Bitext commited on
Commit
4913e79
·
1 Parent(s): 0e8da3b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +166 -91
README.md CHANGED
@@ -1,114 +1,189 @@
1
- Bitext - Customer Service Tagged Training Dataset for LLM-based Chatbots
2
- ========================================================================
3
 
4
- Overview
5
- --------
6
 
7
- This dataset can be used to train chatbots on Large Language Models such as GPT, Llama2 and Falcon.
8
 
9
- The dataset is parallel to our Evaluation dataset (see [Customer Service Tagged Evaluation Dataset for Intent Detection](https://github.com/bitext/customer-support-intent-detection-evaluation-dataset)). Both datasets can be used in conjunction to first train and then evaluate the accuracy provided by training. The main difference between the two datasets is the number of utterances:
10
 
11
- - The training dataset contains 4,269 utterances (around 200 per intent)
12
- - The evaluation dataset contains around 270,000 utterances (around 10,000 per intent)
 
 
 
 
13
 
14
- Both datasets share the rest of the specifications, so they can be used in conjunction. The training dataset has the following specs, shared with the evaluation dataset:
15
 
16
- - Customer Service domain
17
- - 10 categories or intent groups
18
- - 20 intents assigned to one of the 10 categories
19
- - 6 entity/slot types
20
 
21
- Each utterance is tagged with entities/slots when applicable. Additionally, each utterance is enriched with tags that indicate the type of language variation that the utterance expresses. Examples include:
22
 
23
- - The tag “COLLOQUIAL” indicates that the utterance contains informal expressions: “can u close my account”
24
- - The tag “INTERROGATIVE” indicates that the utterance is a question: “how do I open an account”
25
- - The tag “OFFENSIVE” indicates that the utterance contains offensive expressions: “open my f****** account”
26
 
27
- There are a total of 12 tags. See below for a full list of tags, categories and intents.
28
 
29
- The purpose of these tags is to customize the dataset so the trained bot can easily adapt to different user language profiles. A bot that sells sneakers and targets a younger population should be proficient in colloquial language; while a classical retail banking bot should be able to handle more formal or polite language.
30
 
31
- These intents have been selected from Bitext's collection of 20 domain-specific datasets (banking, retail, utilities...), covering the intents that are common across all 20 domains. For a full list of domains see https://www.bitext.com/chatbot-verticals/.
32
 
33
- Utterances and Linguistic Tags
34
- ------------------------------------
35
- The dataset contains 4,269 training utterances, with 200 utterances per intent. It has been split into training (80%), validation (10%) and testing (10%) sets, preserving the distribution of intents and linguistic phenomena.
36
 
37
- The dataset also reflects commonly occurring linguistic phenomena of real-life chatbots, such as spelling mistakes, run-on words, punctuation errors…
 
 
 
 
38
 
39
- Each entry in the dataset contains the following four fields:
40
 
41
- - context: the domain to which the entry applies
42
- - role: the role (virtual assistant, user...) that the model should adopt
43
- - instruction: a user request from the Customer Service domain
44
- - intent: the intent corresponding to the user instruction
45
- - entity_type: the type of entity contained in the utterance
46
- - entity_value: the entity contained in the utterance
47
- - start_offset: the starting position of the entity
48
- - end_offset: the ending position of the entity
49
- - category: the high-level semantic category for the intent
50
- - tags: different tags that reflect the types of language variations expressed in the utterance
51
- - response_type: identifier for tracking the composition, and version of chatbot responses.
52
- - response: an example expected response from the chatbot
53
 
54
- The dataset contains tags that reflect different language phenomena like colloquial or offensive language. So if an utterance for intent “cancel_order” contains the “COLLOQUIAL” tag, the utterance will express an informal language variation like: “can u cancel my order”
 
 
 
 
 
 
 
 
 
55
 
56
- Each utterance is enriched with one or more of these tags:
57
- - Register tags: colloquial language, polite language…
58
- - Q - Colloquial variation
59
- - P - Politeness variation
60
- - Content tags: offensive language, keyword language…
61
- - W - Offensive language
62
- - K - Keyword language
63
- - Linguistic tags: syntactic and morphological tags (interrogative sentence, coordinated sentence…)
64
- - B - Basic syntactic structure
65
- - C - Coordinated syntactic structure
66
- - I - Interrogative structure
67
- - N - negation (don't, can't…)
68
- - M - Morphological variation (plurals, tenses…)
69
- - L - Lexical variation (synonyms)
70
- - E - Expanded abbreviations (I'm -> I am, I'd -> I would…)
71
- - Real-life errors: spelling errors, punctuation errors…
72
- - Z - Noise phenomena like spelling or punctuation errors
73
 
74
- These tags indicate the type of language variation that the utterance expresses. When associated to each utterance, they allow Conversational Designers to customize training datasets to different user profiles with different uses of language. Through these tags, many different datasets can be created to make the resulting assistant more accurate and robust. A bot that sells sneakers should be mainly targeted to younger population that use a more colloquial language; while a classical retail banking bot should be able to handle more formal or polite language.
75
 
76
- You can find more details about the linguistic tags [here](TAGS.md)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
77
 
78
- Categories and Intents
79
- ----------------------
80
- The categories and intents covered by the dataset are:
81
-
82
- - ACCOUNT: create_account, delete_account, edit_account, recover_password, registration_problems, switch_account
83
- - CANCELLATION_FEE: check_cancellation_fee
84
- - CONTACT: contact_customer_service, contact_human_agent
85
- - DELIVERY: delivery_options, delivery_period
86
- - FEEDBACK: complaint, review
87
- - INVOICE: check_invoice, get_invoice
88
- - NEWSLETTER: newsletter_subscription,
89
- - ORDER: cancel_order, change_order, place_order, track_order
90
- - PAYMENT: check_payment_methods, payment_issue
91
- - REFUND: check_refund_policy, get_refund, track_refund
92
- - SHIPPING_ADDRESS: change_shipping_address, set_up_shipping_address
93
-
94
- Entities
95
- --------
96
- The entities covered by the dataset are:
97
 
98
- - account_type
99
- - Intents: create_account, delete_account, edit_account, switch_account
100
- - Values: Free, Freemium, Gold, Platinum, Premium, Pro, Standard
101
- - delivery_city
102
- - Intent: delivery_options
103
- - delivery_country
104
- - Intent: delivery_options
105
- - invoice_id
106
- - Intents: check_invoice, get_invoice
107
- - order_id
108
- - Intents: cancel_order, change_order, track_order
109
- - person_name
110
- - Intents: check_invoice, get_invoice
111
- - refund_amount
112
- - Intents: get_refund, track_refund
113
 
114
  (c) Bitext Innovations, 2023
 
1
+ # Bitext - Customer Service Tagged Training Dataset for LLM-based Virtual Assistants
 
2
 
3
+ ## Overview
 
4
 
5
+ This dataset can be used to train Large Language Models such as GPT, Llama2 and Falcon, both for Fine Tuning and Domain Adaptation.
6
 
7
+ The dataset has the following specs:
8
 
9
+ - Use Case: Intent Detection
10
+ - Vertical: Customer Service
11
+ - 27 intents assigned to 10 categories
12
+ - 26872 question/answer pairs, around 1000 per intent
13
+ - 30 entity/slot types
14
+ - 12 different types of language generation tags
15
 
16
+ The categories and intents have been selected from Bitext's collection of 20 vertical-specific datasets, covering the intents that are common across all 20 verticals. The verticals are:
17
 
18
+ - Automotive, Retail Banking, Education, Events & Ticketing, Field Services, Healthcare, Hospitality, Insurance, Legal Services, Manufacturing, Media Streaming, Mortgages & Loans, Moving & Storage, Real Estate/Construction, Restaurant & Bar Chains, Retail/E-commerce, Telecommunications, Travel, Utilities, Wealth Management
 
 
 
19
 
20
+ For a full list of verticals and its intents see [https://www.bitext.com/chatbot-verticals/](https://www.bitext.com/chatbot-verticals/).
21
 
22
+ The question/answer pairs have been generated using a hybrid methodology that uses natural texts as source text, NLP technology to extract seeds from these texts, and NLG technology to expand the seed texts. All steps in the process are curated by computational linguists.
 
 
23
 
24
+ ## Dataset Token Count
25
 
26
+ The dataset contains an extensive amount of text data across its 'instruction' and 'response' columns. After processing and tokenizing the dataset, we've identified a total of 3.57 million tokens. This rich set of tokens is essential for training advanced LLMs for AI Conversational, AI Generative, and Question and Answering (Q&A) models.
27
 
28
+ ## Fields of the Dataset
29
 
30
+ Each entry in the dataset contains the following fields:
 
 
31
 
32
+ - flags: tags (explained below in the Language Generation Tags section)
33
+ - instruction: a user request from the Customer Service domain
34
+ - category: the high-level semantic category for the intent
35
+ - intent: the intent corresponding to the user instruction
36
+ - response: an example expected response from the virtual assistant
37
 
38
+ ## Categories and Intents
39
 
40
+ The categories and intents covered by the dataset are:
 
 
 
 
 
 
 
 
 
 
 
41
 
42
+ - ACCOUNT: create_account, delete_account, edit_account, switch_account
43
+ - CANCELLATION_FEE: check_cancellation_fee
44
+ - DELIVERY: delivery_options
45
+ - FEEDBACK: complaint, review
46
+ - INVOICE: check_invoice, get_invoice
47
+ - NEWSLETTER: newsletter_subscription
48
+ - ORDER: cancel_order, change_order, place_order
49
+ - PAYMENT: check_payment_methods, payment_issue
50
+ - REFUND: check_refund_policy, track_refund
51
+ - SHIPPING_ADDRESS: change_shipping_address, set_up_shipping_address
52
 
53
+ ## Entities
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
54
 
55
+ The entities covered by the dataset are:
56
 
57
+ - {{Order Number}}, typically present in:
58
+ - Intents: cancel_order, change_order, change_shipping_address, check_invoice, check_refund_policy, complaint, delivery_options, delivery_period, get_invoice, get_refund, place_order, track_order, track_refund
59
+ - {{Invoice Number}}, typically present in:
60
+ - Intents: check_invoice, get_invoice
61
+ - {{Online Order Interaction}}, typically present in:
62
+ - Intents: cancel_order, change_order, check_refund_policy, delivery_period, get_refund, review, track_order, track_refund
63
+ - {{Online Payment Interaction}}, typically present in:
64
+ - Intents: cancel_order, check_payment_methods
65
+ - {{Online Navigation Step}}, typically present in:
66
+ - Intents: complaint, delivery_options
67
+ - {{Online Customer Support Channel}}, typically present in:
68
+ - Intents: check_refund_policy, complaint, contact_human_agent, delete_account, delivery_options, edit_account, get_refund, payment_issue, registration_problems, switch_account
69
+ - {{Profile}}, typically present in:
70
+ - Intent: switch_account
71
+ - {{Profile Type}}, typically present in:
72
+ - Intent: switch_account
73
+ - {{Settings}}, typically present in:
74
+ - Intents: cancel_order, change_order, change_shipping_address, check_cancellation_fee, check_invoice, check_payment_methods, contact_human_agent, delete_account, delivery_options, edit_account, get_invoice, newsletter_subscription, payment_issue, place_order, recover_password, registration_problems, set_up_shipping_address, switch_account, track_order, track_refund
75
+ - {{Online Company Portal Info}}, typically present in:
76
+ - Intents: cancel_order, edit_account
77
+ - {{Date}}, typically present in:
78
+ - Intents: check_invoice, check_refund_policy, get_refund, track_order, track_refund
79
+ - {{Date Range}}, typically present in:
80
+ - Intents: check_cancellation_fee, check_invoice, get_invoice
81
+ - {{Shipping Cut-off Time}}, typically present in:
82
+ - Intent: delivery_options
83
+ - {{Delivery City}}, typically present in:
84
+ - Intent: delivery_options
85
+ - {{Delivery Country}}, typically present in:
86
+ - Intents: check_payment_methods, check_refund_policy, delivery_options, review, switch_account
87
+ - {{Salutation}}, typically present in:
88
+ - Intents: cancel_order, check_payment_methods, check_refund_policy, create_account, delete_account, delivery_options, get_refund, recover_password, review, set_up_shipping_address, switch_account, track_refund
89
+ - {{Client First Name}}, typically present in:
90
+ - Intents: check_invoice, get_invoice
91
+ - {{Client Last Name}}, typically present in:
92
+ - Intents: check_invoice, create_account, get_invoice
93
+ - {{Customer Support Phone Number}}, typically present in:
94
+ - Intents: change_shipping_address, contact_customer_service, contact_human_agent, payment_issue
95
+ - {{Customer Support Email}}, typically present in:
96
+ - Intents: cancel_order, change_shipping_address, check_invoice, check_refund_policy, complaint, contact_customer_service, contact_human_agent, get_invoice, get_refund, newsletter_subscription, payment_issue, recover_password, registration_problems, review, set_up_shipping_address, switch_account
97
+ - {{Live Chat Support}}, typically present in:
98
+ - Intents: check_refund_policy, complaint, contact_human_agent, delete_account, delivery_options, edit_account, get_refund, payment_issue, recover_password, registration_problems, review, set_up_shipping_address, switch_account, track_order
99
+ - {{Website URL}}, typically present in:
100
+ - Intents: check_payment_methods, check_refund_policy, complaint, contact_customer_service, contact_human_agent, create_account, delete_account, delivery_options, get_refund, newsletter_subscription, payment_issue, place_order, recover_password, registration_problems, review, switch_account
101
+ - {{Upgrade Account}}, typically present in:
102
+ - Intents: create_account, edit_account, switch_account
103
+ - {{Account Type}}, typically present in:
104
+ - Intents: cancel_order, change_order, change_shipping_address, check_cancellation_fee, check_invoice, check_payment_methods, check_refund_policy, complaint, contact_customer_service, contact_human_agent, create_account, delete_account, delivery_options, delivery_period, edit_account, get_invoice, get_refund, newsletter_subscription, payment_issue, place_order, recover_password, registration_problems, review, set_up_shipping_address, switch_account, track_order, track_refund
105
+ - {{Account Category}}, typically present in:
106
+ - Intents: cancel_order, change_order, change_shipping_address, check_cancellation_fee, check_invoice, check_payment_methods, check_refund_policy, complaint, contact_customer_service, contact_human_agent, create_account, delete_account, delivery_options, delivery_period, edit_account, get_invoice, get_refund, newsletter_subscription, payment_issue, place_order, recover_password, registration_problems, review, set_up_shipping_address, switch_account, track_order, track_refund
107
+ - {{Account Change}}, typically present in:
108
+ - Intent: switch_account
109
+ - {{Program}}, typically present in:
110
+ - Intent: place_order
111
+ - {{Refund Amount}}, typically present in:
112
+ - Intent: track_refund
113
+ - {{Money Amount}}, typically present in:
114
+ - Intents: check_refund_policy, complaint, get_refund, track_refund
115
+ - {{Store Location}}, typically present in:
116
+ - Intents: complaint, delivery_options, place_order
117
+
118
+ ## Language Generation Tags
119
+
120
+ The dataset contains tags that reflect how language varies/changes across different linguistic phenomena like colloquial or offensive language. So if an utterance for intent “cancel_order” contains the “COLLOQUIAL” tag, the utterance will express an informal language variation like: “can u cancel my order”.
121
+
122
+ These tags indicate the type of language variation that the entry expresses. When associated to each entry, they allow Conversational Designers to customize training datasets to different user profiles with different uses of language. Through these tags, many different datasets can be created to make the resulting assistant more accurate and robust. A bot that sells sneakers should be mainly targeted to younger population that use a more colloquial language; while a classical retail banking bot should be able to handle more formal or polite language. The dataset also reflects commonly occurring linguistic phenomena of real-life virtual assistant, such as spelling mistakes, run-on words, punctuation errors…
123
+
124
+ The dataset contains tagging for all relevant linguistic phenomena that can be used to customize the dataset for different user profiles.
125
+
126
+ ### Tags for Lexical variation
127
+
128
+ M - Morphological variation: inflectional and derivational
129
+ “is my SIM card active”, “is my SIM card activated”
130
+
131
+ L - Semantic variations: synonyms, use of hyphens, compounding…
132
+ “what’s my billing date", “what’s my anniversary date”
133
+
134
+ ### Tags for Syntactic structure variation
135
+
136
+ B - Basic syntactic structure:
137
+ “activate my SIM card”, “I need to activate my SIM card”
138
+
139
+ I - Interrogative structure
140
+ “can you activate my SIM card?”, “how do I activate my SIM card?”
141
+
142
+ C - Coordinated syntactic structure
143
+ “I have a new SIM card, what do I need to do to activate it?”
144
+
145
+ N - Negation
146
+ “I do not want this item, where to cancel my order?”
147
+
148
+ ### Tags for language register variations
149
+
150
+ P - Politeness variation
151
+ “could you help me activate my SIM card, please?”
152
+
153
+ Q - Colloquial variation
154
+ “can u activ8 my SIM?”
155
+
156
+ W - Offensive language
157
+ “I want to talk to a f*&%*g agent”
158
+
159
+ ### Tags for stylistic variations
160
+
161
+ K - Keyword mode
162
+ "activate SIM", "new SIM"
163
+
164
+ E - Use of abbreviations:
165
+ “I'm / I am interested in getting a new SIM”
166
+
167
+ Z - Errors and Typos: spelling issues, wrong punctuation…
168
+ “how can i activaet my card”
169
+
170
+ ### Other tags not in use in this Dataset
171
+
172
+ D - Indirect speech
173
+ “ask my agent to activate my SIM card”
174
+
175
+ G - Regional variations
176
+ US English vs UK English: "truck" vs "lorry"
177
+ France French vs Canadian French: "tchatter" vs "clavarder"
178
+
179
+ R - Respect structures - Language-dependent variations
180
+ English: "may" vs "can…"
181
+ French: "tu" vs "vous..."
182
+ Spanish: "tú" vs "usted..."
183
 
184
+ Y - Code switching
185
+ “activer ma SIM card”
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
186
 
187
+ ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
188
 
189
  (c) Bitext Innovations, 2023