Parrot is a paraphrase based utterance augmentation framework purpose built to accelerate training NLU models. A paraphrase framework is more than just a paraphrasing model.
- Why Parrot?
- Getting started
- Scope
- What makes a paraphraser a good augmentor for NLU? (Details)
- Power of Augmentation - Metrics and Comparison
- Current Features
- Roadmap
- Current Limitations/Known issues
- References
- Citation
Huggingface lists 16 paraphrase generation models, (as of this writing) RapidAPI lists 7 fremium and commercial paraphrasers like QuillBot, Rasa has discussed an experimental paraphraser for augmenting text data here, Sentence-transfomers offers a paraphrase mining utility and NLPAug offers word level augmentation with a PPDB (a multi-million paraphrase database). While these attempts at paraphrasing are great, there are still some gaps and paraphrasing is NOT yet a mainstream option for text augmentation in building NLU models....Parrot is a humble attempt to fill some of these gaps.
What is a good paraphrase? Almost all conditioned text generation models are validated on 2 factors, (1) if the generated text conveys the same meaning as the original context (Adequacy) (2) if the text is fluent / grammatically correct english (Fluency). For instance Neural Machine Translation outputs are tested for Adequacy and Fluency. But a good paraphrase should be adequate and fluent while being as different as possible on the surface lexical form. With respect to this definition, the 3 key metrics that measures the quality of paraphrases are:
- Adequacy (Is the meaning preserved adequately?)
- Fluency (Is the paraphrase fluent English?)
- Diversity (Lexical / Phrasal / Syntactical) (How much has the paraphrase changed the original sentence?)
Parrot offers knobs to control Adequacy, Fluency and Diversity as per your needs.
What makes a paraphraser a good augmentor? For training a NLU model we just don't need a lot of utterances but utterances with intents and slots/entities annotated. Typical flow would be:
- Given an input utterance + input annotations a good augmentor spits out N output paraphrases while preserving the intent and slots.
- The output paraphrases are then converted into annotated data using the input annotations that we got in step 1.
- The annotated data created out of the output paraphrases then makes the training dataset for your NLU model.
But in general being a generative model paraphrasers doesn't guarantee to preserve the slots/entities. So the ability to generate high quality paraphrases in a constrained fashion without trading off the intents and slots for lexical dissimilarity makes a paraphraser a good augmentor. More on this in section 3 below
pip install git+https://github.com/PrithivirajDamodaran/Parrot_Paraphraser.git
from parrot import Parrot
import torch
import warnings
warnings.filterwarnings("ignore")
'''
uncomment to get reproducable paraphrase generations
def random_state(seed):
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
random_state(1234)
'''
#Init models (make sure you init ONLY once if you integrate this to your code)
parrot = Parrot(model_tag="prithivida/parrot_paraphraser_on_T5")
phrases = ["Can you recommed some upscale restaurants in Newyork?",
"What are the famous places we should not miss in Russia?"
]
for phrase in phrases:
print("-"*100)
print("Input_phrase: ", phrase)
print("-"*100)
para_phrases = parrot.augment(input_phrase=phrase, use_gpu=False)
for para_phrase in para_phrases:
print(para_phrase)
----------------------------------------------------------------------
Input_phrase: Can you recommed some upscale restaurants in Newyork?
----------------------------------------------------------------------
list some excellent restaurants to visit in new york city?
what upscale restaurants do you recommend in new york?
i want to try some upscale restaurants in new york?
recommend some upscale restaurants in newyork?
can you recommend some high end restaurants in newyork?
can you recommend some upscale restaurants in new york?
can you recommend some upscale restaurants in newyork?
----------------------------------------------------------------------
Input_phrase: What are the famous places we should not miss in Russia
----------------------------------------------------------------------
what should we not miss when visiting russia?
recommend some of the best places to visit in russia?
list some of the best places to visit in russia?
can you list the top places to visit in russia?
show the places that we should not miss in russia?
list some famous places which we should not miss in russia?
You can play with the do_diverse knob (checkout the next section for more knobs). Consider this example: do_diverse = False (default)*
------------------------------------------------------------------------------
Input_phrase: How are the new Macbook Pros with M1 chips?
------------------------------------------------------------------------------
'how do you rate the new macbook pros? '
'how are the new macbook pros? '
'how is the new macbook pro doing with new chips? '
'how do you like the new macbook pro m1 chip? '
'what is the use of the new macbook pro m1 chips? '
do_diverse = True
------------------------------------------------------------------------------
Input_phrase: How are the new Macbook Pros with M1 chips?
------------------------------------------------------------------------------
'what do you think about the new macbook pro m1? '
'how is the new macbook pro m1? '
'how are the new macbook pros? '
'what do you think about the new macbook pro m1 chips? '
'how good is the new macbook pro m1 chips? '
'how is the new macbook pro m1 chip? '
'do you like the new macbook pro m1 chips? '
'how are the new macbook pros with m1 chips? '
para_phrases = parrot.augment(input_phrase=phrase,
use_gpu=False,
diversity_ranker="levenshtein",
do_diverse=False,
max_return_phrases = 10,
max_length=32,
adequacy_threshold = 0.99,
fluency_threshold = 0.90)
In the space of conversational engines, knowledge bots are to which we ask questions like "when was the Berlin wall teared down?", transactional bots are to which we give commands like "Turn on the music please" and voice assistants are the ones which can do both answer questions and action our commands. Parrot mainly foucses on augmenting texts typed-into or spoken-to conversational interfaces for building robust NLU models. (So usually people neither type out or yell out long paragraphs to conversational interfaces. Hence the pre-trained model is trained on text samples of maximum length of 32.)
While Parrot predominantly aims to be a text augmentor for building good NLU models, it can also be used as a pure-play paraphraser.
To enable automatic training data generation, a paraphraser needs to keep the slots in intact. So the end to end process can take input utternaces, augment and convert them into NLU training format goo et al or rasa format (as shown below). The data generation process needs to look for the same slots in the output paraphrases to derive the start and end positions.(as shown in the json below)
Ideally the above process needs an UI like below to collect to input utternaces along with annotations (Intents, Slots and slot types) which then can be agumented and converted to training data.
{
"rasa_nlu_data": {
"common_examples": [
{
"text": "i would like to find a flight from charlotte to las vegas that makes a stop in st. louis",
"intent": "flight",
"entities": [
{
"start": 35,
"end": 44,
"value": "charlotte",
"entity": "fromloc.city_name"
},
{
"start": 48,
"end": 57,
"value": "las vegas",
"entity": "toloc.city_name"
},
{
"start": 79,
"end": 88,
"value": "st. louis",
"entity": "stoploc.city_name"
}
]
},
...
]
}
}
- Original: I would like a list of round trip flights between indianapolis and orlando florida for the 27th
- Paraphrase useful for augmenting: what are the round trip flights between indianapolis and orlando for the 27th
- Paraphrase not-so-useful for augmenting: what are the round trip flights between chicago and orlando for the 27th.
Experimental setup: From each dataset increasing number of random utternaces per intent were taken to form the raw training data. The same data was then augmented with parrot paraphraser for Nx times(where N =10 or 15 depending the dataset) to form the augmented training data. Now models are trained on both raw data and augmented data to compare the performance. Being a multiclass classification model weighted F1 was used as a metric. The experiment was repeated 4 times for each number of utterance and F1 has been averaged to remove randomness in the trend. I have used 6 prominent NLU datasets from across domains. Below charts reveal that with a "very modest number" utterances and paraphrase augmentation we can achieve good classfication performance on day 1. "Very modest" varies between 4 to 6 utterances per intent in some datasets and 5 to 7 for some datasets.
TBD
TBD
TBD
- The diversity scores are not normalised each of the diversity rankers scores paraphrases differently
- Some command style input phrases generate less adequate paraphrases
TBD
To cite Parrot in your work, please use the following bibtex reference:
@misc{prithivida2021parrot,
author = {Prithiviraj Damodaran},
title = {Parrot: Paraphrase generation for NLU.},
year = 2021,
version = {v1.0}
}