Document expansion: You generate for your paragraphs 20-40 queries and index the paragraphs and the generates queries in a standard BM25 index like Elasticsearch, OpenSearch, or Lucene. The generated queries help to close the lexical gap of lexical search, as the generate queries contain synonyms. Further, it re-weights words giving important words a higher weight even if they appear seldomn in a paragraph. In our BEIR paper we showed that BM25+docT5query is a powerful search engine. In the BEIR repository we have an example how to use docT5query with Pyserini.
Domain Specific Training Data Generation: It can be used to generate training data to learn an embedding model. In our GPL-Paper / GPL Example on SBERT.net we have an example how to use the model to generate (query, text) pairs for a given collection of unlabeled texts. These pairs can then be used to train powerful dense embedding models.

Usage

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

model_name = 'doc2query/msmarco-german-mt5-base-v1'
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

text = "Python ist eine universelle, üblicherweise interpretierte, höhere Programmiersprache. Sie hat den Anspruch, einen gut lesbaren, knappen Programmierstil zu fördern. So werden beispielsweise Blöcke nicht durch geschweifte Klammern, sondern durch Einrückungen strukturiert."


def create_queries(para):
    input_ids = tokenizer.encode(para, return_tensors='pt')
    with torch.no_grad():
        # Here we use top_k / top_k random sampling. It generates more diverse queries, but of lower quality
        sampling_outputs = model.generate(
            input_ids=input_ids,
            max_length=64,
            do_sample=True,
            top_p=0.95,
            top_k=10, 
            num_return_sequences=5
            )
        
        # Here we use Beam-search. It generates better quality queries, but with less diversity
        beam_outputs = model.generate(
            input_ids=input_ids, 
            max_length=64, 
            num_beams=5, 
            no_repeat_ngram_size=2, 
            num_return_sequences=5, 
            early_stopping=True
        )


    print("Paragraph:")
    print(para)
    
    print("\nBeam Outputs:")
    for i in range(len(beam_outputs)):
        query = tokenizer.decode(beam_outputs[i], skip_special_tokens=True)
        print(f'{i + 1}: {query}')

    print("\nSampling Outputs:")
    for i in range(len(sampling_outputs)):
        query = tokenizer.decode(sampling_outputs[i], skip_special_tokens=True)
        print(f'{i + 1}: {query}')

create_queries(text)

Note: model.generate() is non-deterministic for top_k/top_n sampling. It produces different queries each time you run it.

Training

This model fine-tuned google/mt5-base for 66k training steps (4 epochs on the 500k training pairs from MS MARCO). For the training script, see the train_script.py in this repository.

The input-text was truncated to 320 word pieces. Output text was generated up to 64 word pieces.

This model was trained on a (query, passage) from the mMARCO dataset.