数据集:
silicone
语言:
en计算机处理:
monolingual语言创建人:
expert-generated批注创建人:
expert-generated源数据集:
original预印本库:
arxiv:2009.11152许可:
cc-by-sa-4.0The Sequence labellIng evaLuatIon benChmark fOr spoken laNguagE (SILICONE) benchmark is a collection of resources for training, evaluating, and analyzing natural language understanding systems specifically designed for spoken language. All datasets are in the English language and covers a variety of domains including daily life, scripted scenarios, joint task completion, phone call conversations, and televsion dialogue. Some datasets additionally include emotion and/or sentimant labels.
[More Information Needed]
English.
For the dyda_da configuration one example from the dataset is:
{ 'Utterance': "the taxi drivers are on strike again .", 'Dialogue_Act': 2, # "inform" 'Dialogue_ID': "2" }DailyDialog Act Corpus (Emotion)
For the dyda_e configuration one example from the dataset is:
{ 'Utterance': "'oh , breaktime flies .'", 'Emotion': 5, # "sadness" 'Dialogue_ID': "997" }Interactive Emotional Dyadic Motion Capture (IEMOCAP) database
For the iemocap configuration one example from the dataset is:
{ 'Dialogue_ID': "Ses04F_script03_2", 'Utterance_ID': "Ses04F_script03_2_F025", 'Utterance': "You're quite insufferable. I expect it's because you're drunk.", 'Emotion': 0, # "ang" }HCRC MapTask Corpus
For the maptask configuration one example from the dataset is:
{ 'Speaker': "f", 'Utterance': "i think that would bring me over the crevasse", 'Dialogue_Act': 4, # "explain" }Multimodal EmotionLines Dataset (Emotion)
For the meld_e configuration one example from the dataset is:
{ 'Utterance': "'Push 'em out , push 'em out , harder , harder .'", 'Speaker': "Joey", 'Emotion': 3, # "joy" 'Dialogue_ID': "1", 'Utterance_ID': "2" }Multimodal EmotionLines Dataset (Sentiment)
For the meld_s configuration one example from the dataset is:
{ 'Utterance': "'Okay , y'know what ? There is no more left , left !'", 'Speaker': "Rachel", 'Sentiment': 0, # "negative" 'Dialogue_ID': "2", 'Utterance_ID': "4" }ICSI MRDA Corpus
For the mrda configuration one example from the dataset is:
{ 'Utterance_ID': "Bed006-c2_0073656_0076706", 'Dialogue_Act': 0, # "s" 'Channel_ID': "Bed006-c2", 'Speaker': "mn015", 'Dialogue_ID': "Bed006", 'Utterance': "keith is not technically one of us yet ." }BT OASIS Corpus
For the oasis configuration one example from the dataset is:
{ 'Speaker': "b", 'Utterance': "when i rang up um when i rang to find out why she said oh well your card's been declined", 'Dialogue_Act': 21, # "inform" }SEMAINE database
For the sem configuration one example from the dataset is:
{ 'Utterance': "can you think of somebody who is like that ?", 'NbPairInSession': "11", 'Dialogue_ID': "59", 'SpeechTurn': "674", 'Speaker': "Agent", 'Sentiment': 1, # "Neutral" }Switchboard Dialog Act (SwDA) Corpus
For the swda configuration one example from the dataset is:
{ 'Utterance': "but i 'd probably say that 's roughly right .", 'Dialogue_Act': 33, # "aap_am" 'From_Caller': "1255", 'To_Caller': "1087", 'Topic': "CRIME", 'Dialogue_ID': "818", 'Conv_ID': "sw2836", }
For the dyda_da configuration, the different fields are:
For the dyda_e configuration, the different fields are:
For the iemocap configuration, the different fields are:
For the maptask configuration, the different fields are:
For the meld_e configuration, the different fields are:
For the meld_s configuration, the different fields are:
For the mrda configuration, the different fields are:
For the oasis configuration, the different fields are:
For the sem configuration, the different fields are:
For the swda configuration, the different fields are: Utterance : Utterance as a string. Dialogue_Act : Dialogue act label of the utterance. It can be "sd" (0) [Statement-non-opinion], "b" (1) [Acknowledge (Backchannel)], "sv" (2) [Statement-opinion], "%" (3) [Uninterpretable], "aa" (4) [Agree/Accept], "ba" (5) [Appreciation], "fc" (6) [Conventional-closing], "qw" (7) [Wh-Question], "nn" (8) [No Answers], "bk" (9) [Response Acknowledgement], "h" (10) [Hedge], "qy^d" (11) [Declarative Yes-No-Question], "bh" (12) [Backchannel in Question Form], "^q" (13) [Quotation], "bf" (14) [Summarize/Reformulate], 'fo_o_fw_" by_bc' (15) [Other], 'fo_o_fw_by_bc "' (16) [Other], "na" (17) [Affirmative Non-yes Answers], "ad" (18) [Action-directive], "^2" (19) [Collaborative Completion], "b^m" (20) [Repeat-phrase], "qo" (21) [Open-Question], "qh" (22) [Rhetorical-Question], "^h" (23) [Hold Before Answer/Agreement], "ar" (24) [Reject], "ng" (25) [Negative Non-no Answers], "br" (26) [Signal-non-understanding], "no" (27) [Other Answers], "fp" (28) [Conventional-opening], "qrr" (29) [Or-Clause], "arp_nd" (30) [Dispreferred Answers], "t3" (31) [3rd-party-talk], "oo_co_cc" (32) [Offers, Options Commits], "aap_am" (33) [Maybe/Accept-part], "t1" (34) [Downplayer], "bd" (35) [Self-talk], "^g" (36) [Tag-Question], "qw^d" (37) [Declarative Wh-Question], "fa" (38) [Apology], "ft" (39) [Thanking], "+" (40) [Unknown], "x" (41) [Unknown], "ny" (42) [Unknown], "sv_fx" (43) [Unknown], "qy_qr" (44) [Unknown] or "ba_fe" (45) [Unknown]. From_Caller : identifier of the from caller as a string. To_Caller : identifier of the to caller as a string. Topic : Topic as a string. Dialogue_ID : identifier of the dialogue as a string. Conv_ID : identifier of the conversation as a string.
Dataset name | Train | Valid | Test |
---|---|---|---|
dyda_da | 87170 | 8069 | 7740 |
dyda_e | 87170 | 8069 | 7740 |
iemocap | 7213 | 805 | 2021 |
maptask | 20905 | 2963 | 2894 |
meld_e | 9989 | 1109 | 2610 |
meld_s | 9989 | 1109 | 2610 |
mrda | 83944 | 9815 | 15470 |
oasis | 12076 | 1513 | 1478 |
sem | 4264 | 485 | 878 |
swda | 190709 | 21203 | 2714 |
[More Information Needed]
[More Information Needed]
Who are the source language producers?[More Information Needed]
[More Information Needed]
Who are the annotators?[More Information Needed]
[More Information Needed]
[More Information Needed]
[More Information Needed]
[More Information Needed]
Emile Chapuis, Pierre Colombo, Ebenge Usip.
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 Unported License .
@inproceedings{chapuis-etal-2020-hierarchical, title = "Hierarchical Pre-training for Sequence Labelling in Spoken Dialog", author = "Chapuis, Emile and Colombo, Pierre and Manica, Matteo and Labeau, Matthieu and Clavel, Chlo{\'e}", booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2020", month = nov, year = "2020", address = "Online", publisher = "Association for Computational Linguistics", url = "https://www.aclweb.org/anthology/2020.findings-emnlp.239", doi = "10.18653/v1/2020.findings-emnlp.239", pages = "2636--2648", abstract = "Sequence labelling tasks like Dialog Act and Emotion/Sentiment identification are a key component of spoken dialog systems. In this work, we propose a new approach to learn generic representations adapted to spoken dialog, which we evaluate on a new benchmark we call Sequence labellIng evaLuatIon benChmark fOr spoken laNguagE benchmark (SILICONE). SILICONE is model-agnostic and contains 10 different datasets of various sizes. We obtain our representations with a hierarchical encoder based on transformer architectures, for which we extend two well-known pre-training objectives. Pre-training is performed on OpenSubtitles: a large corpus of spoken dialog containing over 2.3 billion of tokens. We demonstrate how hierarchical encoders achieve competitive results with consistently fewer parameters compared to state-of-the-art models and we show their importance for both pre-training and fine-tuning.", }