tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
55 Zexuan Zhong Dan Friedman and Danqi Chen 2021 Factual probing is mask
Learning vs learning to recall arXiv preprint arXiv210405240 2021
56 Jason Zhu Yanling Cui Yuming Liu Hao Sun Xue Li Markus Pelger Tianqi
Yang Liangjie Zhang Ruofei Zhang and Huasha Zhao 2021 Textgnn Improving
text encoder via graph neural network in sponsored search In Proceedings of
theWeb Conference 2021 28482857
57 Jing Zhu Xiang Song Vassilis N Ioannidis Danai Koutra and Christos Falout
sos 2023 TouchUpG Improving feature representation through graphcentric
finetuning arXiv preprint arXiv230913885 2023
A APPENDIX
A1 The Batch Processing Algorithm
In typical NLP tasks on sequential data discrepancies in the in
put length of batched instances are commonly managed through
padding with dummy tokens However in the context of graph data
this complexity is magnified due to the variability in the numberHierarchical Compression of TextRich Graphs via Large Language Models Conference acronym XX June 0305 2018 Woodstock NY
Algorithm 3 Rearrange and Trim Batched Token Sequences
1Input Token sequences 𝑿 attention masks 𝑴
2Output Compact token sequences 𝑿 Compact attention
masks 𝑴
3𝐵𝑿shape0 batch size
4𝑇 max 𝑴sumdim1 maximum number of nondummy
tokens per sequence
5𝑿 A placeholder of all zeros with shape 𝐵𝑇
6𝑴 A placeholder of all zeros with shape 𝐵𝑇
7for𝑖1to𝐵do
8𝑥𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝒙𝑖𝑚𝑖bool rearrange
9𝑇𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑥𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 shape0
10 𝑿i 𝑇𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑥𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 trim
11 𝑴i 𝑇𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 1
12end for
13return 𝑿𝑴
of neighbors each node possesses Although the predefined fanout
parameters aim to sample equal numbers of neighbors for each
node they actually only set upper bounds because some nodes
may not have enough neighbors to reach the target fanout count
To address these challenges a more sophisticated padding strat
egy is needed to uniform both the length of the text sequence on
each node and the number of neighbors of each node for all nodes
in the same level of the hierarchy As illustrated in Figure 2 text
sequences of nodes are padded with dummy tokens to the same
length eg 𝑥1𝑥5𝑥8 and𝑥9are padded to length three and the
sampled neighbors for each node are padded to match the fanout
of the corresponding level eg 𝑣2is padded to match fanout3 for
level one
Algorithm 2 outlines HiCom for processing nodes in batches
to process them efficiently For a batch of target nodes V𝐵V
indexed by𝐵 lines 3 to 6 correspond to the hierarchy construction
phase In particular line 4 samples all the nodes V𝐵𝑙that form
the level𝑙forV𝐵 and the nodes from the output level 𝐿equal
toV𝐵 Line 5 gathers the node text and pads them to the same
length𝑡 Note that unlike the neighbor padding that needs to
be performed dynamically individual node text padding can be
performed beforehand during data preprocessing to save time
which is what we do in practice In the pseudocode we explicitly
list out the padding step for completeness Lines 9 to 15 show
the compression step for level 𝑙 To implement neighbor padding
we first initialize placeholders 𝑿𝐵𝐶
𝑙and𝑺𝐵𝐶
𝑙 which are filled with
dummy tokens and have shapes according to the fanouts parameters
line 9 and line 10 Then we insert text and summary vectors
corresponding to nodes from the previous level ie 𝑿𝐵𝑙1and
𝑺𝐵𝑙1 into the placeholder following node index 𝐵𝑙1line 11 and
line 12 Finally the populated placeholders will be reshaped which
effectively concatenates text and summary vectors for nodes in the
same neighborhood and they are then fed to the compressor to get
the summary vectors 𝑺𝐵𝑙for nodes inV𝐵𝑙line 13 to line 15
A2 The Rearrange and Trim Algorithm
We show the pseudocode of the rearrange and trim function for
batched token sequences in Algorithm 3A3 Detailed Experiment Settings
Hyperparameters Since different graphs emphasize the impor
tance of neighbors from different hops we consider several cases
for each method For HiCom and GNN we consider three different
fanouts number of nodes to sample from neighboring nodes in
each hop 44 28 and 82 and report the best result for each
method in Table 3 For NC since the total input length of the LLM
is limited eg 2048 for OPT the LLM can usually only fit the text
on the target node plus text from a few neighbors We consider NC
by sampling 3 neighbors from the 2hop neighborhood of the target
node The full results of all settings with different hyperparameters
are shown in Table 6
Implementation and Hardware We implement our method and
LM baselines with the Hugging Face platform 40 We train these
models for 20 to 60 epochs depending on the graph size until
convergence on 8 NVIDIA A10 GPUs with 24G memory each We
perform LoRA 15 and gradient checkpointing to optimize GPU
memory We use DeepSpeed 30 to perform distributed training
A4 Detailed Experiment Results
In Table 6 we report the detailed results for HiCom and baselines
with different hyperparameters The best result for each method is
reported in Section 5
A5 Experiments for Nodes in All Regions
In Section 55 we discussed the results of an ablation study on
nodes in all regions We now describe the detailed experiment set
ting and provide the detailed results For the experiment setting
we consider a different way to split the datasets compared to the
main experiment We select all the training validation and test
ing nodes are selected randomly from the graph with the size of
each set staying the same as the main experiment ie 20 nodes
per class for training 1000 for validation and up to 10000 for

55 Zexuan Zhong Dan Friedman and Danqi Chen 2021 Factual probing is mask

Learning vs learning to recall arXiv preprint arXiv210405240 2021

56 Jason Zhu Yanling Cui Yuming Liu Hao Sun Xue Li Markus Pelger Tianqi

Yang Liangjie Zhang Ruofei Zhang and Huasha Zhao 2021 Textgnn Improving

text encoder via graph neural network in sponsored search In Proceedings of

theWeb Conference 2021 28482857

57 Jing Zhu Xiang Song Vassilis N Ioannidis Danai Koutra and Christos Falout

sos 2023 TouchUpG Improving feature representation through graphcentric

finetuning arXiv preprint arXiv230913885 2023

A APPENDIX

A1 The Batch Processing Algorithm

In typical NLP tasks on sequential data discrepancies in the in

put length of batched instances are commonly managed through

padding with dummy tokens However in the context of graph data

this complexity is magnified due to the variability in the numberHierarchical Compression of TextRich Graphs via Large Language Models Conference acronym XX June 0305 2018 Woodstock NY

Algorithm 3 Rearrange and Trim Batched Token Sequences

1Input Token sequences 𝑿 attention masks 𝑴

2Output Compact token sequences 𝑿 Compact attention

masks 𝑴

3𝐵𝑿shape0 batch size

4𝑇 max 𝑴sumdim1 maximum number of nondummy

tokens per sequence

5𝑿 A placeholder of all zeros with shape 𝐵𝑇

6𝑴 A placeholder of all zeros with shape 𝐵𝑇

7for𝑖1to𝐵do

8𝑥𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝒙𝑖𝑚𝑖bool rearrange

9𝑇𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑥𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 shape0

10 𝑿i 𝑇𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 𝑥𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 trim

11 𝑴i 𝑇𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 1

12end for

13return 𝑿𝑴

of neighbors each node possesses Although the predefined fanout

parameters aim to sample equal numbers of neighbors for each

node they actually only set upper bounds because some nodes

may not have enough neighbors to reach the target fanout count

To address these challenges a more sophisticated padding strat

egy is needed to uniform both the length of the text sequence on

each node and the number of neighbors of each node for all nodes

in the same level of the hierarchy As illustrated in Figure 2 text

sequences of nodes are padded with dummy tokens to the same

length eg 𝑥1𝑥5𝑥8 and𝑥9are padded to length three and the

sampled neighbors for each node are padded to match the fanout

of the corresponding level eg 𝑣2is padded to match fanout3 for

level one

Algorithm 2 outlines HiCom for processing nodes in batches

to process them efficiently For a batch of target nodes V𝐵V

indexed by𝐵 lines 3 to 6 correspond to the hierarchy construction

phase In particular line 4 samples all the nodes V𝐵𝑙that form

the level𝑙forV𝐵 and the nodes from the output level 𝐿equal

toV𝐵 Line 5 gathers the node text and pads them to the same

length𝑡 Note that unlike the neighbor padding that needs to

be performed dynamically individual node text padding can be

performed beforehand during data preprocessing to save time

which is what we do in practice In the pseudocode we explicitly

list out the padding step for completeness Lines 9 to 15 show

the compression step for level 𝑙 To implement neighbor padding

we first initialize placeholders 𝑿𝐵𝐶

𝑙and𝑺𝐵𝐶

𝑙 which are filled with

dummy tokens and have shapes according to the fanouts parameters

line 9 and line 10 Then we insert text and summary vectors

corresponding to nodes from the previous level ie 𝑿𝐵𝑙1and

𝑺𝐵𝑙1 into the placeholder following node index 𝐵𝑙1line 11 and

line 12 Finally the populated placeholders will be reshaped which

effectively concatenates text and summary vectors for nodes in the

same neighborhood and they are then fed to the compressor to get

the summary vectors 𝑺𝐵𝑙for nodes inV𝐵𝑙line 13 to line 15

A2 The Rearrange and Trim Algorithm

We show the pseudocode of the rearrange and trim function for

batched token sequences in Algorithm 3A3 Detailed Experiment Settings

Hyperparameters Since different graphs emphasize the impor

tance of neighbors from different hops we consider several cases

for each method For HiCom and GNN we consider three different

fanouts number of nodes to sample from neighboring nodes in

each hop 44 28 and 82 and report the best result for each

method in Table 3 For NC since the total input length of the LLM

is limited eg 2048 for OPT the LLM can usually only fit the text

on the target node plus text from a few neighbors We consider NC

by sampling 3 neighbors from the 2hop neighborhood of the target

node The full results of all settings with different hyperparameters

are shown in Table 6

Implementation and Hardware We implement our method and

LM baselines with the Hugging Face platform 40 We train these

models for 20 to 60 epochs depending on the graph size until

convergence on 8 NVIDIA A10 GPUs with 24G memory each We

perform LoRA 15 and gradient checkpointing to optimize GPU

memory We use DeepSpeed 30 to perform distributed training

A4 Detailed Experiment Results

In Table 6 we report the detailed results for HiCom and baselines

with different hyperparameters The best result for each method is

reported in Section 5

A5 Experiments for Nodes in All Regions

In Section 55 we discussed the results of an ablation study on

nodes in all regions We now describe the detailed experiment set

ting and provide the detailed results For the experiment setting

we consider a different way to split the datasets compared to the

main experiment We select all the training validation and test

ing nodes are selected randomly from the graph with the size of

each set staying the same as the main experiment ie 20 nodes

per class for training 1000 for validation and up to 10000 for