SEQ2SEQ-SC END-TO-END SEMANTIC COMMUNICATION SYSTEMS WITH PRE-TRAINED LANGUAGE MODEL Ju-Hyung LeeDong-Ho LeeEunsoo Sheen Thomas Choi Jay Pujara

2025-05-03 0 0 589.1KB 5 页 10玖币

侵权投诉

SEQ2SEQ-SC: END-TO-END SEMANTIC COMMUNICATION SYSTEMS WITH

PRE-TRAINED LANGUAGE MODEL

Ju-Hyung Lee∗Dong-Ho Lee∗Eunsoo Sheen Thomas Choi Jay Pujara

University of Southern California

ABSTRACT

In this work, we propose a realistic semantic network called

seq2seq-SC, designed to be compatible with 5G NR and

capable of working with generalized text datasets using a

pre-trained language model. The goal is to achieve unprece-

dented communication efﬁciency by focusing on the mean-

ing of messages in semantic communication. We employ

a performance metric called semantic similarity, measured

by BLEU for lexical similarity and SBERT for semantic

similarity. Our ﬁndings demonstrate that seq2seq-SC outper-

forms previous models in extracting semantically meaningful

information while maintaining superior performance. This

study paves the way for continued advancements in semantic

communication and its prospective incorporation with future

wireless systems in 6G networks.

Index Terms—Semantic communication, natural lan-

guage processing (NLP), link-level simulation.

I. INTRODUCTION

The recent rise of deep learning-based techniques to infer

semantics (i.e., the meaning of the message rather than the

message itself) from texts, speeches, and videos, as well

as the ever-increasing quality of service requirements for

extremely data-hungry applications such as extended reality

(XR), have motivated the use of semantic communication [1]

for a new generation of wireless systems (6G). Focusing on

semantics allows forgoing unnecessary data (e.g., articles in

a sentence or background in a portrait photo), which can

increase communication efﬁciency.

While semantic communication may bring unprecedented

beneﬁts, many challenges remain to realize it for actual

usage. First, it must be compatible with existing commu-

nication infrastructure; a “link-level” simulation is hence

required to verify its realistic end-to-end (E2E) performance.

Second, the semantic network has to be generalized to work

with any dataset rather than a particular dataset. Third,

since there is no universal performance metric for semantic

communication yet, metrics such as semantic similarity

must be reﬁned to evaluate performance of the semantic

network from the semantic point of view. Lastly, since classic

communication cannot be completely replaced by semantic

communication, the network must be able to deliver infor-

mation both as-is or modiﬁed with high semantic similarity

dependent upon the communication scenario.

Contributions. We revisit questions raised by DeepSC [2]

regarding semantic communication:

Q 1: How do we design the semantic and channel coding

jointly?

Q 2: How do we measure semantic error (similarity) be-

tween transmitted and received sentences?

Our main contributions, which address these questions, are

summarized as follows:

•We employ E2E link-level simulation compliant to

5G NR (NVIDIA Sionna [3]), which contains features

like Polar codes. Through such method, we validate

semantic network performance in real-world settings

and answer Q 1.

•We integrate the pre-trained encoder-decoder transform-

ers with the E2E semantic communication systems,

which efﬁciently extract the semantic (meaningful)

information with reduced computation effort, dubbed

seq2seq-SC. This network is “generalized”, meaning

it works with all (general) text corpus in comparison

to Deep-SC which has limitations dependent upon a

particular datasets.

•To answer Q 2 and evaluate performance of a semantic

network in a semantic way, we introduce a metric

called semantic similarity. In order to make the network

ﬂexible with respect to the communication scenario, the

network may either prioritize delivering a message as

perfect as possible or on semantic similarity only.

II. PRE-TRAINED MODEL FOR LANGUAGE

Contextualized embeddings from pre-training in a self-

supervised manner (e.g., masked language modeling) with

transformers [4] are extremely effective in providing ini-

tial representations that can be reﬁned to attain acceptable

performance on numerous downstream tasks. Recent studies

on text semantic communication exploit transformer archi-

tectures [4] to extract semantics at transmitter and recover

the original information at receiver [2], [5]. However, such

frameworks may have following challenges: (1) Training

an E2E semantic communication pipeline requires a huge

computational effort due to the many randomly initialized

parameters of semantic encoder/decoder to be trained on; (2)

Difﬁculty in handling out-of-vocabulary (OOV) since they

only use a set of whitespace-separated tokens in the training

data. In this work, we use a pre-trained encoder-decoder

arXiv:2210.15237v2 [eess.SP] 18 Oct 2023

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

SEQ2SEQ-SC:END-TO-ENDSEMANTICCOMMUNICATIONSYSTEMSWITHPRE-TRAINEDLANGUAGEMODELJu-HyungLee∗Dong-HoLee∗EunsooSheenThomasChoiJayPujaraUniversityofSouthernCaliforniaABSTRACTInthiswork,weproposearealisticsemanticnetworkcalledseq2seq-SC,designedtobecompatiblewith5GNRandcapableofworkingwithgeneralizedtextda...

展开>> 收起<<

SEQ2SEQ-SC END-TO-END SEMANTIC COMMUNICATION SYSTEMS WITH PRE-TRAINED LANGUAGE MODEL Ju-Hyung LeeDong-Ho LeeEunsoo Sheen Thomas Choi Jay Pujara.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

SEQ2SEQ-SC END-TO-END SEMANTIC COMMUNICATION SYSTEMS WITH PRE-TRAINED LANGUAGE MODEL Ju-Hyung LeeDong-Ho LeeEunsoo Sheen Thomas Choi Jay Pujara

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: