Unveiling water allocation dynamics: a text analysis of 25 years of stakeholder meetings

Managing water resources in regions with high climate variability and frequent extreme weather events poses challenges for policymakers. To facilitate water allocation in these cases, participatory and collaborative decision-making approaches have become common. However, the evaluation of these approaches is hindered by the lack of structured methods and data to understand them. To address this knowledge gap, we propose a novel methodology that leverages text data to identify key topics, conflicts, and influential actors that shape water allocation dynamics. Our methodology is tested using records of 1020 water basin committee meetings held between 1997 and 2021 across 12 basin committees in Ceará, Brazil—a region known for its extensive history of droughts that have impacted water governance. To uncover key water management issues discussed during these meetings, we employed a three-step topic modeling framework: (1) sentence embedding, (2) dimensionality reduction, and (3) sentence clustering. Furthermore, we used entity recognition, dependency parsing, and network graphs to identify powerful actors influencing these meetings and, ultimately, the decisions taken. Our findings revealed stakeholders’ heightened concern for urban water supply over agricultural demand during droughts. We found that ‘reservoir operation’ was the most recurring topic, especially in basins where the strategic reservoirs are located. Discussions related to ‘climate information’ became significantly more important over time, which indicates that water allocation decisions are increasingly based on the seasonal forecast and data on oceanic indices provided by the meteorology agency. Despite the presence of local users in the committees, governmental representatives dominated the discussions and were central in all river basins. In conclusion, our proposed approach harnesses existing text data to uncover spatiotemporal patterns related to participatory water allocation. This study opens new avenues for investigating water governance using text-based analysis.


Introduction
Managing water resources in scenarios where their availability is uncertain and their uses are conflicting requires a coordinated response from stakeholders and decision-makers [1].Water management is a complex and political task, so its social context cannot be disengaged [2,3].Acknowledging this challenge, frameworks such as integrated water resources management (IWRM) [4] and adaptive management [5,6] have been developed.The IWRM framework [7] focuses on integrating management across multiple scales and resources (i.e.water and land) while attending to the needs of multiple users.Adaptive management, on the other hand, incorporates uncertainty assessment by promoting a flexible, continuous learning approach [8,9].
In recent decades, collaborative and participatory approaches have been widely applied to promote IWRM and adaptive management.However, their utilization in the Global South [10,11] has been limited compared to the United States and Europe [12].
Notable examples exist in Brazil and Chile [13][14][15][16].The appeal of these approaches lies in their effectiveness when implemented within a participatory framework.Specifically, such frameworks have proven to be effective in managing conflict [17], reaching agreed targets [18], integrating local knowledge [19], and improving water planning [20].
The advantage of participatory water management approaches is rooted in their transparency, thus fostering trust and promoting social learning [21].By involving relevant actors, they can incorporate the perspectives of vulnerable and marginalized groups-a dimension often overlooked in modeldriven approaches.Nevertheless, they also carry the risks of power imbalances [22], and the prevalence of hierarchical governance approaches [19].Additionally, the lack of systematic analysis of the process poses a risk to the effectiveness of participatory approaches [23].Therefore, there is a pressing need to develop evaluation tools to assess the impact of participatory water management initiatives.
The evaluation of participatory water management is, however, often hindered by the lack of structured methods and data.Typically, the analysis of water management approaches relies on inductive methods based on document analysis, cursory analysis (e.g.[24]), and interviews (e.g.[25,26]).For instance, Ferreyra and Beard [27] conducted 25 interviews with water managers in Canada to evaluate the perceived collaborative water management outcomes.Minutes of meetings have also been analyzed qualitatively to provide insights into the effectiveness of water management policies [28,29].Within this context, Fischhendler and Katz [30] investigated the minutes of 47 meetings concerning water negotiations between Israel and Palestine.Despite providing in-depth details, these methods are limited in terms of comprehensiveness, as only a few documents or interviews can be analyzed.
In the wake of digitalization, much of the information about water management decisions has moved to virtual platforms, including databases of meeting records (e.g.European Council [31] and UN meetings) and legislative and policy documents (e.g. the FAOLEX database: http://faolex.fao.org/faolex/index.htm).This digital landscape provides vast amounts of text on an unprecedented scale, breadth, and depth.At the same time, advancements in machine learning and natural language processing (NLP) have opened previously unforeseen opportunities for using the unprecedented abundance of texts to understand water management practices.
Here, we seek to exploit these new developments systematically by proposing an automated approach to investigate participatory water management practices based on NLP and social network analysis.More specifically, we consider the minutes of 1020 water basin committee meetings held in Ceará, Northeast Brazil, to provide insights into the participatory IWRM activities conducted in the past 20 years.The goal was threefold: (1) to identify the water management topics that concern stakeholders in the decisionmaking process, (2) to map their spatiotemporal dynamics and trajectories, and (3) to identify influential actors involved in water allocation dynamics, their linkages (or disconnections) and how they are related to the water management topics.To the best of our knowledge, this approach is the first to perform an automated text analysis of water committee meeting records as well as an automated assessment of stakeholder networks based on text data.

Methods
In this study, we used the minutes of water committee meetings to examine water allocation dynamics in Ceará, Brazil.We employed a series of NLP and social network analysis tools as described in the following sections.A detailed workflow of our methodology is provided in the supplementary information.Our approach is generalizable for other case study areas as long as documents with enough coverage (in time and space) are available.

Case study area
To illustrate how text data can be used to assess participatory water management, we considered the case of Ceará, Northeast Brazil.This state, predominantly located (>90%) in the Brazilian semiarid region, is frequently affected by severe and long-lasting meteorological and hydrological droughts [32,33].The state is divided into 12 hydrographic basins (figure 1), with extensive water infrastructure, including 155 monitored artificial reservoirs with a storage capacity of 18.67 billion m 3 .
Water availability in Ceará is not only seriously affected by frequent and recurrent droughts but is also subject to a high temporal and spatial climate variability-one of the highest in the world [34].To address these challenges, water governance in Ceará has been regulated since the early 90s through national laws and state legislation.Notably, the establishment of river basin committees and negotiated water allocation strategies have emerged as pivotal mechanisms for fostering public participation.The first committee was established in 1992, whereas the others were implemented in subsequent years, according to the need for urgent collective decisionmaking on water allocation (figure 1(A)).
The committees meet regularly to discuss water management issues.After the rainy season, they also hold annual water allocation meetings, which work as a negotiation forum in single or multiple reservoir water systems, allowing for participatory reservoir operation.Currently, the river basin committees are composed of representatives of governmental and non-governmental institutions, with the following distribution and percentage of participation: local

Text data collection and pre-processing
Minutes from committee meetings held between 1997 and 2022 in Ceará were scrapped from the water basins' websites (https://portal.cogerh.com.br/comites-de-bacias-hidrograficas).These minutes are publicly available and are uploaded after being approved by committee members.A total of 1020 documents were obtained and converted into plain text.They were then classified according to the corresponding water basin committee (figure 1(B)) and the meeting date.
We pre-processed the text data by removing stop words, numbers, and special characters.Furthermore, to avoid introducing noise in the identification of key topics discussed, we removed proper nouns (e.g.names of persons or locations) and structural elements of the minutes (e.g.list of participants' names and meeting agenda).To automatically remove this information, we used regex to detect in which part of the text the minute starts (e.g.'the meeting started with' or 'the meeting was opened by').Municipality names and people's names detected with the named entity recognition (NER) tool from SpaCy [36] were also removed from the texts.Finally, we segmented the text into sentences using the sentence tokenizer module from the Python package nltk [37].We further removed sentences composed of less than five words, as we considered these to be uninformative.To validate this assumption, we randomly selected a sample of ten sentences with less than five words for each committee and read them.This process resulted in a clean corpus of 49 111 sentences.

Topic modeling
To detect the main topics discussed in the water basin meetings, we created a topic modeling approach that (1) converts the sentences into high-dimensional vectors (i.e.sentence embedding), (2) reduces the dimensionality of these embeddings, and (3) clusters the resulting dimensionality reductions.We opted to work at the sentence level to capture the context and the overall message of the discussions more effectively.An unsupervised deep-learning framework based on the sentence-level BERT (sBERT) algorithm [38] was used for sentence embedding.Unlike text topic models such as word2vec [39], sBERT enriches word representations with contextual information.
In the second step, we reduced the dimensionality of these embeddings to decrease the computational cost of clustering them while keeping their characteristics, i.e. sentences with similar contexts would have similar vector representations.Dimensionality reduction was performed with uniform manifold approximation and projection (UMAP) [40], which is commonly applied for text embeddings [41].We also tested principal components analysis and t-SNE for this task, but the topics obtained with these methods were not as informative.
In the final step, we clustered the reducedembedded sentences using the mini-batch k-means model-a faster implementation of k-means that updates cluster centers with mini-batches.Compared to other clustering methods, such as HDBSCAN and hierarchical clustering, k-means produced more coherent and well-distributed topics.We used batches of size 100 and a maximum of ten iterations over the dataset.As a result, we obtained 40 clusters corresponding to different topics discussed in the water committee meetings.

Evaluation of the topic modeling clusters
The quality of the clusters obtained with the topic modeling was evaluated quantitatively and qualitatively.For the quantitative evaluation of how sensitive the clusters are to parameter choices, we used the silhouette coefficient [42].This coefficient measures how similar each sentence is to its cluster compared to other clusters.Hence, it quantifies how well each sample was classified.To obtain an overall measure of cluster quality, we calculated the average silhouette coefficient for all clusters.
For hyperparameter tuning, we used a grid search approach.We varied the number of neighbors in UMAP (5-20, varying in windows of 5) and the number of clusters in k-means (30-80, varying in windows of 10) and calculated the average silhouette coefficient for all parameter combinations.The analysis revealed little variation in the clustering quality (the silhouette coefficient varied between 0.25 and 0.32).Hence, we chose the parameter combination that provided the most coherent and informative topics following a qualitative analysis.

Labeling of the topic modeling clusters
To label the topics and reduce the subjectivity, which is inherent in this task [43], we followed a hierarchical clustering approach, combining quantitative and qualitative analyses.First, we inspected the 40 topics generated by the mini-batch k-means model, considering the ten most frequently occurring words in each topic.We further examined the ten sentences closest to the cluster's centroids.Throughout the labeling process, irrelevant topics were identified and thus removed from the analysis.For example, some topics were mainly composed of unimportant information, such as metadata or the approval of previous meeting minutes.Topics encompassing multiple distinct topics were partitioned into two or more clusters by rerunning k-means (see table S2 for a full list of labeled topics and the divisive approach).To evaluate the newly created clusters, we used cosine similarity to measure the degree of similarity among the clustered sentences.Furthermore, in some cases, the topics identified by our model were qualitatively similar enough that they were given identical labels and merged.As a result of this analysis, a total of 35 topics were selected for spatiotemporal analyses.

Social network analysis
In addition to investigating the key topics discussed, we also examined influential actors, water management instruments, and how they relate.To assess participation and analyze the social networks that pertain to the water allocation process, we used NER to extract the actors mentioned in each document in the corpus that did not undergo proper noun removal (see section 2.2).Here, we define actors as entities of people with one common goal, i.e. organizations, companies, associations, government bodies, and public, private, and third-sector entities.We also considered management instruments, such as water permits and river basin plans.While these do not constitute actors per se, we assume them as nonhuman entities that can relate to other actors-and which often hold account for water management [44].
After automatically extracting actors' names, we mapped them into 18 groups (table 1) using a regular expression rule-based search [45].Details on the criteria used to classify the actors into organizations are provided in table S3.This analysis did not include people's names mentioned in the meetings, as accurately attributing them to their respective actor group is challenging given that the minutes typically provide only first names (e.g.'José' could either refer to 'José Medeiros' or 'José Xavier').
After identifying the stakeholders and constructing their co-occurrence network, we computed different centrality measures (namely, degree centrality, betweenness, and closeness).

Statistical analyses
To assess spatial differences in topic frequency across water basins, we applied the Kruskal-Wallis test [46].This test compares groups of independent observations of a variable to detect differences between them.Here, the groups were the 35 topics identified in the meeting minutes, and the observations were the frequency of topic mentions over time.
To explore the relationship between actors and topics associated with agriculture, water quality, and climate, we used a linear mixed-effects regression model.In this model, the response variable was the topic frequency, while the frequency of actors' mentions in the minutes served as the independent variable.For this analysis, we regarded the meeting dates as a unit of time.Additionally, we considered the 12 water basin committees as a grouping variable, treating the data from each committee as independent.This allowed us to explore the connections between topics and actors while considering the unique characteristics of each committee.

Dynamics of water management topics over time and space
Using topic modeling, we mapped key discussions related to water supply, water use, agriculture, water quality, climate, and the participatory water management process in each of the basins in Ceará state to 35 topics (figure 2).Overall, topics related to 'participatory water management' , 'water supply and use' , and 'water allocation' themes dominated the discussions over the last 20 years (31%, 18%, and 16% of all sentences, respectively).Indeed, 'participatory water management' was identified as a key topic for all basins investigated (figure 3).Conversely, topics such as 'pesticide use and cattle raising' and 'agriculture water use' , which are crucial for the management of irrigation water demands-the second highest in the state, were not as widely discussed.Likewise, we found that few discussions were centered on environmental problems.We found a significant spatial difference regarding the distribution of some topics across the basins (figure 4).This indicates that stakeholders have different concerns depending on the basin they act on, reinforcing the need for adaptive management practices.For example, the topics 'water infrastructure' , 'municipal power participation' , and 'water allocation decisions' are distributed differently across the basins.Since 2013, a substantial infrastructure project called Cinturão das Águas has been implemented.The project will add over 1300 km of channels, tunnels, and siphons in Ceará so that the entire state can be supplied by surface reservoirs.The committees with a high frequency of mentions associated with 'water infrastructure' (i.e.Alto and Médio Jaguaribe) coincide with the water basins where most of the construction has been initiated or completed.Conversely, topics associated with management instruments have an almost homogeneous distribution across all the committees.
Regarding the temporal distribution, a noticeable upward trend in the frequency of topics such as 'alternative water sources' can be observed (figure 5).This indicates a possible increased search for capacity expansion of water supply.Indeed, when qualitatively examining these sentences, we found that potable water trucks are frequently mentioned (n = 102 word mentions), besides wells (n = 302 mentions) and cisterns (n = 20 mentions).Desalination (n = 12) is also mentioned as an alternative to surface reservoirs.
Also, discussions regarding infrastructure frequently mention wells, surface reservoirs, and channels (table S2), indicating the significant role decentralized water sources play in the supply system.Interestingly, water management and allocation seem to be increasingly based on 'climate information'or at least stakeholders have discussed more it in the past years (p < 0.01).'Water allocation decisions' and 'water allocation parameters' gained special attention over time (p < 0.01), especially after the 2012-2018 drought started.This reinforces the hypothesis that the occurrence and frequency of hydrological extremes also drive the management mechanisms in Ceará.As expected, topics related to the creation and administration of the committees, including 'participation in national committee meetings' and 'management committees' , became less relevant over time (p < 0.01).

Actors' involvement and social network analysis
To investigate the role actors played in each of the different water allocation periods, we performed a social network analysis.
In the user-based water allocation period (1997-2001), state water companies, state power, and the water basin committees dominated the discourse, and institutions such as agriculture users were not represented.In fact, up until 2006, most of the actor groups were not even mentioned in the meetings.This lack of representation can be explained by the prevailing governance approach at that time; if water is easily available, fewer people tend to be interested in contributing to its management.Alternatively, this can also be explained by the limited number of minutes between 1997 and 2001, as not all committees had been established during that time (0.5% of all sentences).
From 2001 to 2002, the State Water Company initiated an experimental program that, among other measures, enabled small agriculture producers to exchange water rights with large producers during the incurring drought [47].This program opened the possibility of a water marketlike governance mode.In 2005, the second version of the State Water Plan was elaborated by representatives of the State Water Company, the Water Resources Secretariat, and other water-related state agencies.According to the legislation, the plan required approval from the water basin committeesand this can explain the increase in participation and interconnectedness between actors from 2005 onward.
During the scarcity-based stage, especially after the 2012-2018 drought, agriculture and industry users became more relevant.One interesting aspect of Ceará's water management policy is that water availability is conditioned to the negotiated water allocation, i.e. the decisions taken during the meetings can restrict water withdrawals.Consequently, when water quantity diminishes, water users are more likely to be interested in participating in the meetings.
Over the entire period, statutory entities were the actors that could most easily reach all others in the network, as depicted by the betweenness metric in figure 6(B).While the closeness measure is homogeneous for all actors, actors representing the state power and the water basin committee are the only to have a significant betweenness value.This indicates that although participation has increased recently, government entities still control the information flow and strongly influence decision-making during these meetings.
Previous studies in Brazil revealed that committee members perceive an unequal power distribution in water basins where techno-scientific knowledge is more strongly used in decision-making [13,48,49].On the other hand, stakeholders appreciate expert support, indicating a successful trust-building process [14,50].

Relationship between actors and different agriculture, water quality, and climate topics
The analysis of the relationships between actors and the topics (figure 2) showed several significant relationships (table 2).Specifically, topics related to environmental problems showed a strong connection with actors from educational institutions (p < 0.01).This was expected as researchers affiliated with local universities and research institutes have a heightened concern about water quality in semiarid reservoirs, as shown by several studies published in the past years [51][52][53].
International organizations were frequently mentioned together with topics on agriculture water use (p < 0.001) and pesticides (p < 0.001).These associations may be attributed to the stringent requirements imposed by the European Union, Brazil's largest market, to import agricultural products [54].
Finally, climate information seems to be a concern not only to municipal power (p < 0.05) but also to consulting companies (p < 0.05).This does not necessarily mean that they provide this information but that their associated tasks (e.g.water supply projects and environmental diagnosis) are often contextualized with seasonal forecasts and climate projections.

Discussion
The water resources system of Ceará has been historically affected by severe and frequent droughts [55], which encouraged the implementation of several management approaches and instruments, e.g.public participation in water management decisions through basin committees and the negotiated water allocation.In this study, we systematically investigate the shift from IWRM toward adaptive management practices in Ceará by analyzing the minutes of 1020 meetings.
Our results show that discussions on water management seem to be transitioning to a climateinformed approach but still focused on quantity rather than quality.This is reflected by the highly frequent, uniform mentions of topics associated with infrastructure and urban supply and the reduced frequency of topics related to environmental problems (which entails water quality issues).Indeed, water quality concerns such as 'pesticide use' seem only to be discussed in times of crisis (e.g. after the increase in algae bloom episodes in 2015 [56], figure S4).This topic, alongside 'agriculture water use' shows an increased frequency during drought periods (e.g. the 2012-2018 drought).
The increased frequency of 'climate information' in water allocation meetings can not only be associated with the development of climate services in Brazil and the effort by the local meteorology agency to produce seasonal forecasts and monitor drought in the past decades [57] but also with an increasingly level of averseness to risk adopted by water managers [58].
We found that topic distribution across water basins is not uniform-mirroring the needs of each basin.For example, 'reservoir operation' has a stronger focus in the basins where the state's strategic reservoirs are located (i.e. the largest ones).On the other hand, 'environmental conservation and education' is prominent in the 'Metropolitana' basin, which holds the state's capital and is almost entirely urbanized [59].This underscores the need for a flexible, collaborative water management approach.
Furthermore, we learned that although the participation of multiple actors has increased in the past decade, technical knowledge and information flow are still mainly dominated by government entities and representatives of the water basin committees (figure 6(B)).In all committees, independent of the water allocation stage, state actors had higher connectedness and could communicate more easily than other actors.Previous empirical studies have indicated that this is indeed the case: multi-stakeholders can vote on the reservoir operation strategy, but the water availability and allocation scenarios are usually prepared in advance by the water management company and some committee representatives [50].Risk tolerance and the expected future user behavior are rarely agreed upon during the water allocation process.To improve the adaptive capacity of water systems, all organizations and actors must be able to interpret and use technical information [6,13].
It is also worth noting that organizations beyond the water sector (e.g.energy, environment, health) seem to be missing from the participative process.
A cross-sectoral approach to water management is important to ensure water security and reduce the impacts of climate hazards and the uncertainties associated with the dependence on decisions made in other sectors [60].A detachment from other sectors was also observed by the stakeholders of the water basin committees from another Brazilian state [25].To improve collaboration across sectors and align their interests, it is necessary to increase institutional capacity so that actors from different governance levels and sectors work together to develop strategies and measures [60].

Limitations and further research
The proposed NLP-based approach allows for a low-cost and efficient assessment of unstructured data, requiring minimal computer and human The size of the nodes is proportional to its degree.(B) Network centrality measures per year for each organization group.Empty spaces indicate that the corresponding organization group was not identified in the documents in that period.Circle size and color intensity are proportional to the measured value, i.e. organizations with darker, bigger circles have a higher influence on information flow (betweenness), are more well connected to other actors (centrality), and take longer to communicate and spread information (closeness).
resources.Furthermore, it provides a strategy to extract longitudinal information on participatory processes that would only be possible if panel surveys had been performed.Importantly, the adaptability of the proposed approach extends beyond the current context and language, making it applicable to various scenarios.For instance, widely available parliamentary discourse corpora (e.g.[61]) and meeting minutes from water basins in Canada [62] and Uruguay [63] could be used to examine water allocation discussions in other contexts.However, leveraging text information from meeting registers has some drawbacks, such as the intrinsic subjectivity of the minute taker, the inconsistency of organizations' acronyms, and the inability to adequately capture social interactions, non-verbal cues, tone of voice, and cultural references from human interaction [64].Despite these limitations, the documents still provide systematic and reliable information on the meeting content, as it is usually approved by the participants upon publication.
In addition to limitations related to the type of data used, methodological considerations should also be acknowledged.While language models like BERT can effectively learn context from text data, they may inherit biases from their training data [65].Additionally, the interpretation of the topics identified with our method requires \ conducting similar analyses together with local specialists and community stakeholders to ensure comprehensive interpretation.Within this context, the text-based analysis would benefit from combining qualitative research data, including surveys, interviews, or observational studies.For example, the topics identified with our approach could be further investigated with qualitative content analysis and grounded theory.

Conclusion
A key element for the success of water management strategies lies in the consistent evaluation and systematic adaptation of ongoing practices.This should preferably be done systematically rather than organically to ensure that external factors such as political change will not discontinue the process.However, obtaining information on this process is difficult, as continuous surveys or interviews are not readily available or feasible for a long-term assessment spanning multiple years or decades.Valuable information on decision-making, task distribution, and actor engagement are often hidden in minutes of meetings, which nowadays are usually publicly available.Hence, automated analysis strategies such as the one presented here represent a way forward to evaluate participatory approaches and learn about water management strategies.The proposed approach can be adapted to different contexts and has the potential to be combined with qualitative research (e.g. by linking it with grounded theory), aiming at providing context and background to our findings.By considering the case of the water systems in Ceará, Brazil, we evaluated how power networks developed and changed with time and what the stakeholders are worried about.We learned that the management process has become more participative and diverse in Ceará, but it still fails to incorporate actors from sectors other than the water resources and that practical information still mainly originates from the state power and the water basin committee.In this context, moving to a polycentric governance approach might be a way to deal better with the complexity of the water resources system and climate and social uncertainties.
Regardless of the climate and environmental context, our findings suggest that participatory water management processes might fail in sharing information and actively including stakeholders in the decision-making, despite their increased attendance in the meetings.The advantages of a stronger interaction between actors are the development of social capital and the capacity of participants to resolve conflicts through self-established networks.
Our methodology presents an opportunity for stakeholders from the water sector to evaluate the dynamics of the participatory process and to verify which areas or actors deserve closer attention or to be better incorporated into it.Ultimately, it provides a strategy to improve our understanding regarding the implementation of water management strategies.

Figure 1 .
Figure 1.Location of the study area and distribution of the meeting documents in time and space.(A) Total number of documents per basin.Since the committees were created at different periods, there is an uneven frequency of meetings across them.(B) Ceará (dark gray) location in Northeast Brazil.(C) Water basins of Ceará and histograms with the yearly number of documents available in each basin.

Figure 2 .
Figure 2. Visualization of the topics in the minute's corpora.Each dot corresponds to a sentence from the meeting minutes.The colors represent the associated topic.Similar sentences are close together, and dissimilar are apart.

T M Nunes Carvalho et al 3 .
Total frequency of topic mentions in the corpus of meeting minutes, grouped by their main theme.The gray lines represent the maximum (n = 6143) and the average (n = 1403) number of mentions for each topic.

Figure 4 .
Figure 4. Normalized topic frequency per water basin.we represent only the topics for which there is a statistically significantly different topic frequency distribution across water basins ( * indicates p < 0.01 and * * indicates p < 0.05).

Figure 5 .
Figure 5. Normalized topic frequency over time for all committees.The area charts represent topic frequency over time.For each time series, we fitted a regression model to identify the overall trend (dashed gray line).

Figure 6 .
Figure 6.(A) Actors' network graphs across different allocation stages (demand-based, availability-based, and scarcity-based).The size of the nodes is proportional to its degree.(B) Network centrality measures per year for each organization group.Empty spaces indicate that the corresponding organization group was not identified in the documents in that period.Circle size and color intensity are proportional to the measured value, i.e. organizations with darker, bigger circles have a higher influence on information flow (betweenness), are more well connected to other actors (centrality), and take longer to communicate and spread information (closeness).

Table 1 .
Classification criteria of the actors identified in the document.

2 .
Results of the mixed linear regression model fitted to the topics related to agriculture, water quality, and climate.