text
stringlengths
0
89.3k
For this line of research the main challenge is the high complexity
and memory demands of attention computation which limits the
input context length of transformers This issue is particularly pro
nounced when applying transformers to textrich graphs where
the context to encode in a neighborhood often far exceeds the
model limit Several methods have been developed for encoding
long sequential inputs Examples include limiting the attention
window 810 using recurrent memory 4 and employing ap
proximate computation of attention 92954 The most relevant
to this work is AutoCompressor 6 where a long sequential input
is segmented into smaller pieces for stepbystep encoding We
review its technical details in Section 3 A detailed survey suggests
more methods 35 While these studies are relevant from the model
perspective they do not apply to graph data Our work bridges this
gap enhancing the ability of transformerbased models to process
rich text with graph structures
Foundation LLMs and Graphs Foundation LLMs have shown re
markable versatility and adaptability 2 A defining characteristic
of these models is their ability to tackle a wide array of tasks through
prompting and incontext learning ICL without the need for task
specific training Several works have tried to leverage these foun
dation LLMs for solving graph tasks through ICL The ways graph
data is handled include using LLMs to summarize node sampled
from a neighborhood 5 using pure natural language to describe
each nodes multihop connectivity and metafeatures 123946
expanding LLMs vocabulary by creating a new token for every
node 12 and instruction tuning via multiple graph tasks 34
These explorations are promising but are not directly comparable
to this work Our work aims to provide scalable finetuned LLMs
that can be deployed for various textrich graphs without concerns
of privacy or inference cost In contrast many of these works rely
on huge blackbox LLMs and query those models via API which
causes concerns about scalability inference cost and privacy3 NOTATIONS AND PRELIMINARIES
This section establishes a common ground in terms of notation and
foundational concepts before diving into the core methodologies
and results of our study We summarize the notations that will
be used throughout the paper in Table 1 and discuss them in the
following In this work a graph is denoted as GV𝑿 where
Vrepresents the set of nodes and 𝑿represents the text features
on nodes Each node 𝑣𝑖V is associated with a text feature rep
resented in 𝑥𝑖𝑿 We use superscripts to index tokens of text
sequence ie 𝑥𝑖𝑥1
𝑖𝑥𝑡
𝑖 For the node classification task the
goal is to predict the label 𝑦𝑖of a node𝑣𝑖based on its feature and
its neighborhood context As we progress through the paper fur
ther notations will be introduced as required to provide clarity on
specific algorithms and methodologies
Soft Prompts A prompt for LLMs refers to a set of instructions
or questions designed to elicit responses or actions from the model
usually in terms of natural text and is concatenated with the input
text data Soft prompting refers to a technique where instead of
providing textual prompts learnable embeddings are provided as
soft prompts to guide the models response 232455 Similarly
the soft prompts are often concatenated with the tokenized input
text data but they are learnable and will be optimized for the best
model response while freezing the LLM parameters
AutoCompressor 6To make LLMs handle long inputs Auto
Compressor applies soft prompts and encodes a long input by break
ing it into shorter pieces AutoCompressor incorporates a series
of new learnable tokens 𝑝1𝑝𝑘as soft prompts into LLMs
existing vocabulary These tokens when appended to the input
text instruct the LLM to output summaries in a vector format For
instance the input data 𝑥1𝑥𝑡is appended to include the
special tokens resulting in 𝑥1𝑥𝑡𝑝1𝑝𝑘 This augmented
input is processed by the LLM to update the representation of all
tokens including the special tokens The updated representations
of𝑝1𝑝𝑘 denoted as 𝒔𝑠1𝑠𝑘 act as the summaryConference acronym XX June 0305 2018 Woodstock NY Zhang et al
Table 1 Notations
Notations Descriptions
GV𝑿A graph with text nodes Vand text features 𝑿
𝑣𝑖V Nodes inG
𝑥𝑖𝑿 Raw text on node 𝑣𝑖 with𝑥𝑗
𝑖stands for the 𝑗th token
𝒔𝑖 Summary vectors of 𝑣𝑖s neighborhood context
𝑝 The soft prompt with 𝑝𝑗stands for the 𝑗th token
𝐶𝑜𝑚𝑝 The LLMbased compressor that maps text and optional
lowerlevel summary vectors to summary vectors
𝐿 Number of hierarchy levels
𝑛1𝑛𝐿Budget of nodes to sample for each level fanouts
𝑿𝐶𝑿𝐶 Concatenated text 𝑿𝐶means reshaped
𝑺𝐶𝑺𝐶 Concatenated summary vectors 𝑺𝐶means reshaped
vectors These summary vectors distill the meaning from the origi
nal text input while substantially shortening it to length 𝑘 where
𝑘𝑡 Given such length reduction a long sequential input can be
segmented into smaller pieces and compressed to length 𝑘pieces
step by step The learning of these soft prompt tokens slightly alters
the standard decoderbased LLMs nexttoken prediction objective
A standard objective focuses on maximizing the log probability
of the next token conditioned on the previously observed tokens
ielog𝑝𝑥𝑡𝑖𝑥1𝑥𝑡𝑖1 AutoCompressor soft prompt learn
ing involves substituting context tokens with summary vectors
ie changing the objective to log𝑝𝑥𝑡𝑖𝑠1𝑠𝑘𝑥𝑡1𝑥𝑡𝑖1
where the original context 𝑥1𝑥𝑡is replaced by𝑠1𝑠𝑘
4 METHOD
Hierarchical compression HiCom is a framework designed to inte
grate LLMs with textrich graphs Its motivation lies in leveraging
the strong textunderstanding ability of LLMs to extract contex
tual information from an extensive neighborhood The primary
challenge is to process the complex graph structure and the abun