max utterance length: 449
• Mean utterance length: 9.66
• Total Number of dialogues: 636
• Max dialogue length: 153
• Mean dialogue length: 23.69
• Vocabulary size: 2230
• Number of labels: 42
• Number of speakers: 2

Train set

• Number of dialogues: 508
• Max dialogue length: 153
• Mean dialogue length: 23.77
• Number of utterances: 12076

Test set

• Number of dialogues: 64
• Max dialogue length: 92
• Mean dialogue length: 23.27
• Number of utterances: 1489

Val set

• Number of dialogues: 64
• Max dialogue length: 78
• Mean dialogue length: 23.47
• Number of utterances: 1502

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 Oasis data.
• speakers = List of all speakers.
• num_speakers = Number of speakers in the Oasis data.

Each data set also has:

• <setname>_num_utterances = Number of utterances in the set.
• <setname>_num_dialogues = Number of dialogues in the set.
• <setname>_max_dialogue_len = Length of the longest dialogue in the set.
• <setname>_mean_dialogue_len = Mean length of dialogues in the set.