The Selection of Relevance: Content-Selection Strategies in Semantic and Task-Oriented V2X Communications

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE The Selection of Relevance: Content-Selection Strategies in Semantic and Task-Oriented V2X Communications Luca Lusvarghi, Javier Gozalvez UWICORE Laboratory, Universidad Miguel Hernandez de Elche, Elche, Spain. Email: {llusvarghi, j.gozalvez}@umh.es Abstract—Semantic and task-oriented Vehicle-toEverything (V2X) communications have been recently proposed to address the scalability challenges of future V2X networks by focusing on the relevance of the exchanged information. Relevance captures the impact of the transmitted information meaning on the intended receiver’s task by taking into account the intended receiver’s context, e.g., its position, speed, and planned driving intentions. Due to its context-dependent nature, relevance can greatly vary across different driving scenarios and different intended receivers, as each vehicle experiences unique context conditions. Hence, selecting the content of the transmitted messages to provide all intended receivers with the required relevant information can be a challenging task, especially in the broadcast V2X communication domain where each transmitting vehicle has multiple intended receivers. This work analyses the impact of different content-selection strategies on the performance of semantic and task-oriented V2X communications. The obtained results show that contentselection strategies play a critical role, particularly under constrained communication scenarios, and that their impact can significantly vary under different driving scenarios. Our analysis highlights key tradeoffs in the content-selection task and offers valuable insights for the design of effective contentselection strategies able to maximize the performance of semantic and task-oriented V2X communications. Keywords—semantic communications, task-oriented communications, V2X, 6G, relevance, content selection. I. INTRODUCTION Vehicle-to-Everything (V2X) communications are a key enabler in the Connected and Autonomous Driving (CAD) vision of safer, more efficient, and more sustainable mobility. V2X communications allow vehicles to enhance their situational awareness beyond the line-of-sight of their onboard sensors and develop new cooperative driving capabilities through the exchange of a broad range of information. For instance, vehicles can leverage V2X communications to share their position and speed (cooperative awareness), detected objects (cooperative perception), and planned trajectories (cooperative driving). However, existing communication paradigms are not ready to scalably support the large-scale deployment of V2X communications in future 6G and beyond networks. Currently, V2X communications are mainly designed to guarantee the reliable and timely delivery of information. This is the typical design principle of the majority of existing communication systems and corresponds to a Level A (technical level) design according to Shannon and Weaver’s three-level communication model [1]. A Level A design pays limited attention to the content of the transmitted messages and, therefore, it may lead to the transmission of information that is not needed at the receiver. This can potentially lead to an inefficient usage of communication resources that can ultimately compromise the scalability of V2X communications. To address current V2X scalability challenges, recent works have focused on the design of beyond-Level A communication paradigms that focus on the content of the transmitted messages. Level B (semantic level) semantic communications have recently been proposed to improve communication efficiency by focusing on the transmission of meaning instead of raw data. Meaning represents the essential set of semantic features needed to correctly reconstruct the raw data at the receiver. Applications of Level B semantic communications to the V2X domain mainly concentrate on vehicle-to-vehicle image transmission tasks. For example, an Image Segmentation Semantic Communication (ISSC) system that transmits only the essential semantic features extracted from raw onboard camera images was proposed in [2]. The work in [3] demonstrated that semantic V2X communications can push image compression ratios beyond the limits of conventional lossy compression techniques while reducing the end-to-end latency and the energy consumption. In [4], the authors put forth an autoencoder-based semantic V2X communication system that focuses meaning extraction on regions of the input image associated with important object categories (e.g., vehicles and trucks). Existing works demonstrated that semantic communication systems can greatly reduce the amount of exchanged information in one-to-one unicast communication settings. However, the V2X domain is a broadcast communication scenario where each transmitting vehicle has multiple intended receivers. In a (one-to-many) broadcast setting, curating the content of the transmitted messages with the specific meaning required by each intended receiver is a challenging task that has not yet been addressed. On a parallel research line, a number of existing works has concentrated on the design of Level C (effectiveness level) task-oriented communication paradigms centred on the Age of Information (AoI), Age of Incorrect Information (AoII), or Value of Information (VoI) metrics. Task-oriented paradigms aim to improve communication efficiency by transmitting only the information needed by the intended receiver’s task. Existing proposals rely on the assumption that the employed metrics can accurately capture the impact of the transmitted information on the receiver’s task. In the V2X domain, the majority of task-oriented communication systems relies on VoI and has been framed in the cooperative perception context. State-of-the-art works leverage VoI definitions based on the relative entropy [5], the dynamics [6] (e.g., position, speed, and heading variations), or the detection accuracy [7][8] of the detected objects. The cooperative perception standard includes additional VoI definitions based on the redundancy and classification confidence of the detected objects [9]. Existing VoI definitions attempt to capture the impact of the transmitted information on the intended receivers’ task without considering their context. Context refers to all the circumstances under which a piece of This work was supported by the European Union under the 2023 MSCA Postdoctoral Fellowship program (project no. 101153845) and by MCIN/AEI/10.13039/501100011033 (PID2023 -150308OB-I00). information is exchanged or processed. Context is fundamental to accurately estimate the impact of a piece of information on the intended receivers’ task. Information about an unseen (high entropy-VoI [5]), high-speed (high dynamicsVoI [6]), and accurately detected (high accuracy-VoI [7][8]) object might have no impact on the driving task of an intended receiver. The detected object’s impact depends on, for example, the intended receiver’s current state (position, heading, speed) and planned actions (stop, accelerate, turn), i.e., its context. Hence, VoI-based communication systems cannot completely prevent the transmission of unnecessary information and may not be able to support V2X communications at scale [10]. A more recent and promising solution for addressing the scalability challenges of future V2X networks is the semantic and task-oriented V2X communication paradigm proposed by the authors in [11]. Differently with respect to the state-of-theart, the proposed semantic and task-oriented V2X communication paradigm features a joint Level B and Level C design centred on the relevance of the exchanged information. Relevance is a joint Level B-and-C concept that captures the impact of a piece of information on the intended receiver’s task (Level C) by intertwining its meaning (Level B) with the intended receiver’s context. By leveraging context, relevance can more accurately estimate the impact of the transmitted information on the intended receivers’ task with respect to existing metrics (AoI, AoII, and VoI). As a result, the proposed relevance-based semantic and taskoriented V2X communication paradigm can limit the transmission of unnecessary (or irrelevant) information and significantly improve the communication efficiency compared to existing communication paradigms [11]. Semantic and task-oriented V2X communications rely on the capability of the transmitting vehicles to include the most relevant information for their intended receivers in the transmitted messages. Yet, relevance can significantly vary across different driving scenarios and conditions due to its context-dependent nature. In broadcast V2X communications, relevance can also substantially vary across different intended receivers as each vehicle experiences unique context conditions. Hence, selecting the content of the transmitted messages in order to provide all intended receivers with the required relevant information is a challenging task. To this end, this work analyses the impact of different content-selection strategies on the performance of semantic and task-oriented V2X communications. The analysis shows that employing an adequate content-selection strategy is especially important in constrained communication scenario. It also reveals that the performance of each content-selection strategy can greatly vary across different driving scenarios and that no strategy can maximize the performance of semantic and task-oriented V2X communications in all respects. The outcomes of our analysis shed light on important tradeoffs that characterize the content-selection task and offer valuable insights for the design of more effective content-selection strategies. II. A SEMANTIC AND TASK-ORIENTED V2X COMMUNICATION PARADIGM A. The Vision The relevance theory [12] identifies the search for relevance as the key communication principle that allows human interactions to achieve the greatest possible communication efficiency with minimal communication effort. According to the relevance theory, relevance is defined as the impact that a piece of information has on the listener’s perception of the world when processed within a context of existing assumptions. The vision of a semantic and task-oriented V2X communication paradigm stems from the relevance theory. It aims to bring the highly efficient relevance-based human communication principles to the V2X domain, making the search for relevance the core principle of vehicular communications. In the semantic and task-oriented V2X communication paradigm, the concept of relevance is extended to jointly integrate Level B and Level C design principles. Relevance is defined as the context-dependent impact that the meaning of a piece of information has on the receiving vehicle’s digital representation of the driving environment and, ultimately, on its driving tasks. By focusing on relevance, semantic and task-oriented V2X communications can avoid the transmission of unnecessary (or irrelevant) information. This is key to improve the communication efficiency and address V2X scalability challenges in future 6G and beyond networks [11]. The potential of semantic and task-oriented V2X communications relies on the transmitting vehicles’ ability to identify and transmit the most relevant information for their intended receivers. In the V2X domain, the relevance of a piece of information strongly depends on context. Context refers to all the circumstances (e.g., driving scenario and conditions, transmitter/receiver planned actions, previously exchanged messages) under which a piece of information is exchanged or processed. For example, a pedestrian crossing an intersection is highly relevant for a vehicle approaching the same intersection at high speed (context 1). It can strongly influence the approaching vehicle’s driving decisions. Conversely, the pedestrian’s relevance can substantially reduce if the vehicle is approaching the intersection at low speed (context 2) and it can become negligible if the vehicle is leaving the intersection (context 3). In addition, relevance cannot be univocally defined in the broadcast V2X communication scenario where each transmitting vehicle has multiple intended receivers. Each intended receiver has a unique context and, hence, the relevance of a piece of information can greatly vary across the set of intended receivers. B. Cooperative Perception Example The potential of the proposed semantic and task-oriented V2X communication paradigm is qualitatively illustrated by the cooperative perception example depicted in Fig. 1. Fig. 1 represents a T-shaped intersection with one transmitting vehicle (𝑉𝑉1), one receiving vehicle (𝑉𝑉2), and four objects (𝑥𝑥1, 𝑥𝑥2, 𝑥𝑥3, and 𝑥𝑥4). The four objects represent non-connected (a) (b) Fig. 1. Cooperative perception without (a) and with (b) semantic and task - oriented V2X communications. Object Vehicle Transmitted Object 𝑉𝑉 1 𝑥𝑥 4 𝑉𝑉 2 𝑥𝑥 1 𝑥𝑥 2 𝑥𝑥 3 𝑉𝑉 1 𝑥𝑥 4 𝑉𝑉 2 𝑥𝑥 1 𝑥𝑥 2 𝑥𝑥 3 vehicles1 and we assume that they can be locally detected by 𝑉𝑉1 with its onboard sensors. In Fig. 1(a), 𝑉𝑉1 selects the content of the transmitted message based on the redundancy of the locally detected objects. As long as they have not already been reported by another vehicle (i.e., are not redundant), 𝑉𝑉1 will include all four objects in its transmitted message. A redundancy-based VoI definition is currently adopted in the cooperative perception standard to curate the content of the generated messages [9]. In Fig. 1(b), 𝑉𝑉1 selects the transmitted message content based on the relevance of the locally detected objects for its intended receiver (𝑉𝑉2). In this case, only objects 𝑥𝑥2 and 𝑥𝑥3 are included in the transmitted message. These objects are relevant for 𝑉𝑉2 because they are approaching the intersection and might have an impact on its driving decisions, as 𝑉𝑉2 is also approaching the intersection. Conversely, objects 𝑥𝑥1 and 𝑥𝑥4 are irrelevant for 𝑉𝑉2 as they are leaving the intersection and cannot influence its driving. The comparison in Fig. 1 qualitatively demonstrates that redundancy (hence, redundancy-based VoI) alone cannot accurately capture the intended receiver’s task-oriented needs. When focusing on redundancy, the transmitting vehicle can include irrelevant information in the transmitted message (objects 𝑥𝑥1 and 𝑥𝑥4), leading to an inefficient usage of the available communication resources. Instead, semantic and task-oriented V2X communications can reduce the amount of consumed resources and improve communication efficiency while providing the intended receivers with all required relevant data. The intuitions offered by this qualitative analysis are quantitatively confirmed through numerical simulations in [11]. III. SEMANTIC EVALUATION FRAMEWORK The analysis performed in this study relies on a custom semantic evaluation framework initially presented in [11] that consists of a driving scenario populated by 𝑁𝑁 vehicles and 𝐾𝐾 exogenous variables. Exogenous variables are randomly distributed and represent detected objects. The set of all exogenous variables is denoted with 𝒦𝒦 and each exogenous variable is indicated with 𝑥𝑥𝑘𝑘, 𝑘𝑘= 1, … , 𝐾𝐾. Each vehicle 𝑉𝑉 𝑛𝑛 (𝑛𝑛= 1, … , 𝑁𝑁) in the scenario collects exogenous variables either locally, through its onboard sensors, or from other vehicles via V2X communications. We model the probability that a vehicle 𝑉𝑉 𝑛𝑛 locally detects an exogenous variable 𝑥𝑥𝑘𝑘 using its onboard sensors with the following logistic function [11]: 𝑃𝑃�𝐷𝐷𝑘𝑘,𝑛𝑛�=1 1 + 𝑐𝑐1∗ 𝑒𝑒−𝑐𝑐2(𝐷𝐷𝑘𝑘,𝑛𝑛−𝑐𝑐3) , (1) where 𝐷𝐷𝑘𝑘,𝑛𝑛 represents the distance between 𝑥𝑥𝑘𝑘 and 𝑉𝑉 𝑛𝑛. In the semantic evaluation framework, vehicles periodically generate – and transmit – a new message every 𝑇𝑇 ms. The content of each generated message consists of a selected subset of the exogenous variables locally detected by the transmitting vehicle through its onboard sensors. Exogenous variables are selected based on their relevance for the intended receivers. A vehicle 𝑉𝑉 𝑛𝑛 is estimated as an intended receiver if the transmitter can decode at least one of the last two messages transmitted by 𝑉𝑉 𝑛𝑛. 1 In cooperative perception, detected objects can represent static obstacles or different types of connected and non-connected road users (vehicles, two-wheelers, pedestrians, etc.). The transmitting vehicle estimates the relevance of locally detected exogeneous variables in two steps. First, it performs redundancy estimation by analysing the content of the messages exchanged on the V2X network. Any exogenous variable 𝑥𝑥𝑘𝑘 correctly received by the transmitter via V2X is estimated to be redundant for all intended receivers. Note that this is the redundancy estimation approach employed in cooperative perception studies and standards [9]. Exogenous variables estimated to be redundant are considered irrelevant for the intended receivers 2. Then, it performs relevance estimation, which consists of estimating the relevance of the locally collected non-redundant exogenous variables for each intended receiver 𝑉𝑉𝑅𝑅. At each intended receiver, the context-dependent relevance of each exogenous variable 𝑥𝑥𝑘𝑘∈ 𝒦𝒦 is modelled by a semantic value 𝑤𝑤𝑅𝑅,𝑘𝑘∈[0,1] that represents the contextdependent relevance of each 𝑥𝑥𝑘𝑘 for 𝑉𝑉𝑅𝑅. Relevant exogenous variables are assigned a semantic value larger than zero (𝑤𝑤𝑅𝑅,𝑘𝑘> 0), while irrelevant variables are assigned a semantic value 𝑤𝑤𝑅𝑅,𝑘𝑘= 0 . At each intended receiver, the maximum distance from a relevant exogenous variable is limited by the relevance range 𝐷𝐷𝑚𝑚𝑚𝑚𝑚𝑚 . All exogenous variables located outside the 𝐷𝐷𝑚𝑚𝑚𝑚𝑚𝑚 range are considered irrelevant for 𝑉𝑉𝑅𝑅. Within 𝐷𝐷𝑚𝑚𝑚𝑚𝑚𝑚 , the total number of relevant exogenous variables is 𝑀𝑀, and the remaining variables are irrelevant for 𝑉𝑉𝑅𝑅. Relevant exogenous variables are further divided into short-term and medium-term relevant variables. The semantic value (i.e., the relevance) of short-term relevant exogenous variables (𝑤𝑤𝑅𝑅,𝑘𝑘 𝑆𝑆) is 2 times larger than the semantic value of medium-term relevant variables ( 𝑤𝑤𝑅𝑅,𝑘𝑘 𝑀𝑀). At each intended receiver, the probability of shortand medium-term relevant variables is denoted by 𝑝𝑝𝑆𝑆 and 𝑝𝑝𝑀𝑀, with 𝑝𝑝𝑆𝑆+𝑝𝑝𝑀𝑀= 1. The transmitter employs the estimated multi-receiver relevance of each locally collected exogenous variable to select the content of the generated messages. In the semantic evaluation framework, vehicles transmit the generated messages employing the C-V2X sidelink communication technology. C-V2X sidelink performance is modelled through the analytical models in [13]. These models capture the message reception probability under co-channel interference (packet collisions), propagation impairments (pathloss, shadowing, fast fading), and half-duplex transceiver constraints. The analytical models express the message decoding probability as a function of the transmitter-receiver distance and the Channel Busy Ratio (CBR) at the receiver. The CBR is a channel load indicator that measures the fraction of channel resources sensed as occupied within the last 𝑇𝑇 ms. The sensing of channel resources is modelled through the sensing probability model also openly released in [13]. IV. CONTENT SELECTION A. Content-Selection Challenges Let’s consider the cooperative perception example depicted in Fig. 2. In Fig. 2, the transmitting vehicle (𝑉𝑉1) has three intended receivers ( 𝑉𝑉2, 𝑉𝑉3 and 𝑉𝑉4) and is locally detecting objects 𝑥𝑥1 and 𝑥𝑥2 with its onboard sensors. In an unconstrained communication scenario, 𝑉𝑉1 would include both 𝑥𝑥1 and 𝑥𝑥2 in its transmitted message as they are both relevant for its intended receivers. Object 𝑥𝑥1 is short2 Redundant information is assumed to have no impact on the intended receiver’s understanding of the driving environment and, ultimately, on its driving tasks. Note that this is a common assumption employed also in standardized cooperative perception message generation rules. term relevant information for 𝑉𝑉4. Its reception has an immediate impact on 𝑉𝑉4’s current driving decisions, as it prompts 𝑉𝑉4 to urgently perform a strong deceleration to avoid the potential collision with 𝑥𝑥1 as it merges into its lane. Object 𝑥𝑥2 is medium-term relevant information for 𝑉𝑉2 and 𝑉𝑉3. It does not influence their immediate driving decisions, but it has an impact on their planned driving actions. The reception of 𝑥𝑥2 allows 𝑉𝑉2 and 𝑉𝑉3’s to anticipate a potentially safety-critical situation, i.e., the collision with the approaching object (𝑥𝑥2), and accordingly plan their future trajectories to maximize driving safety. In a constrained communication scenario where the maximum number of objects per transmitted message is equal to one, 𝑉𝑉1 would be forced to select either 𝑥𝑥1 or 𝑥𝑥2. In this case, a natural question arises: should 𝑉𝑉1 select the short-term relevant object (𝑥𝑥1) required by a single intended receiver or the medium-term relevant object (𝑥𝑥2) required by multiple intended receivers? Note that both shortand medium-term relevant information is required to maximize the safety and efficiency of the intended receivers’ driving task. Constrained communication scenarios shed light on the inherent contentselection challenges that characterize semantic V2X communications due to the broadcast nature of the V2X domain. The design of effective content-selection strategies able to provide multiple intended receivers with all required relevant information is a challenging yet crucial task to maximize the performance of semantic V2X communications. B. Content-Selection Strategies We analyse the impact of different content-selection strategies on the performance of semantic V2X communications. In semantic V2X communications, contentselection strategies are articulated into three phases. First, the transmitting vehicle filters out all exogenous variables that are estimated as irrelevant for all intended receivers. Then, it ranks the remaining exogenous variables following one of the following approaches. 1) Random: exogenous variables are randomly ranked. 2) Count: exogenous variables are ranked based on the number of intended receivers for which they are relevant, regardless of their relevance type (shortor medium-term). 3) Average: exogenous variables are ranked based on the average relevance that they have for the intended receivers. The average relevance is computed considering only semantic values larger than zero. 4) Maximum: exogenous variables are ranked based on their maximum relevance for the intended receivers. 5) Hybrid: exogenous variables are ranked by combining the Maximum and Count approaches. Exogenous variables are first ranked based on their maximum relevance for the intended receivers. Then, exogenous variables with the same maximum relevance are further ranked based on the number of intended receivers for which they are relevant. Last, after ranking, the transmitter includes in the transmitted message only the top Г exogenous variables. Г defines the maximum number of exogenous variables that can be included in a transmitted message (a proxy of the message size). It represents the communication constraints that could be enforced by congestion control mechanisms to control the channel load in constrained communication scenarios. C. Simulation Assumptions We perform our analysis considering a 2 km x 2 km driving scenario. We set the coefficients 𝑐𝑐1, 𝑐𝑐2, and 𝑐𝑐3 in (1) to 0.08, -0.08, and 60, respectively. As a result, the perception range of each vehicle is 150 m. The perception range is the distance at which the probability that a vehicle can locally detect an exogenous variable with its onboard sensors is equal to zero. The average number of exogenous variables locally detected by each vehicle with its own onboard sensors is 20, a value that corresponds to the object density measured in a 5lanes low-density highway [14]. The relevance range, 𝐷𝐷𝑚𝑚𝑚𝑚𝑚𝑚, is set to 400 m and the total number of relevant variables (either shortor medium-term) within the relevance range, 𝑀𝑀, is set to 20. The 𝐷𝐷𝑚𝑚𝑚𝑚𝑚𝑚 setting is in line with the V2X service requirements defined by ETSI in [15]. We set the channel bandwidth to 10 MHz and the message generation (and transmission) period to 𝑇𝑇 = 40 ms. This setting is representative of a V2X scenario where multiple services (cooperative awareness, perception, and maneuvering) coexist within the same channel [10]. Vehicles communicate employing a QPSK-0.7 Modulation and Coding Scheme (MCS). The size of exogenous variables is set to 52 bytes, a configuration that corresponds to the typical size of a detected object in cooperative perception [16]. D. Numerical Results We start our analysis by considering a driving scenario where the probability of shortand medium-term relevant exogenous variables are respectively set to 𝑝𝑝𝑆𝑆= 0.1 and 𝑝𝑝𝑀𝑀= 0.9. This relevance profile corresponds to a driving scenario where the majority of driving environment information has medium-term relevance. Fig. 3 reports the Available Relevant variables Ratio (ARR) under different communication constraints (Г) as a function of the number of communicating vehicles (𝑁𝑁). The ARR is defined as the ratio between the number of relevant exogenous variables available at the intended receiver and 𝑀𝑀, the total number of relevant variables within the receiver’s relevance range. A relevant exogenous variable is available when it is either locally collected through the onboard sensors or it is received via V2X from other vehicles. The ARR is averaged across all intended receivers and it does not distinguish between shortand medium-term relevant variables. Fig. 3(a) shows that the Count strategy maximizes the ARR performance of semantic V2X communications under stringent communication constraints (Г= 2). A higher ARR indicates that transmitting vehicles can serve their intended receivers with a larger fraction of the required (shortor medium-term) relevant information. Note that the gap between the Count strategy and the other strategies increases with 𝑁𝑁. As 𝑁𝑁 increases, the number of intended receivers at each transmitting vehicle also increases, emphasizing the impact of an effective contentselection at the transmitter. The analysis of Fig. 3(a) through Fig. 3(d) reveals that the ARR performance gap between content-selection strategies reduces as communication constraints get looser (i.e., for Fig. 2. Semantic and taskoriented V2X communications with multiple intended receivers. Cooperative perception example. 𝑉𝑉 4 𝑥𝑥 2 Object Vehicle 𝑥𝑥 1 𝑉𝑉 1 𝑉𝑉 3 𝑉𝑉 2 larger Г values). Less constrained communication scenarios allow the transmitting vehicles to include larger amounts of relevant information in each message. This relaxes the need for an effective content-selection strategy. In Fig. 3(d), where Г= 8 , vehicles can include all locally detected relevant variables in the transmitted message and the baseline Random strategy leads to the same ARR performance as other solutions 3. Fig. 3(a)-(d) show that properly selecting the content of the transmitted messages is particularly important in constrained communication scenarios (see Section IV.A). For this reason, the remainder of this work will focus the analysis of content-selection strategies in the Г= 2 setting. The analysis of Fig. 3(a) is extended by Fig. 4(a) and Fig. 4(b), which report the ARR measured on the subsets of shortterm (ARR-short) and medium-term relevant (ARR-medium) exogenous variables when Г= 2. Fig. 4(a) shows that the Count strategy cannot guarantee the delivery of short-term relevant variables to the intended receivers. It leads to an ARR-short performance that is better than the Random baseline, but significantly worse than the Hybrid, Maximum, and Average strategies. The Count strategy prioritizes the transmission of information that is relevant (either shortor medium-term relevant) to a larger number of intended receivers. In the 𝑝𝑝𝑆𝑆= 0.1 (and 𝑝𝑝𝑀𝑀= 0.9) profile case, the majority of exogenous variables has a medium-term relevance for the intended receivers. Hence, the Count strategy prioritizes the transmission of medium-term relevant information to maximize the number of served intended receivers (see Fig. 3(a)). Semantic V2X communications achieve the best ARR-medium performance with the Count strategy in Fig. 4(b). On the other hand, the Hybrid and Maximum contentselection strategies prioritize the transmission of short-term relevant variables, as they have a larger relevance (or semantic value) for the intended receivers. By design, they maximize the ARR-short performance (see Fig. 4(a)) sacrificing the transmission of medium-term relevant variables (see Fig. 3 Figs. 3(a)-(d) also show that the ARR performance exhibits a decreasing trend at medium-large vehicular densities ( 𝑁𝑁). As the number of communicating vehicles increases, so does the risk of packet collisions, 4(b)). Smaller ARR-medium values directly translate into fewer intended receivers being served with the required relevant information in the 𝑝𝑝𝑆𝑆= 0.1 scenario (see Fig. 3(a)). It is worth highlighting that the Hybrid strategy leads to a better ARR-medium performance than the Maximum, Average, and Random strategies. This is the case because the Hybrid strategy prioritizes the transmission of exogenous variables that have short-term relevance for one intended receiver and medium-term relevance for other intended receivers. Instead, the Maximum and Average strategies deprioritize the transmission of medium-term relevant variables by design, as they have a lower relevance (or semantic value), thus yielding ARR-medium values lower than the Random baseline. The Maximum strategy exclusively focuses on the transmission of variables with short-term (hence, higher) relevance. The Average strategy favours the transmission of exogenous variables that have short-term relevance for one intended receiver and medium-term relevance for a smaller number of intended receivers. Exogenous variables with short-term relevance for one intended receiver and medium-term relevance for a smaller number of intended receivers have a larger average semantic value and are prioritized in constrained communication scenarios. As a result, the Average strategy leads to the same ARR-short performance of the Hybrid and Maximum strategies (Fig. 4(a)) while yielding the lowest ARR-medium values (see Fig. 4(b)). Next, we consider two additional relevance profiles with [𝑝𝑝𝑆𝑆, 𝑝𝑝𝑀𝑀] respectively set to [0.4, 0.6] and [0.7, 0.3]. These relevance profiles represent driving scenarios with increasing amounts of information with short-term (hence, higher) relevance. Fig. 5 reports the ARR performance of semantic V2X communications in the 𝑝𝑝𝑆𝑆= 0.4 and 𝑝𝑝𝑆𝑆= 0.7 profile scenarios under stringent communication constraints (Г= 2). The comparison between Fig. 3(a) and Fig. 5 reveals that the ARR performance achieved with the Count strategy does not change under different driving scenarios. It also illustrates that the ARR performance of the Hybrid, Maximum, and Average strategies exhibit opposite trends as the amount of short-term relevant information in the driving scenario increases. The Hybrid strategy gradually approaches – and eventually matches (see Fig. 5(b)) – the ARR obtained with the Count strategy. Instead, the Maximum and Average strategies yield progressively lower ARR values, approaching (and even surpassing) the Random baseline. In Fig. 5(b), where 𝑝𝑝𝑆𝑆= 0.7, the Average strategy leads to the worst ARR performance. The same trends reported in Fig. 3 (𝑝𝑝𝑆𝑆= 0.1) for varying Г ultimately increasing the likelihood that relevant information cannot be correctly decoded at the receiver. This effect is particularly evident under large Г values. (a) Г = 2 (b) Г = 4 (c) Г = 6 (d) Г = 8 Fig. 3. ARR under different communication constraints (Г). 𝑝𝑝𝑆𝑆= 0.1. (a) (b) Fig. 4. ARR-short (a) and ARR-medium (b), Г= 2. 𝑝𝑝𝑆𝑆= 0.1. values have also been observed with the 𝑝𝑝𝑆𝑆= 0.4 and 𝑝𝑝𝑆𝑆= 0.7 relevance profiles: the performance of semantic V2X communications becomes less sensitive to the employed content-selection strategy as communications constraints become more relaxed. Fig. 6(a) and Fig. 6(b) complete the 𝑝𝑝𝑆𝑆= 0.7 relevance profile analysis by focusing on the ARR-short and ARRmedium metrics. In Fig. 6(a), semantic V2X communications achieve the same ARR-short performance when either the Count or Hybrid strategy is employed. This is a remarkable difference compared to the 𝑝𝑝𝑆𝑆= 0.1 case, where the Count strategy could not guarantee the delivery of short-term relevant exogenous variables (see Fig. 4(a)). The majority of exogenous variables has short-term relevance for the intended receivers when 𝑝𝑝𝑆𝑆= 0.7 . Hence, the Count strategy prioritizes the transmission of short-term relevant variables to maximize the number of receivers served with relevant information. For this reason, Count and Hybrid strategies also yield nearly identical ARR-medium values in Fig. 6(b). The comparison between Fig. 4 and Fig. 6 highlights the degradation suffered by the Maximum and Average strategies in the 𝑝𝑝𝑆𝑆= 0.7 case. In Fig. 6(a), they are no longer able to maximize the ARR-short performance, yielding lower values compared to the Count and Hybrid strategies, especially at large vehicular densities (𝑁𝑁). The ARR-short degradation is inherent to their design, as neither the Maximum nor the Average strategy take into account the number of receivers for which an exogenous variable has a short-term (hence, higher) relevance. In addition, Fig. 6(b) shows that the ARR-medium performance obtained with the Maximum and Average strategy drops even further when 𝑝𝑝𝑆𝑆= 0.7. V. CONCLUSIONS Selecting the content of the transmitted messages to provide all intended receivers with the required relevant information is a critical yet challenging task in semantic and task-oriented V2X communications. The delivery of all required relevant information is fundamental to maximize the intended receivers’ driving safety and efficiency, regardless of its short-term or medium-term nature. The obtained results revealed that content-selection strategies play a particularly critical role especially in constrained communication settings and at high vehicular densities, i.e., when the number of intended receivers is large. They also showed that the impact of content-selection strategies can greatly vary across driving scenarios characterized by different amounts of shortand medium-term relevant information. Our analysis highlighted that no single content-selection strategy could maximize the performance of semantic and task-oriented V2X communications in all respects. In particular, it demonstrated that selecting the locally collected information based on its maximum or average relevance is not an effective content-selection strategy across all considered settings. Moreover, the analysis shed light on fundamental tradeoffs between serving a larger set of intended receivers with medium-term lower-relevance information and ensuring the delivery of short-term higherrelevance information to fewer receivers. These tradeoffs underline the need for adaptive and collaborative contentselection strategies where the transmitting vehicles jointly optimize the content of their messages to provide the intended receivers with all required relevant information under varying vehicular densities, communication constraints, and scenarios. REFERENCES [1] C. Shannon and W. Weaver, “The Mathematical Theory of Communication,” University of Illinois Press, 1962. [2] Q. Pan et al., “Image Segmentation Semantic Communication over Internet of Vehicles,” in Proc. 2023 IEEE WCNC, UK, 2023, pp. 1-6. [3]J. M. Gimenez-Guzman, I. Leyva-Mayorga and P. Popovski, “Semantic V2X Communications for Image Transmission in 6G Systems,” IEEE Network, vol. 38, no. 6, pp. 48-54, Nov. 2024. [4] J. Löhdefink et al., “Focussing Learned Image Compression to Semantic Classes for V2X Applications,” in Proc. 2020 IEEE Int. Veh. (IV) Symposium, Las Vegas, NV, USA, 2020, pp. 1641-1648. [5] T. Higuchi et al., “Value-Anticipating V2V Communications for Cooperative Perception,” in Proc. 2019 IEEE Int. Veh. (IV) Symposium, Paris, France, 2019, pp. 1947-1952. [6] G. Thandavarayan et al., “Redundancy Mitigation in Cooperative Perception for Connected and Automated Vehicles,” in Proc. IEEE VTC2020-Spring, Antwerp, Belgium, 2020, pp. 1-5. [7] V. A. Wolff et al., “Uncertainty and Prioritization: Empirical Evaluation of a VoI-Based CPM Generation Pipeline Using RealWorld Data,” in Proc. 2025 IEEE VNC Conf., Porto, Portugal, 2025. [8] T. Lyu et al., “Accuracy and Relevance: A Value of Information based Prioritisation of Perceived Objects for the ETSI Collective Perception Service,” in Proc. 2025 IEEE VNC Conf., Porto, Portugal, 2025. [9] ETSI, “Intelligent Transport System (ITS); Vehicular Communications; Basic Set of Applications; Collective Perception Service; Release 2,” ETSI TS 103 324 V2.1.1, June 2023. [10] M. Sepulcre et al., “LTE-V2X Scalability and Spectrum Requirements to support Multiple V2X Services”, in Proc. VTC2024-Fall, Washington DC, USA, 2024, pp. 1-6. [11] L. Lusvarghi et al., “The Search for Relevance: A Context-Aware Paradigm Shift in Semantic and Task-Oriented V2X Communications,” 2025, arXiv:2508.07394. [12] D. Sperber and D. Wilson, “Relevance: Communications and Cognition,” Blackwell Publishers Inc. 1995. [13] M. Gonzalez-Martín et al., “Analytical Models of the Performance of C-V2X Mode 4 Vehicular Communications,” IEEE Transactions on Vehicular Technology, vol. 68, no. 2, pp. 1155-1166, Feb. 2019. [14] F. Schiegg, I. Llatser and T. Michalke, “Object Detection Probability for Highly Automated Vehicles: An Analytical Sensor Model,” in 2019 Proc. VEHITS Conf., Crete, Greece, 2019, pp. 223-231. [15] ETSI, “5G; Service requirements for enhanced V2X scenarios,” ETSI TS 122 186 V17.0.0, April 2024. [16] E. Xhoxhi and F. A. Schiegg, “A First Study on the Spectrum Needs for Release 2 V2X Services,” in Proc. VTC2023-Fall, Hong Kong, Hong Kong, 2023, pp. 1-6. (a) (b) Fig. 5. ARR, Г = 2. 𝑝𝑝𝑆𝑆= 0.4 (a) and 𝑝𝑝𝑆𝑆= 0.7 (b) profiles. (a) (b) Fig. 6. ARR-short (a) and ARR-medium (b), Г= 2. 𝑝𝑝𝑆𝑆= 0.7.