seq2seq_constrained_decoding

This project includes constrained-decoding utilities for structured text generation using Huggingface seq2seq models.

Requirements

pip install torch transformers

The package is tested with transformers==4.15.0, might break for past or future versions.

Use cases

Several use-cases leverage pretrained sequence-to-sequence models, such as BART or T5, for generating a (maybe partially) structured text sequence. For example, you may want to finetine the model to generate a set of key-points summarizing an input paragraph, or to produce a text sequence that follows some strict regularities.

Below we detail about the use-cases supported by this project.

Word lists

The word_list_decoding.py module suggests simple LogitsProcessor classes which enforce allowed / forbidden words during text generation (see the WhiteListLogitsProcessor and BlackListLogitsProcessor classes respectively).

Example usage:

from transformers import LogitsProcessorList, AutoTokenizer, AutoModel
from constrained_decoding.word_list_decoding import BlackListLogitsProcessor
tokenizer = AutoTokenizer.from_pretrained('t5-small')
model = AutoModel.from_pretrained('t5-small')

bad_words = "here are words that should not occur in the generated text"
bad_word_ids = tokenizer.encode(bad_words)
black_list_processor = BlackListLogitsProcessor(bad_word_ids)
good_words = "only these words can occur in the generated text"
good_word_ids = tokenizer.encode(good_words)
white_list_processor = WhiteListLogitsProcessor(good_word_ids)

input_seq = "here are the input words to condition generated text upon"
input_ids = tokenizer.encode(input_seq, return_tensors='pt')
out = model.generate(input_ids, num_beams=10)
print(tokenizer.batch_decode(out))
# ['<pad> Hier are the input words to condition generated text upon</s>']
out = model.generate(input_ids, num_beams=10, logits_processor=[black_list_processor])
print(tokenizer.batch_decode(out))
# ['<pad> Voici voici les input mots to condition a condition a condition a condition a']
out = model.generate(input_ids, num_beams=10, logits_processor=[white_list_processor])
print(tokenizer.batch_decode(out))
# ['<pad> in the words in the words in the words in the words in the words in the words in']

Set decoding

In some scenarios, you would like to regard the output sequence as expressing a set of elements comprised of sub-sequences. For example, you might finetune your Seq2Seq model on a multi-label document classification task (e.g. generating the set of relation types occuring in the input document).

The set_decoding.SetDecodingLogitProcessor class can gurantee that no subsequence (e.g. a relation type) would occur more than once. Output subsequences are assumed to be defined using a special separator token.

DFA-based constrained decoding

The most powerful and generic constrained decoding algorithm we propose is using a Deterministic Finite Automata (DFA). You can instanciate a DFA with the dfa.DFA class, defined over a dictionary of dictionaries. For example, the following represents an automaton that accepts only binary numbers that are multiples of 3 (see illustration in the Wikipedia article on DFA):

from dfa import DFA
transitions = {0:{'0':0, '1':1},
               1:{'0':2, '1':0},
               2:{'0':1, '1':2}} 
dfa = DFA(transitions, s0=0, accept_states=[0])

For defining constrained decoding using a DFA, the automaton's alphabet should correspond to tokens in the model's vocabulry. The DFA class supports translating a dfa that uses regular words or phrases as alphabet into a tokenizer-adjusted dfa -

transitions = {0:{'John':1, 'Mike':1, 'Dan':1},
               1:{'went':2, 'ran':2, 'jogged':2},
               2:{'to':3, 'in':3},
               3:{'the':4, 'a':4},
               4:{'park':5}} 
words_dfa = DFA(transitions, s0=0, accept_states=[5])

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("t5-small")
tokens_dfa = words_dfa.adjust_for_tokenizer(tokenizer)

Eventually, generation constraining is achieved by replacing the model's beam_search method with an adapted version (in dfa_constrained_beam_search.py) that enforces every beam to follow the automaton transitions. The probability of vocabulary entries that are not accessible according to the dfa will be set to minus infinity.

For example:

from transformers import AutoModel
model = AutoModel.from_pretrained("t5-small")
from dfa_constrained_beam_search import set_model_beam_search_to_dfa_constrained
set_model_beam_search_to_dfa_constrained(model, tokens_dfa)

#   The previous two steps can equivalently be done in one call:
# set_model_beam_search_to_dfa_constrained(model, words_dfa, tokenizer) 

Other supported utility functions within the DFA class include:

• DFA.from_slots - for constructing a linear DFA out of a list of "slots", where each slot is represented by a list of allowed words / phrases.
• DFA.concat_two and DFA.concat - for concatenating two or more (linear) DFAs into a long DFA.
• as_cyclic - for converting a linear DFA into a cyclic one, by merging or connecting some end-states with the initial state.

QA-SRL

Our motivational use-case is seq2seq-based QA-SRL parsing. In that project, we finetune BART/T5 on the Question-Answer driven Semantic Role Labeling task. Given a verb or nominalization in context, the task is to generate Question-Answer pairs capturing the participants of the verbal event.

To model the task using a seq2seq paradigm, the QAs are linearized into a single target sequence, using separators between QA pairs, between question and answer, and between multiple answers. Furthermore, QASRL questions must adhere a slot-based template, while answers could only be continuous spans copied from the input sentence.

Check out the qasrl.py module to see how we leverage the DFA utilities for enforcing a valid QASRL output sequence.