YAML Metadata Warning:The pipeline tag "text2text-generation" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other
Model Summary
This is a generative model designed specifically for search query rewriting, employing a sequence-to-sequence architecture for generating reformulated queries. It leverages a Reinforcement Learning framework to further boost performance, integrating a policy gradient algorithm. The model is trained with reward functions aimed at diversifying the generated queries by paraphrasing keywords. It can be integrated with sparse retrieval methods, such as bm25-based retrieval, to enhance document recall in search.
Intended use cases
Query rewriting for search (web, e-commerce), Virtual assistants and chatbots, Information retrieval
Model Description
Training Procedure
- The training process begins by initializing the sequence-to-sequence model with Google's T5-base model .
- Initially, the model undergoes supervised training using the MS-MARCO query pairs dataset
- Subsequently, the model is fine-tuned using a reinforcement learning (RL) framework to enhance its ability to generate queries that are both diverse and relevant.
- It uses a policy gradient approach to fine-tune the model. For a given input query, a set of trajectories (reformulated queries) are sampled from the model and reward is computed. Policy gradient algorithm is applied to update the model.
- Rewards are heuristically computed to enhance the model's paraphrasing capability. However, these rewards can be substituted with other domain-specific or goal-specific reward functions as needed.
Refer here for more details.
Model Sources
How to use
For optimal utilization of this model, use sampling with repetition penalty to generate diverse samples. Below is the provided sample code.
import torch
from transformers import T5ForConditionalGeneration, T5Tokenizer
MODEL_ID = "prhegde/t5-query-reformulation-RL"
tokenizer = T5Tokenizer.from_pretrained(MODEL_ID)
model = T5ForConditionalGeneration.from_pretrained(MODEL_ID)
model.eval()
input_sequence = "how to bake great cookie"
input_ids = tokenizer(input_sequence, return_tensors="pt").input_ids
print(f'Input: {input_sequence}')
nsent = 4
with torch.no_grad():
for i in range(nsent):
output = model.generate(input_ids, max_length=35, num_beams=1, do_sample=True, repetition_penalty=1.8)
target_sequence = tokenizer.decode(output[0], skip_special_tokens=True)
print(f'Target: {target_sequence}')
- Downloads last month
- 20