tensorkelechi/arxiv-cstext · Datasets at Hugging Face

text stringlengths 0 89.3k
raw text directly In the situation of textrich graphs the raw text is
usually processed by a text encoder eg a language model LM to
transform it into a feature vector before GNNs can be applied How
ever as evidenced by previous studies 757 separately encoding
text and performing messagepassing is suboptimal in capturing
semantic meanings of the neighborhood context and can result
in inferior performance When textual data volume is large and
contextually rich there is a clear demand for strategies that can
seamlessly integrate the graph structures with the textual features
The exceptional ability of large language models LLMs 32831
3650 in solving language tasks positions them as promising tools
for handling textrich graph problems Their ability to contextualize
text and draw semantic relationships provides potential solutions to
address the challenges where GNNs fall short However integrating
LLMs with textrich graphs presents significant challenges and re
mains underexplored One major challenge is that LLMs originally
designed for onedimensional sequential text face difficulties when
dealing with complex topological structures of graphs Additionally
LLMs are constrained by a token limit that is inadequate for the
dense and abundant text in realworld graphs While advanced
models such as GPT4 1 can handle very long inputs the compu
tational demands are substantial necessitating extensive hardware
resources Commonly available opensource LLMs face limitations
only able to handle a few thousand tokens 3313650 which
is insufficient given the extensive text data contained within the
neighborhoods of actual graphs For example an ecommerce graph
involving sports equipments on Amazon has an average node de
gree of 814 and about 13093 text tokens per node 26 Processing
even a twohop neighborhood quickly becomes computationally
infeasible Therefore adapting LLMs to handle textrich graphs
necessitates a method that can effectively compress neighborhood
text while preserving structural information
To address the above challenges we present Hierarchical Com
pression HiCom enabled by LLMs which can both model graph
structures and manage long text We draw inspiration from the
messagepassing MP algorithm of GNNs 14 as well as the trans
former compression techniques for handling long documents es
pecially AutoCompressor 6 HiCom first constructs a hierarchy
according to the graph structure for processing a large neighbor
hood and then compresses the context level by level to avoid input
explosion In our experiments we apply HiCom to finetune an
OPT 50 model to efficiently capture the semantic meaning of rich
neighborhoods The HiComOPT model outperforms both GNNs
and the LLM backbone for node classification on five textrich
graphs covering academic and ecommerce applications We test
HiCom under both normal settings and also a challenging setting
16oz Glass Tumbler with Straw and Lid Heat Resistant Coffee Milk Smoothie
Plastic Water Bottle with Carry Handle Perfect for Travel Outdoors and Gym Health Fitness Tracker Workout Intensity Sleep Tracking 247 Heart Rate
Kitchen DiningSports OutdoorsWater Bottle 20 oz Glass Tumbler Straw Silicone Protective Sleeve Bamboo Lid Water Bottle Made with Material Derived From 50 Plastic Waste
Bubble Blends Strawberry Popping Boba 22lbs 100 FatFree Figure 1 Category classification of two water bottles in the
middle from the Amazon productcoviewing graph Their
categories in red are not clear solely from the product de
scriptions in green boxes but will more likely be correctly
classified through the neighborhood context
focusing on nodes from a dense region in a graph kcore graph
with densely connected nodes HiCom achieves an 348 average
performance improvement Meanwhile HiCom is more efficient
than vanilla LLMs by breaking long inputs and compressing them
levelbylevel Our ablation studies show that HiCom maintains
its superior performance in both dense and sparse regions in the
graph as well as when the training data is either scarce or sufficient
In summary the advantages of HiCom is multifaceted
Adaptability HiCom crafts LLMs for easy adaptation to graph
inputs by transforming the vast context of a nodes neighbor
hood into a more manageable and structured format
Effectiveness HiCom outperforms both GNNs and vanilla
LLMs for node classification on textrich graphs
Efficiency HiCom introduces memoryefficient implementa
tion techniques that facilitate LLM finetuning on graph data
and makes HiCom more scalable than vanilla LLMs
2 RELATED WORK
Learning on TextRich Graphs Learning on textrich graphs re
quires integration of textual features and graph structures and is
gaining increased attention in graph data mining 720434456
GNNs have shown an exceptional ability to capture structural infor
mation and can transform and aggregate node features according to
the graph structure via message passing 142138 The problem
with applying GNNs on textrich graphs is that they are not inher
ently equipped to handle raw text Therefore GNNs usually need to
be combined with a text encoder to transform the text into vector
representations and the encoder may be combined with GNNs
through a twintower structure 56 nested layers 43 a cascade
structure with joint training 41 or iterative training via varia
tional inference 53 Another approach for learning on textrich
graphs emphasizes pretraining LMs over graph data to preserve
relation information and enrich node representations with graph
structures 7204449 The pretraining objective can be neigh
borhood prediction 7 feature and structure recovery 49 masked
modeling at both tokenlevel and documentlevel 20 or aligning
masked language modeling with knowledge graph reasoning 44Hierarchical Compression of TextRich Graphs via Large Language Models Conference acronym XX June 0305 2018 Woodstock NY
10523674891010
5236748910Hierarchy Construction
𝑥 𝑥 𝑥 𝑥 𝑥 𝑥 𝑥Compressor 𝑠 𝑠Compressor 𝑥 𝑥 𝑥 𝑥 𝑥 𝑠 𝑠Compressor 𝑥 𝑥 𝑥 𝑥 𝑥 𝑠 𝑠 𝑠 𝑠 𝑠 𝑠Hierarchical CompressionPrediction with TheCompressed SummaryText TokensSummary VectorsDummy Tokens 𝑠 𝑥Target NodeOnehop Samples Twohop SamplesFanouts 32 𝑠 𝑠 𝑥 𝑥 𝑥Predictor 𝑦
Figure 2 An illustration of the hierarchical compression framework with LLMs The left part corresponds to the hierarchy
construction step for a target node 𝑣0 with fanouts 32 indicating the budget of nodes to sample in each level The lower
right part shows how the neighborhood context is compressed to a summary vector 𝑠0following the hierarchy The upper right
part shows the final prediction is made with the target node text 𝑥0as input and the summary vector 𝑠0
Transformer Models for Long Inputs Transformers 37 have
shown their effectiveness in various NLP tasks Relevant to our
research is how transformers handle long inputs 48102954

raw text directly In the situation of textrich graphs the raw text is

usually processed by a text encoder eg a language model LM to

transform it into a feature vector before GNNs can be applied How

ever as evidenced by previous studies 757 separately encoding

text and performing messagepassing is suboptimal in capturing

semantic meanings of the neighborhood context and can result

in inferior performance When textual data volume is large and

contextually rich there is a clear demand for strategies that can

seamlessly integrate the graph structures with the textual features

The exceptional ability of large language models LLMs 32831

3650 in solving language tasks positions them as promising tools

for handling textrich graph problems Their ability to contextualize

text and draw semantic relationships provides potential solutions to

address the challenges where GNNs fall short However integrating

LLMs with textrich graphs presents significant challenges and re

mains underexplored One major challenge is that LLMs originally

designed for onedimensional sequential text face difficulties when

dealing with complex topological structures of graphs Additionally

LLMs are constrained by a token limit that is inadequate for the

dense and abundant text in realworld graphs While advanced

models such as GPT4 1 can handle very long inputs the compu

tational demands are substantial necessitating extensive hardware

resources Commonly available opensource LLMs face limitations

only able to handle a few thousand tokens 3313650 which

is insufficient given the extensive text data contained within the

neighborhoods of actual graphs For example an ecommerce graph

involving sports equipments on Amazon has an average node de

gree of 814 and about 13093 text tokens per node 26 Processing

even a twohop neighborhood quickly becomes computationally

infeasible Therefore adapting LLMs to handle textrich graphs

necessitates a method that can effectively compress neighborhood

text while preserving structural information

To address the above challenges we present Hierarchical Com

pression HiCom enabled by LLMs which can both model graph

structures and manage long text We draw inspiration from the

messagepassing MP algorithm of GNNs 14 as well as the trans

former compression techniques for handling long documents es

pecially AutoCompressor 6 HiCom first constructs a hierarchy

according to the graph structure for processing a large neighbor

hood and then compresses the context level by level to avoid input

explosion In our experiments we apply HiCom to finetune an

OPT 50 model to efficiently capture the semantic meaning of rich

neighborhoods The HiComOPT model outperforms both GNNs

and the LLM backbone for node classification on five textrich

graphs covering academic and ecommerce applications We test

HiCom under both normal settings and also a challenging setting

16oz Glass Tumbler with Straw and Lid Heat Resistant Coffee Milk Smoothie

Plastic Water Bottle with Carry Handle Perfect for Travel Outdoors and Gym Health Fitness Tracker Workout Intensity Sleep Tracking 247 Heart Rate

Kitchen DiningSports OutdoorsWater Bottle 20 oz Glass Tumbler Straw Silicone Protective Sleeve Bamboo Lid Water Bottle Made with Material Derived From 50 Plastic Waste

Bubble Blends Strawberry Popping Boba 22lbs 100 FatFree Figure 1 Category classification of two water bottles in the

middle from the Amazon productcoviewing graph Their

categories in red are not clear solely from the product de

scriptions in green boxes but will more likely be correctly

classified through the neighborhood context

focusing on nodes from a dense region in a graph kcore graph

with densely connected nodes HiCom achieves an 348 average

performance improvement Meanwhile HiCom is more efficient

than vanilla LLMs by breaking long inputs and compressing them

levelbylevel Our ablation studies show that HiCom maintains

its superior performance in both dense and sparse regions in the

graph as well as when the training data is either scarce or sufficient

In summary the advantages of HiCom is multifaceted

Adaptability HiCom crafts LLMs for easy adaptation to graph

inputs by transforming the vast context of a nodes neighbor

hood into a more manageable and structured format

Effectiveness HiCom outperforms both GNNs and vanilla

LLMs for node classification on textrich graphs

Efficiency HiCom introduces memoryefficient implementa

tion techniques that facilitate LLM finetuning on graph data

and makes HiCom more scalable than vanilla LLMs

2 RELATED WORK

Learning on TextRich Graphs Learning on textrich graphs re

quires integration of textual features and graph structures and is

gaining increased attention in graph data mining 720434456

GNNs have shown an exceptional ability to capture structural infor

mation and can transform and aggregate node features according to

the graph structure via message passing 142138 The problem

with applying GNNs on textrich graphs is that they are not inher

ently equipped to handle raw text Therefore GNNs usually need to

be combined with a text encoder to transform the text into vector

representations and the encoder may be combined with GNNs

through a twintower structure 56 nested layers 43 a cascade

structure with joint training 41 or iterative training via varia

tional inference 53 Another approach for learning on textrich

graphs emphasizes pretraining LMs over graph data to preserve

relation information and enrich node representations with graph

structures 7204449 The pretraining objective can be neigh

borhood prediction 7 feature and structure recovery 49 masked

modeling at both tokenlevel and documentlevel 20 or aligning

masked language modeling with knowledge graph reasoning 44Hierarchical Compression of TextRich Graphs via Large Language Models Conference acronym XX June 0305 2018 Woodstock NY

10523674891010

5236748910Hierarchy Construction

𝑥 𝑥 𝑥 𝑥 𝑥 𝑥 𝑥Compressor 𝑠 𝑠Compressor 𝑥 𝑥 𝑥 𝑥 𝑥 𝑠 𝑠Compressor 𝑥 𝑥 𝑥 𝑥 𝑥 𝑠 𝑠 𝑠 𝑠 𝑠 𝑠Hierarchical CompressionPrediction with TheCompressed SummaryText TokensSummary VectorsDummy Tokens 𝑠 𝑥Target NodeOnehop Samples Twohop SamplesFanouts 32 𝑠 𝑠 𝑥 𝑥 𝑥Predictor 𝑦

Figure 2 An illustration of the hierarchical compression framework with LLMs The left part corresponds to the hierarchy

construction step for a target node 𝑣0 with fanouts 32 indicating the budget of nodes to sample in each level The lower

right part shows how the neighborhood context is compressed to a summary vector 𝑠0following the hierarchy The upper right

part shows the final prediction is made with the target node text 𝑥0as input and the summary vector 𝑠0

Transformer Models for Long Inputs Transformers 37 have

shown their effectiveness in various NLP tasks Relevant to our

research is how transformers handle long inputs 48102954