Datasets:

Languages:
English
ArXiv:
License:
zhuqi commited on
Commit
c03ac92
1 Parent(s): 9b483a3

Upload README.md

Browse files
Files changed (1) hide show
  1. README.md +61 -0
README.md ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Dataset Card for Taskmaster-1
2
+
3
+ - **Repository:** https://github.com/google-research-datasets/Taskmaster/tree/master/TM-2-2020
4
+ - **Paper:** https://arxiv.org/pdf/1909.05358.pdf
5
+ - **Leaderboard:** None
6
+ - **Who transforms the dataset:** Qi Zhu(zhuq96 at gmail dot com)
7
+
8
+ ### Dataset Summary
9
+
10
+ The Taskmaster-2 dataset consists of 17,289 dialogs in the seven domains. Unlike Taskmaster-1, which includes both written "self-dialogs" and spoken two-person dialogs, Taskmaster-2 consists entirely of spoken two-person dialogs. In addition, while Taskmaster-1 is almost exclusively task-based, Taskmaster-2 contains a good number of search- and recommendation-oriented dialogs, as seen for example in the restaurants, flights, hotels, and movies verticals. The music browsing and sports conversations are almost exclusively search- and recommendation-based. All dialogs in this release were created using a Wizard of Oz (WOz) methodology in which crowdsourced workers played the role of a 'user' and trained call center operators played the role of the 'assistant'. In this way, users were led to believe they were interacting with an automated system that “spoke” using text-to-speech (TTS) even though it was in fact a human behind the scenes. As a result, users could express themselves however they chose in the context of an automated interface.
11
+
12
+ - **How to get the transformed data from original data:**
13
+ - Download [master.zip](https://github.com/google-research-datasets/Taskmaster/archive/refs/heads/master.zip).
14
+ - Run `python preprocess.py` in the current directory.
15
+ - **Main changes of the transformation:**
16
+ - Remove dialogs that are empty or only contain one speaker.
17
+ - Split each domain dialogs into train/validation/test randomly (8:1:1).
18
+ - Merge continuous turns by the same speaker (ignore repeated turns).
19
+ - Annotate `dialogue acts` according to the original segment annotations. Add `intent` annotation (`==inform`). The type of `dialogue act` is set to `non-categorical` if the `slot` is not in `anno2slot` in `preprocess.py`). Otherwise, the type is set to `binary` (and the `value` is empty). If there are multiple spans overlapping, we only keep the shortest one, since we found that this simple strategy can reduce the noise in annotation.
20
+ - Add `domain`, `intent`, and `slot` descriptions.
21
+ - Add `state` by accumulate `non-categorical dialogue acts` in the order that they appear.
22
+ - Keep the first annotation since each conversation was annotated by two workers.
23
+ - **Annotations:**
24
+ - dialogue acts, state.
25
+
26
+ ### Supported Tasks and Leaderboards
27
+
28
+ NLU, DST, Policy, NLG
29
+
30
+ ### Languages
31
+
32
+ English
33
+
34
+ ### Data Splits
35
+
36
+ | split | dialogues | utterances | avg_utt | avg_tokens | avg_domains | cat slot match(state) | cat slot match(goal) | cat slot match(dialogue act) | non-cat slot span(dialogue act) |
37
+ |------------|-------------|--------------|-----------|--------------|---------------|-------------------------|------------------------|--------------------------------|-----------------------------------|
38
+ | train | 13838 | 234321 | 16.93 | 9.1 | 1 | - | - | - | 100 |
39
+ | validation | 1731 | 29349 | 16.95 | 9.15 | 1 | - | - | - | 100 |
40
+ | test | 1734 | 29447 | 16.98 | 9.07 | 1 | - | - | - | 100 |
41
+ | all | 17303 | 293117 | 16.94 | 9.1 | 1 | - | - | - | 100 |
42
+
43
+ 7 domains: ['flights', 'food-ordering', 'hotels', 'movies', 'music', 'restaurant-search', 'sports']
44
+ - **cat slot match**: how many values of categorical slots are in the possible values of ontology in percentage.
45
+ - **non-cat slot span**: how many values of non-categorical slots have span annotation in percentage.
46
+
47
+ ### Citation
48
+
49
+ ```
50
+ @inproceedings{byrne-etal-2019-taskmaster,
51
+ title = {Taskmaster-1:Toward a Realistic and Diverse Dialog Dataset},
52
+ author = {Bill Byrne and Karthik Krishnamoorthi and Chinnadhurai Sankar and Arvind Neelakantan and Daniel Duckworth and Semih Yavuz and Ben Goodrich and Amit Dubey and Kyu-Young Kim and Andy Cedilnik},
53
+ booktitle = {2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint Conference on Natural Language Processing},
54
+ address = {Hong Kong},
55
+ year = {2019}
56
+ }
57
+ ```
58
+
59
+ ### Licensing Information
60
+
61
+ [**CC BY 4.0**](https://creativecommons.org/licenses/by/4.0/)