Intra-Document Cascading (IDCM)

We provide a retrieval trained IDCM model. Our model is trained on MSMARCO-Document with up to 2000 tokens.

This instance can be used to re-rank a candidate set of long documents. The base BERT architecure is a 6-layer DistilBERT.

If you want to know more about our intra document cascading model & training procedure using knowledge distillation check out our paper: https://arxiv.org/abs/2105.09816 🎉

For more information, training data, source code, and a minimal usage example please visit: https://github.com/sebastian-hofstaetter/intra-document-cascade

Configuration

• Trained with fp16 mixed precision
• We select the top 4 windows of size (50 + 2*7 overlap words) with our fast CK model and score them with BERT
• The published code here is only usable for inference (we removed the training code)

Model Code

from transformers import AutoTokenizer,AutoModel, PreTrainedModel,PretrainedConfig
from typing import Dict
import torch
from torch import nn as nn

class IDCM_InferenceOnly(PreTrainedModel):
    '''
    IDCM is a neural re-ranking model for long documents, it creates an intra-document cascade between a fast (CK) and a slow module (BERT_Cat)
    This code is only usable for inference (we removed the training mechanism for simplicity)
    '''

    config_class = IDCM_Config
    base_model_prefix = "bert_model"

    def __init__(self,
                 cfg) -> None:
        super().__init__(cfg)

        #
        # bert - scoring
        #
        if isinstance(cfg.bert_model, str):
            self.bert_model = AutoModel.from_pretrained(cfg.bert_model)
        else:
            self.bert_model = cfg.bert_model

        #
        # final scoring (combination of bert scores)
        #
        self._classification_layer = torch.nn.Linear(self.bert_model.config.hidden_size, 1)
        self.top_k_chunks = cfg.top_k_chunks
        self.top_k_scoring = nn.Parameter(torch.full([1,self.top_k_chunks], 1, dtype=torch.float32, requires_grad=True))

        #
        # local self attention
        #
        self.padding_idx= cfg.padding_idx
        self.chunk_size = cfg.chunk_size
        self.overlap = cfg.overlap
        self.extended_chunk_size = self.chunk_size + 2 * self.overlap

        #
        # sampling stuff
        #
        self.sample_n = cfg.sample_n
        self.sample_context = cfg.sample_context

        if self.sample_context == "ck":
            i = 3
            self.sample_cnn3 = nn.Sequential(
                        nn.ConstantPad1d((0,i - 1), 0),
                        nn.Conv1d(kernel_size=i, in_channels=self.bert_model.config.dim, out_channels=self.bert_model.config.dim),
                        nn.ReLU()
                        ) 
        elif self.sample_context == "ck-small":
            i = 3
            self.sample_projector = nn.Linear(self.bert_model.config.dim,384)
            self.sample_cnn3 = nn.Sequential(
                        nn.ConstantPad1d((0,i - 1), 0),
                        nn.Conv1d(kernel_size=i, in_channels=384, out_channels=128),
                        nn.ReLU()
                        ) 

        self.sampling_binweights = nn.Linear(11, 1, bias=True)
        torch.nn.init.uniform_(self.sampling_binweights.weight, -0.01, 0.01)
        self.kernel_alpha_scaler = nn.Parameter(torch.full([1,1,11], 1, dtype=torch.float32, requires_grad=True))

        self.register_buffer("mu",nn.Parameter(torch.tensor([1.0, 0.9, 0.7, 0.5, 0.3, 0.1, -0.1, -0.3, -0.5, -0.7, -0.9]), requires_grad=False).view(1, 1, 1, -1))
        self.register_buffer("sigma", nn.Parameter(torch.tensor([0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.1]), requires_grad=False).view(1, 1, 1, -1))
        

    def forward(self,
                query: Dict[str, torch.LongTensor],
                document: Dict[str, torch.LongTensor],
                use_fp16:bool = True,
                output_secondary_output: bool = False):

        #
        # patch up documents - local self attention
        #
        document_ids = document["input_ids"][:,1:]
        if document_ids.shape[1] > self.overlap:
            needed_padding = self.extended_chunk_size - (((document_ids.shape[1]) % self.chunk_size)  - self.overlap)
        else:
            needed_padding = self.extended_chunk_size - self.overlap - document_ids.shape[1]
        orig_doc_len = document_ids.shape[1]

        document_ids = nn.functional.pad(document_ids,(self.overlap, needed_padding),value=self.padding_idx)
        chunked_ids = document_ids.unfold(1,self.extended_chunk_size,self.chunk_size)

        batch_size = chunked_ids.shape[0]
        chunk_pieces = chunked_ids.shape[1]


        chunked_ids_unrolled=chunked_ids.reshape(-1,self.extended_chunk_size)
        packed_indices = (chunked_ids_unrolled[:,self.overlap:-self.overlap] != self.padding_idx).any(-1)
        orig_packed_indices = packed_indices.clone()
        ids_packed = chunked_ids_unrolled[packed_indices]
        mask_packed = (ids_packed != self.padding_idx)

        total_chunks=chunked_ids_unrolled.shape[0]

        packed_query_ids = query["input_ids"].unsqueeze(1).expand(-1,chunk_pieces,-1).reshape(-1,query["input_ids"].shape[1])[packed_indices]
        packed_query_mask = query["attention_mask"].unsqueeze(1).expand(-1,chunk_pieces,-1).reshape(-1,query["attention_mask"].shape[1])[packed_indices]

        #
        # sampling
        # 
        if self.sample_n > -1:
            
            #
            # ck learned matches
            #
            if self.sample_context == "ck-small":
                query_ctx = torch.nn.functional.normalize(self.sample_cnn3(self.sample_projector(self.bert_model.embeddings(packed_query_ids).detach()).transpose(1,2)).transpose(1, 2),p=2,dim=-1)
                document_ctx = torch.nn.functional.normalize(self.sample_cnn3(self.sample_projector(self.bert_model.embeddings(ids_packed).detach()).transpose(1,2)).transpose(1, 2),p=2,dim=-1)
            elif self.sample_context == "ck":
                query_ctx = torch.nn.functional.normalize(self.sample_cnn3((self.bert_model.embeddings(packed_query_ids).detach()).transpose(1,2)).transpose(1, 2),p=2,dim=-1)
                document_ctx = torch.nn.functional.normalize(self.sample_cnn3((self.bert_model.embeddings(ids_packed).detach()).transpose(1,2)).transpose(1, 2),p=2,dim=-1)
            else:
                qe = self.tk_projector(self.bert_model.embeddings(packed_query_ids).detach())
                de = self.tk_projector(self.bert_model.embeddings(ids_packed).detach())
                query_ctx = self.tk_contextualizer(qe.transpose(1,0),src_key_padding_mask=~packed_query_mask.bool()).transpose(1,0)
                document_ctx = self.tk_contextualizer(de.transpose(1,0),src_key_padding_mask=~mask_packed.bool()).transpose(1,0)
        
                query_ctx =   torch.nn.functional.normalize(query_ctx,p=2,dim=-1)
                document_ctx= torch.nn.functional.normalize(document_ctx,p=2,dim=-1)

            cosine_matrix = torch.bmm(query_ctx,document_ctx.transpose(-1, -2)).unsqueeze(-1)

            kernel_activations = torch.exp(- torch.pow(cosine_matrix - self.mu, 2) / (2 * torch.pow(self.sigma, 2))) * mask_packed.unsqueeze(-1).unsqueeze(1)
            kernel_res = torch.log(torch.clamp(torch.sum(kernel_activations, 2) * self.kernel_alpha_scaler, min=1e-4)) * packed_query_mask.unsqueeze(-1)
            packed_patch_scores = self.sampling_binweights(torch.sum(kernel_res, 1))

            
            sampling_scores_per_doc = torch.zeros((total_chunks,1), dtype=packed_patch_scores.dtype, layout=packed_patch_scores.layout, device=packed_patch_scores.device)
            sampling_scores_per_doc[packed_indices] = packed_patch_scores
            sampling_scores_per_doc = sampling_scores_per_doc.reshape(batch_size,-1,)
            sampling_scores_per_doc_orig = sampling_scores_per_doc.clone()
            sampling_scores_per_doc[sampling_scores_per_doc == 0] = -9000

            sampling_sorted = sampling_scores_per_doc.sort(descending=True)
            sampled_indices = sampling_sorted.indices + torch.arange(0,sampling_scores_per_doc.shape[0]*sampling_scores_per_doc.shape[1],sampling_scores_per_doc.shape[1],device=sampling_scores_per_doc.device).unsqueeze(-1)

            sampled_indices = sampled_indices[:,:self.sample_n]
            sampled_indices_mask = torch.zeros_like(packed_indices).scatter(0, sampled_indices.reshape(-1), 1)

            # pack indices

            packed_indices = sampled_indices_mask * packed_indices
    
            packed_query_ids = query["input_ids"].unsqueeze(1).expand(-1,chunk_pieces,-1).reshape(-1,query["input_ids"].shape[1])[packed_indices]
            packed_query_mask = query["attention_mask"].unsqueeze(1).expand(-1,chunk_pieces,-1).reshape(-1,query["attention_mask"].shape[1])[packed_indices]

            ids_packed = chunked_ids_unrolled[packed_indices]
            mask_packed = (ids_packed != self.padding_idx)

        #
        # expensive bert scores
        #
        
        bert_vecs = self.forward_representation(torch.cat([packed_query_ids,ids_packed],dim=1),torch.cat([packed_query_mask,mask_packed],dim=1))
        packed_patch_scores = self._classification_layer(bert_vecs) 

        scores_per_doc = torch.zeros((total_chunks,1), dtype=packed_patch_scores.dtype, layout=packed_patch_scores.layout, device=packed_patch_scores.device)
        scores_per_doc[packed_indices] = packed_patch_scores
        scores_per_doc = scores_per_doc.reshape(batch_size,-1,)
        scores_per_doc_orig = scores_per_doc.clone()
        scores_per_doc_orig_sorter = scores_per_doc.clone()

        if self.sample_n > -1:
            scores_per_doc = scores_per_doc * sampled_indices_mask.view(batch_size,-1)
        
        #
        # aggregate bert scores
        #

        if scores_per_doc.shape[1] < self.top_k_chunks:
            scores_per_doc = nn.functional.pad(scores_per_doc,(0, self.top_k_chunks - scores_per_doc.shape[1]))

        scores_per_doc[scores_per_doc == 0] = -9000
        scores_per_doc_orig_sorter[scores_per_doc_orig_sorter == 0] = -9000
        score = torch.sort(scores_per_doc,descending=True,dim=-1).values
        score[score <= -8900] = 0

        score = (score[:,:self.top_k_chunks] * self.top_k_scoring).sum(dim=1)

        if self.sample_n == -1:
            if output_secondary_output:
                return score,{
                    "packed_indices": orig_packed_indices.view(batch_size,-1),
                    "bert_scores":scores_per_doc_orig
                }
            else:
                return score,scores_per_doc_orig    
        else:
            if output_secondary_output:
                return score,scores_per_doc_orig,{
                    "score": score,
                    "packed_indices": orig_packed_indices.view(batch_size,-1),
                    "sampling_scores":sampling_scores_per_doc_orig,
                    "bert_scores":scores_per_doc_orig
                }

            return score

    def forward_representation(self, ids,mask,type_ids=None) -> Dict[str, torch.Tensor]:
        
        if self.bert_model.base_model_prefix == 'distilbert': # diff input / output 
            pooled = self.bert_model(input_ids=ids,
                                     attention_mask=mask)[0][:,0,:]
        elif self.bert_model.base_model_prefix == 'longformer':
            _, pooled = self.bert_model(input_ids=ids,
                                        attention_mask=mask.long(),
                                        global_attention_mask = ((1-ids)*mask).long())
        elif self.bert_model.base_model_prefix == 'roberta': # no token type ids
            _, pooled = self.bert_model(input_ids=ids,
                                        attention_mask=mask)
        else:
            _, pooled = self.bert_model(input_ids=ids,
                                        token_type_ids=type_ids,
                                        attention_mask=mask)

        return pooled

tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased") # honestly not sure if that is the best way to go, but it works :)
model = IDCM_InferenceOnly.from_pretrained("sebastian-hofstaetter/idcm-distilbert-msmarco_doc")

Effectiveness on MSMARCO Passage & TREC Deep Learning '19

We trained our model on the MSMARCO-Document collection. We trained the selection module CK with knowledge distillation from the stronger BERT model.

For re-ranking we used the top-100 BM25 results. The throughput of IDCM should be ~600 documents with max 2000 tokens per second.

MSMARCO-Document-DEV

TREC-DL'19 (Document Task)

For MRR we use the recommended binarization point of the graded relevance of 2. This might skew the results when compared to other binarization point numbers.

For more metrics, baselines, info and analysis, please see the paper: https://arxiv.org/abs/2105.09816

Limitations & Bias

• The model inherits social biases from both DistilBERT and MSMARCO.

• The model is only trained on longer documents of MSMARCO, so it might struggle with especially short document text - for short text we recommend one of our MSMARCO-Passage trained models.

Citation

If you use our model checkpoint please cite our work as:

@inproceedings{Hofstaetter2021_idcm,
    author = {Sebastian Hofst{\"a}tter and Bhaskar Mitra and Hamed Zamani and Nick Craswell and Allan Hanbury},
    title = {{Intra-Document Cascading: Learning to Select Passages for Neural Document Ranking}},
    booktitle = {Proc. of SIGIR},
    year = {2021},
}