Citizens' demands for improved farm animal welfare are becoming increasingly urgent worldwide (European Commission, 2023; Buller et al., 2018). It is an important and relevant topic that drives research. For several decades, applied animal scientists have generated a growing body of scientific knowledge about animal behaviour, stress, health, and production parameters related to farm animal husbandry systems. For science-based policy-making, the available scientific information should be compiled. Performing a meta-analysis or systematic review is one established way to synthesise multivariate data. However, those efforts have thus far been performed for and are probably suited to only a limited number of animal welfare topics, e.g. different floor types in beef cattle (Keane et al., 2018), enrichment materials in laying hens (Van Staaveren et al., 2021), or the combination of both for fattening pigs (Averós et al., 2010). Semantic modelling is an alternative approach that is suitable to provide a more generalised synthesis in the new area of integrated animal welfare research. It concerns a procedure to assess overall welfare using a wider range of the available knowledge in a transparent and “best-possible” way based on a systematic decomposition of the literature (Bracke, 2001). Scientific statements typically describe how different treatments affect the welfare of animals. This information can be used to rank (aspects of) housing systems and to identify knowledge gaps. So far, semantic modelling has been applied for overall welfare assessment of housing systems for several species of farm animals including pregnant sows (Bracke et al., 2002a, b), laying hens (Mol et al., 2006; Shimmura et al., 2011), dairy cows (Ursinus et al., 2009), and Atlantic salmon (Tschirren et al., 2021; Stien et al., 2013; Pettersen et al., 2014; Folkedal et al., 2016). In addition, semantic models have been used to assess specific welfare issues, such as the risk of tail biting in pigs (Bracke et al., 2004a, b) and the welfare value of enrichment materials for growing pigs (Bracke et al., 2006; Bracke, 2008).

Semantic modelling of animal welfare differs from semantic modelling in computer science (Rishe, 1992). The former derives its name from the role semantics, i.e. the meaning of words in scientific statements, plays in deriving overall welfare scores. The attention to the meaning of words is also illustrated by the fact that animal welfare is defined as the quality of life as perceived by the animals themselves (Bracke et al., 1999a). This is a feelings-based conception of animal welfare (Fraser et al., 1997). It is implied by the meaning of the word welfare, which refers to “faring well”, i.e. animal happiness (Webb et al., 2019). While feelings define welfare, the biological-functioning conception – and, to a limited, extent also the natural-living conception – identified by Fraser et al. (1997) is used in the semantic-modelling procedures to operationalise welfare assessment. These procedures have thus been developed based on an understanding of the meaning (semantics) of the available science, where peer-reviewed papers provide the norm. Housing and management systems have been studied using animal-based welfare measures. The scientific literature thus describes how environment-based measures affect the animal's welfare (Bracke et al., 1999b, 2001). Based on the systematic decomposition of scientific statements, a semantic model can be generated with a set of welfare-relevant attributes. These are welfare-relevant properties of a housing system. Each attribute has two or more levels and a weighting factor (WF) that allow for the calculation of a weighted overall welfare score (OWSc) on a scale from 0 to 1 as output (for a glossary see Table 1).

Despite the various scientific publications on sematic modelling of animal welfare, further formalisation of scientific information processing is needed to generate better welfare scores. We examined the existing models during the development of the ANyWEL model framework in the InKalkTier project, aiming for integrated sustainability assessment (i.e. complementing welfare assessment with economic and environmental (emission) data (Benthin et al., 2023; KTBL, 2024). The ANyWEL model framework was designed to assess the welfare of any type of (farm) animal. It is described in detail in our companion paper (Part 1; see Benthin et al., 2026). During the development of that framework, we discovered that the basic procedure of statement decomposition as used in semantic modelling was not easy to understand for new modellers. Hence, these procedures needed further explanation and formalisation. Preliminary reliability trials showed a robust intra-modeller agreement with a reliability of about 80 %, while the inter-modeller reliability was substantially lower (about 40 %–54 %; see Supplement File S1). Hence, the reproducibility of semantic modelling (and existing welfare models) may be at stake despite its relevance and robust outcomes in validation studies (Bracke et al., 2002b, 2004b, 2007). A lack of reproducibility can slow down scientific progress (Open Science Collaboration, 2015), erode public trust, and have negative implications for stakeholders and policy-makers (Munafò et al., 2017). This generated a need for more transparency regarding the criteria for the identification of statements, steps for their decomposition, and the weighting of welfare effects. Therefore, the objective of this paper is to provide more clear and formalised guidelines and to take a major step forward by describing in more detail how to process scientific information in semantic modelling to enhance uptake by new modellers and to improve inter-modeller reliability and reproducibility. To this end, we specify the rules, supported by real-life examples, to explain the formalised decomposition and welfare weighting in semantic modelling.

The input of a semantic model consists of linguistically meaningful sections of the scientific literature (i.e. quotes), so called statements (Fig. 1). Statements typically describe the relationship between welfare input and output. The input generally consists of environment-based measures, also called attributes, dependent variables, or experimental treatments. The welfare output mainly refers to animal-based measures, also called independent variables or welfare measures. The if–then rule then typically reads as “If animals are exposed to environmental feature E compared to F then they show a (significant) increase or decrease in welfare measure M, implying that E is better or worse than F for welfare (other things being equal)”. Hence, attributes (e.g. enrichment material or stocking density), attribute levels (e.g. straw as enrichment material or a stocking density of 1 m2 per pig) and their positive or negative effect on the animals' welfare (e.g. as measured by the number of skin lesions) can be extracted from scientific statements (see the glossary in Table 1). So-called weighting categories (WCats) classify different kinds of welfare measures into positive and negative measures. Three weighting categories indicate positive welfare, namely natural behaviour, preference, and (measures of) demand, whereas nine weighting categories denote reduced or negative welfare, i.e. (measures of) pain, illness, fitness, survival, aggression, stress physiology (HPA (hypothalamic–pituitary–adrenal) axis activation, SAM (sympathetic adreno-medullary) axis activation), abnormal behaviour, and frustration and avoidance (Table 2) (Bracke et al., 2002a).

To be able to calculate the weighting factor (WF) for each attribute, so-called weighting-category level scores (WCatLevScs) come into play (Tables 2 and 3). The WCats indicating positive welfare have positive WCatLevScs, while the negative WCats have negative scores. Moreover, the 12 above-mentioned WCats have been classified into relatively low-impact and high-impact categories based on their ability to reflect a lower or higher impact on the intensity, duration, and incidence as indicated by a welfare measure. The lower-impact WCats (i.e. presumed to be somewhat less important) use a WCatLevSc scale from ±1 to 3, the higher-impact categories (presumed to be more important) use a scale from ±1 to 5 (Table 3). Here, both the lower- and higher-impact WCats have three primary levels (i.e. respectively, ±1, 2, 3 or ±1, 3, 5) and a number of other possible scores besides these primary ones, especially at the low end of the scale from 0 to 1; see Part 1: Benthin et al., 2026). The higher-impact negative WCats are pain, illness, survival, frustration and avoidance, and HPA axis activation. They receive primarily negative WCatLevScs of -1, -3, or -5. Demand is the only higher-impact positive WCat. It generates positive scores of +1, +3, or +5. By contrast, lower-impact negative WCats that generate a negative primary WCatLevSc (i.e. -1, -2, or -3) include fitness, SAM axis activation, aggression, and abnormal behaviour. Natural behaviour and preference are lower-impact positive WCats that generate WCatLevScs of +1, +2, or +3 (Table 3; Bracke et al., 2002a). The WCatLevScs serve as input for the calculation of WFs for all of the attributes in the model. WFs are derived from the difference between the (sum of the) maximum positive and (the sum of the) minimum negative WCatLevScs for each WCat assigned to the best and worst levels of each attribute in the model, respectively (Table 1; see also Part 1: Benthin et al., 2026, Sect. 3).

Prior to calculating the WFs of attributes, at least two attribute levels, i.e. a best and worst level, need to be defined for each attribute in the model. Attributes need to apply across all housing systems within the assessment domain, and their respective attribute levels need to have a science-based ranking. For example, the attribute of floor type can have levels like fully slatted concrete floor, partly slatted concrete floor, and bedded floor. These levels are assigned only when scientific findings report differences, e.g. with respect to lameness, between these types of floor, such that one type is better than another. When the scientific evidence is inconclusive for a specific attribute, one or more attribute levels should be disqualified or combined to ensure that all attribute levels and their rankings have a scientific basis. For example, when the scientific evidence is inconclusive about whether “50 % slatted floor” is better or worse for animal welfare than “25 % slatted floor”, these attribute levels could be combined into the level “partly slatted floor” or “25 %–50 % slatted floor”. Subsequently, attribute-level scores (ALevScs) can be determined for each attribute by following the science-based ranking of their levels. For example, different levels of the attribute of floor type can be ranked from worst to best based on scientific studies reporting the short- and long-term effects of the different types of floor on animal welfare.

Finally, overall welfare scores are calculated as weighted-average ALevScs considering the fact that, together, all attributes in the model cover all welfare needs of the species of interest and all possible housing systems in the assessment domain (Bracke et al., 2002a; Fig. 1). A detailed explanation of the calculation steps for the overall welfare score is provided in Benthin et al. (2026).

Ultimately, the modelling principles and procedures developed for semantic modelling of animal welfare have been designed to generate the best-possible overall-welfare assessment based on all available information (at a certain point in time). This is achieved by systematically linking (as much as possible of) the available scientific literature about the animals and their living conditions (Bracke et al., 2002a). Interestingly, semantic models are not designed to be static constructs. They are intended to be updated periodically with the most recent scientific findings, and recent developments in AI are likely to facilitate this (Zhang et al., 2026).

In semantic modelling, a relational-database structure was developed with tables listing the scientific references, the scientific statements, and information about the location of the statement within the scientific reference in addition to the attribute(s); the attribute levels; the types of welfare measure; the associated WCats and WCatLevScs; and perhaps a note on, for instance, the interpretation of the statement (Bracke et al., 2001; Anonymous, 2001; see Supplement File S2 for an excerpt from a database file, and see Benthin et al., 2023). This enhances control over the data and allows for periodic updating when new information becomes available (Bracke et al., 2006). The first step in building such a database is to collect scientific statements for the species of interest within the defined assessment domain (i.e. range of housing systems) that meet the modeller's objective (Fig. 2). Literature sources include primarily peer-reviewed publications reporting empirical studies (including both original works and review papers) and scientific reports (e.g. expert-panel reports from the European Food Safety Authority). These sources can contain information on welfare and, in particular, describe relationships between environment-based treatments on the one hand and the behaviour, health status, and/or (stress-related) physiology of the species or animal type of interest on the other hand. Search criteria for literature should be clear, e.g. in accordance with PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analysis; Page et al., 2021) as far as possible. Sources should preferably be available in the form of full text and should report statistically significant findings (Bracke et al., 2006). Papers published under an open-access licence make such sources more readily available (Muñoz-Tamayo et al., 2022). Also, other sources, such as textbooks, industry reports, and magazine articles, may be used if deemed appropriate, i.e. sufficiently trustworthy. In general, however, political or technical recommendations and legal standards should not be selected as they may not primarily rely on science. To fill knowledge gaps in the scientific literature, it is also possible to include statements reflecting expert opinions as another potential source of information (e.g. about new innovative housing concepts; see Part 1: Benthin et al., 2026).

The next step in semantic modelling is to identify from those literature sources the statements that describe the impact of experimental treatments on animal welfare. There is no minimum number of statements to extract; the modeller selects all statements that seem to be relevant and are feasible to process. The scientific statements are used to construct the welfare model by decomposing them into if–then relationships between the attributes of housing systems and reported welfare measures (Bracke et al., 2001). To extract the relevant information for semantic modelling, a semantic analysis, also called a decomposition of a statement, is needed (see also Sect. 4, “Decomposing scientific statements”). This is a language procedure that breaks down the meaning of phrases (i.e. statements) into segments based on logic. The criterion that flags a scientific statement as potentially worth considering for decomposition is the description of a positive or negative effect on the welfare of the relevant species of some property (attribute level) of a housing system within the assessment domain (Fig. 2).

For illustration purposes, here, we provide two examples of scientific texts that contain statements to be selected for making a welfare model. Example 1 is as follows:

In the farrowing pen, sows normally are housed individually in crates with a partly slatted floor with a heated resting area (solid floor with some substrate) for the piglets of at least 0.6 m². In preference tests and operant conditioning tests, sows were highly motivated to farrow in pens with straw. (cited from Anonymous, 2001, p. 26)

Once a scientific statement is deemed to be suitable for decomposition, the first step is to identify (i) the attribute, (ii) the levels of the attribute that are being compared, and (iii) the welfare effect that is being reported. The relationship between the attribute level and the welfare effect can be identified by transforming the original statement into an if–then rule (Bracke et al., 2001, 2002a; Fig. 2). In this procedure, the meaning of both the original statement and the if–then statement remains logically equivalent. The if–then statement can be formulated as follows:

If animals (i.e. the species of interest) are subjected to attribute level L1 of attribute A as compared to its level L2 then there is a (significant) effect (i.e. increase or decrease) on welfare measure M that belongs to weighting category WCat and has a magnitude of WCatLevSc S. (see Sect. 4.2 in Part 1: Benthin et al., 2026)

Sometimes the welfare effect in the statement appears to be mismatched with the positive or negative WCat as defined previously (Bracke et al., 2002a). For example, a statement may present a positive welfare effect (e.g. improved claw health in treatment L1 compared to in L2), whereas the available WCat (pain) describes only negative effects (e.g. increased lameness). The solution here is to flip the statement around as it were (i.e. negate it) when the if–then rule is deduced from the original statement. Thus, the seemingly positive effect mentioned in the original statement (i.e. improved claw health in L1, which, in fact, was only a reduction in negative welfare) is converted into a negative effect in the if–then statement (i.e. increased lameness in treatment L2 compared to in L1). This negative effect can then be classified (and scored) using the negative WCat of pain. The following example illustrates this procedure again with treatment L1 being “grazing on pasture” and L2 being “no grazing possibility available”:

Grazing is known to benefit claw health (less severe disorders and better recovery) and to reduce stereotypies and aggression in the herd. (cited from Cornelissen at al., 2009, p. 7)

Decomposing a scientific statement through the conversion into an if–then statement helps to identify the attribute levels, as well as their welfare rank (and their attribute level scores, ALevScs). Subsequently, a WCatLevSc can be assigned by classifying the reported welfare measure and assigning a score for the degree to which welfare was affected according to (the meaning of) the scientific statement. The WCatLevScs are assigned by the modeller based on the incidence, duration, and intensity of the reported welfare effect, i.e. how many animals are affected, for how long, and how severely (Willeberg, 1991; Fig. 2, Tables 2 and 3). As mentioned earlier, there are positive and negative WCats that have either a relatively low or high impact (with primary WCatLevScs of ±1, 2, 3 and ±1, 3, 5, respectively; Bracke et al., 2002a). Modellers should take note of the use of specific wordings in a statement (e.g. chronic or severe) that say something about the incidence, duration, and/or intensity of the welfare problem when assigning a WCatLevSc. By default, a score of ±1 is given for statistically significant (or equivalent) findings. A higher score of ±2, 3, 5 is used for a strong(er) welfare effect, as indicated by the welfare measure (i.e. classified as a relatively low- or high-impact WCat) or based on the use of specific terms or phrases in the scientific statement that is being analysed. We listed examples of such phrases that are typically associated with a WCatLevSc of ±1 and those associated with the higher scores ±2, 3, 5 (Table 3). The example phrases mentioned in the third column of Table 3 (i.e. phrases like “chronic”, “severe”, and “highly”) are used to generate higher WCatLevScs based on the general rule of thumb that the authors of the statements have used these words only when justified. Consequently, these “power terms” are taken to indicate a strong(er) positive or negative impact of a certain attribute level on the animal's welfare. Besides specific phrases, modellers can also make use of numerical data about the incidence, duration, and/or intensity of a welfare problem. The following example illustrates this:

[adventitious] Bursitis was found to affect 95.5 per cent of pigs reared without straw compared with 3.75 per cent of pigs reared with straw (Smith & Smith 1990). (cited from Arey, 1993, p. 240)

Here, almost all pigs were affected by a condition causing lameness when housed on concrete floors compared to straw bedding; hence, the attribute level of concrete floor for the attribute of floor type would receive a maximum WCatLevSc of -5 for the relatively high-impact WCat of pain.

Besides the two standard scales (plus or minus 1, 2, 3 and 1, 3, 5) for the two kinds of WCats (i.e. with a relatively low and high impact, respectively), we suggest using a lower proxy score of ±0.5. That score should be used for uncertain statements based on weak phrasing (Table 3) or tendencies to reduce the impact of publication bias (Benthin et al., 2026). This may be useful since there is discussion within the scientific community about null-hypothesis significance testing based on arbitrary P-value thresholds and the use of the term “statistical significance” (Muff et al., 2022). Furthermore, a score of 0 was used for statements that needed to be discarded because they had been updated by new statements (such as an expert opinion on an innovative new housing system that is replaced by new scientific findings). Statements may also be discarded from the calculation of WFs when they end up not having a relation to any of the final attribute levels in the model. Moreover, when WCatLevScs are assigned, it is important that all attribute levels within the statement receive a scoring. Thus, the potentially best and worst level of the attribute must receive, respectively, a minimum of one positive and one negative WCatLevSc (Fig. 2). This is a practical requirement to enable the calculation of the overall welfare score (OWSc) using the semantic model (Bracke et al., 2002a; Part 1: Benthin et al., 2026).

An attribute level can receive several WCatLevScs based on one or more statements, which affect the WF for that attribute. In case two or more statements report opposite welfare effects, the WCatLevScs could cancel each other out. In addition, the welfare effect of the treatments (attribute levels) compared within the scientific statement may not always be clear by means of the specified WCats. For instance, we present the following statement: “In preference tests and operant conditioning tests, sows were highly motivated to farrow in pens with straw” from example 1. Here, the word “highly” specifies a strong welfare effect. In addition, welfare measures from both preference and motivation tests (stated also in plural) are mentioned. Hence, a higher WCatLevSc of +2 for the WCat of preference (with the scale +1, 2, 3) and +3 for the WCat of demand (with the scale +1, 3, 5) can be assigned to the attribute level “straw available” of the attribute “type of nest-building material”. Since feed is considered to be the gold standard in motivation tests (Dawkins, 1988) and because the scientific statement does not compare access to straw with access to feed, no maximum WCatLevScs (i.e. +3 and +5, respectively) are assigned. In this statement, the attribute level “straw available” is compared to the level “no straw available”. However, the statement (example 1) does not report a negative welfare effect of not having straw available in the farrowing pen. This attribute level (no straw available) could potentially be the worst level of the attribute “type of nesting material”. In the process of decomposing statements, it may not be known whether more statements will be found about these attribute levels. Therefore, we suggest assigning a WCatLevSc of ±0.01 (i.e. near to 0) to ensure that there is always a difference between the best and worst level of each attribute (see also our companion paper by Benthin et al., 2026, and Table 3). By assigning a WCatLevSc close to 0 (i.e. 0.01) to the complementary attribute level rather than not having assigned any value, the semantic model as implemented in a relational database can still calculate the WFs of all attributes. Consequently, the attribute level “no straw available” in this statement receives a score of -0.01 for the WCat of frustration and avoidance (i.e. assuming that nest-building behaviour is blocked in sows when no nest-building material is available). In the process of further statement decomposition, the modeller is very likely to encounter one or more other statements that more clearly show that the attribute level “no straw available” can cause frustration in farrowing sows. Then, the -0.01 score is overruled by higher (more negative) WCatLevScs. An example is the next statement:

At farrowing, sows are normally confined in crates without bedding. These conditions can lead to acute signs of stress as shown by both the sow's behaviour (Baxter, 1982; Schouten, 1987) and physiology (Baxter and Petherick, 1980; Metz and Oosterlee, 1981; Kilgour and Dalton, 1984; Vestergaard, 1984). Much of the frustration may be caused by the lack of materials to build a nest. (cited from Arey, 1993, p. 238)

Once all of the selected sources have been read and all of the relevant statements have been identified and decomposed, the names and descriptions of the attributes and attribute levels are to be revisited to finalise the database. The modeller needs to ensure that multiple statements referring to the same attribute and/or attribute level are labelled with identical names. This will consolidate the science-based ranking of attribute levels, including the best and worst level of each attribute, such that attribute level scores and WFs can be assigned properly. For the ranking of attribute levels, the modeller has to review the WCatLevScs of each attribute level and (i) sum up the lowest WCatLevScs for each negative WCat and (ii) sum up the highest WCatLevScs for each positive WCat. In this way, the modeller can rank the attribute levels of each attribute from the best to worst level. Here, the best level has the highest sum of the most positive WCatLevScs (and, if any, also added to that the most negative WCatLevScs) per WCat. The worst level has the lowest sum of the most negative WCatLevScs per WCat (and, if any, also added to that the most positive WCatLevScs; see also Part 1: Benthin et al., 2026). Afterwards, the model can be used to calculate the overall welfare scores for housing systems within the assessment domain, e.g. to generate welfare scores of benchmark systems (see Benthin et al., 2026, and Bracke et al., 2002a, for this procedure).

Overall, statements collected from the scientific literature and, if needed, from interviews with experts are key for semantic modelling of animal welfare. To enhance the scientific reproducibility and transparency of overall welfare assessment using a semantic welfare model, we formalised (i) the relevant scientific knowledge sources, (ii) the criterion that flags a potentially meaningful scientific statement for semantic modelling, and (iii) described the if–then procedural rules for the decomposition of statements. This may further explain why we believe semantic modelling is bio-logical: it is based on an understanding of (the semantics of) the biology of the animals, as well as on logic. As to logic, semantic modelling implies something like this:

If n scientific statements (S1, … Sn) are known then, since S1 is true, S2 is true, … and Sn is true, for each S there is an attribute level L1 that is better for welfare than L2 as measured by welfare measure M, and if all of that is the case then we can derive i attributes (A1, … , Ai), each with ranked levels L_worst … L_best and weighting factors (WF1, …, WFi), such that overall welfare scores (OWScs) can be calculated for all housing systems in the assessment domain based on the available science (i.e. the collected statements) by calculating weighted average attribute level scores (which are based on the attribute level ranks and on the intensity, duration, and incidence of the reported welfare effects).

This work provides a set of essential formalised rules regarding how scientific literature can be used to build a semantic welfare model by spelling out the process of semantic modelling. We specified how scientific information can be retrieved, recorded, decomposed, and weighted using semantic-modelling principles. The explanations and recommendations presented in this paper's guidelines could be an important step forward in making the use of scientific literature for integrated welfare assessment more transparent and standardised at the inter-modeller scale.