Add readme
Browse files
README.md
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# A More Natural PersonaChat
|
| 2 |
+
|
| 3 |
+
## Dataset Summary
|
| 4 |
+
|
| 5 |
+
This dataset is a true-cased version of the PersonaChat dataset by Zhang et al. (2018).
|
| 6 |
+
The original PersonaChat dataset is all lower case, and has extra space around each
|
| 7 |
+
clause/sentence separating punctuation mark. This version of the dataset has more of a
|
| 8 |
+
natural language look, with sentence capitalization, proper noun capitalization, and
|
| 9 |
+
normalized whitespace. Also, each dialogue turn includes a pool of distractor
|
| 10 |
+
candidate responses, which can be used by a multiple choice regularization loss during
|
| 11 |
+
training.
|
| 12 |
+
|
| 13 |
+
## Lanugages
|
| 14 |
+
|
| 15 |
+
The text in the dataset is in English (**en**).
|
| 16 |
+
|
| 17 |
+
## Data Fields
|
| 18 |
+
|
| 19 |
+
Each instance of the dataset represents a conversational utterance that a
|
| 20 |
+
crowdworker made, while pretending to have a certain personality. Each instance has
|
| 21 |
+
these fields:
|
| 22 |
+
|
| 23 |
+
| Field Name | Datatype | Description |
|
| 24 |
+
|---------------|----------------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
|
| 25 |
+
| `conv_id` | int | A unique identifier for the instance's conversation. |
|
| 26 |
+
| `utterance_idx` | int | The index of the instance in the conversation. |
|
| 27 |
+
| `personality` | list of string | Sentences describing the personality of the current speaker. |
|
| 28 |
+
| `history` | list of string | The conversation's utterances so far, alternating between speakers with one utterance per speaker. |
|
| 29 |
+
| `candidates` | list of string | A list of utterances including distractor utterances as well as the true utterance the speaker gave, given their personality and the conversation history thus far. The true utterance is always the last utterance in this list. |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## Dataset Curation
|
| 33 |
+
|
| 34 |
+
The dataset was sourced from HuggingFace's version of the dataset used in the code for their
|
| 35 |
+
ConvAI 2018 submission, which was described in their [blog article](https://medium.com/huggingface/how-to-build-a-state-of-the-art-conversational-ai-with-transfer-learning-2d818ac26313)
|
| 36 |
+
on that submission. This version of the dataset has had extra white spaces removed,
|
| 37 |
+
and a StanfordNLP [stanza](https://stanfordnlp.github.io/stanza/) NLP pipeline was
|
| 38 |
+
used to conduct part-of-speech tagging to identify proper nouns, which were then
|
| 39 |
+
capitalized. The pipeline was also used to conduct sentence segmentation, allowing
|
| 40 |
+
the beginning of sentences to then be capitalized. Finally, all instances of the
|
| 41 |
+
pronoun "I" were capitalized, along with its contractions.
|