Public avalanche forecasts, as provided in many mountainous regions, inform readers about snow and avalanche conditions at a regional scale. To effectively communicate different components of the avalanche situation, forecasts often describe expected conditions systematically, using a hierarchical information pyramid communicating the current regional danger level, the most avalanche-prone locations (slope aspects and elevation range, also called the core zone), and the dominant avalanche problem as well as narrative text, a danger description Fig. a and b; e.g.,, or snowpack and weather summaries. Although the communication of danger levels, corresponding core zones, and dominant avalanche problems is often standardized using common terms or icons (e.g., in Europe: five danger levels, , and five avalanche problems, ), the degree of detail and the use of text and graphics vary considerably between forecast products issued by different (usually national) forecasting centers . Important information describing the severity of the avalanche conditions, such as likely triggers required to release an avalanche, the frequency with which such triggering spots will be found in a region, the specifics of the likely locations of these triggering spots, and the expected avalanche sizes, is communicated in various ways (Table ). Even though this information defines the avalanche danger level, its communication is much less standardized. In North America, this information is often provided through a combination of keywords and graphics. In Europe, however, narrative textual danger descriptions are much more common (Table , Fig. b), though in Norway a tabular format has been adopted .

Public avalanche warning services in Europe aim to provide efficient and effective forecasts to their users e.g.,. To achieve this goal, forecasts in general must provide credible information of value to the user e.g.,. Value, however, is directly influenced by forecasts being consistent and accurate . In recent years, this fact has increasingly been recognized in avalanche forecasting. To date, the most effort exploring quality and consistency of avalanche forecasts has focused on the forecast danger level e.g., and more rarely on other forecast elements, such as the avalanche problem . Little research has explored either consistency or quality of icons or the text used by forecasters when assessing avalanche danger. For instance, when asking avalanche professionals to assign a probability to the meaning of the five classes describing the likelihood of avalanches or to rate the size of observed avalanches , considerable variation in responses was noted. Most studies have treated the forecast as a product and, for example, explored usability by testing whether users were aware of different elements of the forecast or whether users understood the information presented . tested the comprehension of text, icons, and pictures among Norwegian avalanche bulletin users and noted that ability to comprehend the information provided in the forecast depends on the competency of the user and the complexity of the avalanche scenario. Recent work in North America used interviews to develop a typology of avalanche forecast users' competency with respect to forecast content . In a similar vein, surveyed the literacy of forecast readers with respect to standardized terms and icons used in North America (such as listing or ordering the danger levels, identifying avalanche problem icons, or applying information in a slope-choice experiment). explored the characterization of the severity of the avalanche problem in Canadian avalanche forecasts, described by the likelihood of avalanches (comparable to the combination of snowpack stability and the frequency distribution of snowpack stability in Europe) and avalanche size for each avalanche problem type separately, in relation to the avalanche danger rating. noted variations in the way the same avalanche danger level was characterized between different avalanche problems.

However, to our knowledge, there has been no systematic exploration of how forecasters actually describe the avalanche conditions in the narrative part of avalanche forecasts, despite such textual descriptions being the main way of communicating the factors describing the severity of the expected avalanche conditions in many forecast products in Europe (Table ).

Standards used in avalanche forecasting, including the avalanche danger scale Table ;, the avalanche problems , the avalanche size classification , or the conceptual model of avalanche hazard , make use of specific terms to describe the stability of the snowpack (or what it takes to trigger an avalanche), the frequency distribution of snowpack stability (or how frequent triggering spots are), and what the expected avalanche size and hence the damage potential are. Nonetheless, these descriptions often include undefined, ambiguous, or hedged statements allowing considerable room for interpretation by both forecasters and forecast users. Furthermore, there are many possible ways for forecasters to incorporate individual terms into the narrative texts often used to communicate details of avalanche danger in Europe (Table ).

Thus, to answer the question of how avalanche danger is described in textual descriptions in avalanche forecasts, focusing on the perspective of the forecaster, we arrive at our first research question: (1) how well do forecasters agree on the meaning of terms characterizing triggers required to release avalanches, frequencies of triggering spots, and expected avalanche sizes? Our second research question builds on this characterization, and asks the following: (2) how does the use of language in danger descriptions relate to avalanche danger? To address these questions, we explored danger descriptions, written using a controlled-language environment referred to as the catalogue of phrases;, and avalanche danger published in more than 1000 avalanche forecasts by the national avalanche warning service in Switzerland during eight forecast seasons. The use of the catalogue of phrases is important as it impacts the forecast product since all forecasters use the same set of words. It also impacts the analysis as the number of possible combinations of words and phrases, though large, is finite and known. We analyzed these texts in a two-step process: first, starting from an iterative annotation process , we annotated textual phrases relating to avalanche danger for further analysis. In a second step, Swiss avalanche forecasters related these text phrases directly to contributing factors of avalanche danger (RQ1). Building on this classification, we then extracted and analyzed text used in published danger descriptions and related it to the reported avalanche danger levels (RQ2).

In Switzerland, the national avalanche warning service at the WSL Institute for Snow and Avalanche Research (SLF), Davos, is responsible for the publication of avalanche forecasts covering the Swiss Alps and the Jura mountains.

Avalanche forecasters in Switzerland use definitions and guidelines provided by the European Avalanche Warning Services (EAWS) when assessing and communicating avalanche danger. Of particular relevance is the European Avalanche Danger Scale EADS;Table , which qualitatively describes the five danger levels (1 (low), 2 (moderate), 3 (considerable), 4 (high), 5 (very high)) in terms of snowpack stability and the likelihood of triggering. The EADS links triggers typically required to release avalanches, the number of potential triggering spots or of avalanches, the probability of avalanche release, and the potential size of avalanches to the ordinal danger levels.

Swiss avalanche forecasts describe expected regional avalanche conditions, always communicating the highest danger level expected during the forecast period . In Switzerland, dry-snow conditions, where they exist, are essentially always summarized by a danger rating, with the danger description describing these dry-snow conditions and relevant avalanche problems. Wet-snow conditions, on the other hand, are often mentioned only as a secondary problem. This in turn means that danger resulting from secondary wet-snow problems is at most as high as the danger level communicated with the primary problem but may often be lower. A danger rating referring to wet-snow conditions is only given if the danger of wet-snow or gliding avalanches exceeds that of dry-snow avalanches.

To communicate spatial variation in avalanche conditions, the forecast area is divided into almost 150 warning regions, the smallest spatial units used in the forecast. These warning regions are aggregated flexibly to danger regions, which are characterized with the same danger level, avalanche problems, avalanche-prone locations, and danger description (see example in Fig. b, which shows the forecast components describing avalanche danger in a danger region). Each danger region contains a textual description of avalanche conditions. In Switzerland, the severity of the avalanche situation is described exclusively in the danger description (see also Table  for a comparison with other forecasts). The trigger, the frequency of triggering spots, and avalanche size described in the text may refer to a specific avalanche problem (as in the example in Fig. b) or to the avalanche conditions as a whole.

Avalanche forecasts are published in the evening at 17:00 CET and are valid until the following day at 17:00 CET During the main winter season forecasts are also updated at 08:00 CET, remaining valid until 17:00 CET on the same day.

Since November 2012, the text of Swiss avalanche forecasts has been prepared using a controlled-language environment, relying on a so-called catalogue of phrases – a collection of predefined sentence templates, permitting instantaneous, automatic translation of the German text into French, Italian, and English. The aim of such a catalogue of phrases is, for one base language, to allow generation of naturalistic texts using a restricted lexicon, base syntax, and semantics . The resulting texts are understandable by speakers of the base language, and as shown by , speakers could not distinguish between danger descriptions written using the catalogue of phrases and freely written danger descriptions. Using this catalogue of phrases also makes automatic translation straightforward, and recently the catalogue has been extended to Spanish and Catalan (since winter 2020/21).

The catalogue of phrases consists of a number of sentences, each containing a number of phrase_options. Phrase_options contain either values, the actual textual content, or up to two additional levels of phrase_options, thus allowing an enormous number of possible combinations. An example, illustrating the creation of a single sentence, is shown and explained in Fig. .

We analyzed 1286 map-based avalanche forecasts published at 17:00 CET during eight winter seasons between 27 November 2012 and 13 February 2020. A total of 5897 danger regions were described by a danger level, an avalanche problem, aspects and elevations where the danger prevails, and a danger description.

For this study, we extracted the forecast danger level (element 1 in Fig. ) and the respective danger description related to the major problem being either dry-snow conditions or wet-snow or gliding avalanches (elements 4 and 3 in Fig. , Table ).

We worked with German danger descriptions as German is both the working language of forecasters in Switzerland and the base language used in the creation of the catalogue of phrases. Since the danger descriptions, written with the catalogue of phrases, are a form of controlled natural language , they can be analyzed using standard natural language processing methods. As we are working with a finite corpus of phrases, and we are interested in how those meanings or phrases can be and are interpreted (the semantics), we manually annotated textual elements in the catalogue of phrases . We followed the iterative annotation proposed by , which they describe as the model–annotate–model–annotate (MAMA) cycle.

Our starting point was the catalogue of phrases and its 9989 unique values. As only some values contain information related to the three key factors characterizing avalanche danger – type of trigger, the frequency of triggering spots, and avalanche size – a two-step approach was used to annotate the text.

First, rules were defined (step 1 in Fig. ) before individual values in the catalogue of phrases were assigned to one of the three key factors in an iterative annotation process using these rules (Sect. ). This initial step allowed us to retain only values contained in the catalogue of phrases judged to describe one of the three key factors, considerably reducing the amount of labeling to be done in the following step. Second (step 2 in Fig. ), the values which characterize the key factors were assigned ordinally ranked thematic labels according to current practice and/or suggested labeling used in avalanche forecasting or recent research (Sect. ). These two steps allowed us to answer research question 1: how well do forecasters agree on the meaning of terms characterizing triggers required to release avalanches, frequencies of triggering spots, and expected avalanche sizes?

Finally, danger descriptions and the labels assigned to them were analyzed with respect to avalanche danger (step 3 in Fig. , Sect. ), permitting the exploration of research question 2: how does the use of language in danger descriptions relate to avalanche danger?

Our approach aimed to better understand how words and phrases from a structured catalogue are used to convey avalanche danger. In contrast to the relatively small number of other studies which have concerned themselves with the communication of avalanche danger through forecasts e.g.,, our starting point was to explore how forecasters interpret (RQ1) and use narrative text to convey avalanche hazard (RQ2). We took advantage of a unique data set to perform our analysis: avalanche forecasts written over eight winter seasons using a structured catalogue in Switzerland. To discuss our results, we introduce here the semiotic triangle , a concept commonly used in linguistics and cartography to understand the relationships between a

referent, an instance in the real world, in our case a (partially observable) avalanche situation;

To answer the first research question (how well do forecasters agree on the meaning of terms characterizing triggers required to release avalanches, frequencies of triggering spots, and expected avalanche sizes?), we asked forecasters to assign labels to text values available in the catalogue of phrases to describe these factors. This annotation process was a necessary step in exploring our second research question since we needed these labels to understand how avalanche danger was described by forecasters. However, equally importantly, it gave us insight into the degree to which a trained team of forecasters uses language to describe different characteristics of avalanche danger. Since the task was performed in isolation – that is to say forecasters classified terms independently of a specific avalanche situation – it relates to one side of the semiotic triangle: the relationship between the symbol (the language used to convey a situation using the varied options available in the catalogue of phrases) and thought (the abstraction of an avalanche situation described using a small number of key factor labels). Although the overall agreement in the assigned labels between forecasters was rather high (Fig. ; κ>0.64), with 50 % of the text symbols being assigned to the same class by all forecasters, it is important to note that these values are based on expert annotation by a team working together on a daily basis.

Zooming in on the individual classifications, it is possible to identify areas for discussion in the forecasting team with regard to three issues. First, the terms used most consistently were those taken directly from definitions. For example, there was 100 % agreement about the use of terms used in the definition of avalanche size classes (e.g., small avalanche for a size 1 avalanche or very large avalanche for a size 4 avalanche). Second, other terms, especially those which hedge, were considered more ambiguous by the forecasters, with for example rather small avalanches being considered by four forecasters to be representative of size class 1 and by four forecasters to be representative of size class 1–2 avalanches. This difference matters since size 1 avalanches are typically not associated with burials, while size 2 may “bury, injure or kill a person” and as size 2 avalanches more often lead to burials of recreationalists in Switzerland. Third, we also identified a number of terms present in the structured catalogue which were never used by the forecasters. In general, annotating and assigning words and phrases to particular situations gave valuable insights into the ways in which avalanche forecasters describe avalanche situations and help identify areas where consistency could be improved.

Our second research question asked how the use of language in danger descriptions relates to avalanche danger. The answer to this question provides knowledge as to how forecasters take a referent, in this case the expected evolution of the avalanche situation over the next 24 h, and represents it through language. The annotations of words and phrases used in the avalanche forecast allow us to first characterize how avalanche danger is described and second explore the consistency of descriptions of similar avalanche danger.

The description of the three elements characterizing the danger level – the trigger required to release an avalanche, the frequency of triggering spots, and the expected avalanche size – varied in their degree of completeness. Danger level 2 (moderate) regarding avalanche danger in dry-snow conditions was characterized by all three factors only 21 % of the time, and 30 % of descriptions only mentioned one factor (most often the frequency and location of the likely triggering spots). Since in Switzerland many avalanche accidents happen at this level of forecast avalanche danger e.g.,, characterizing the likely consequences and triggers of these avalanches more often may be useful in communicating the situation. For danger levels 3 (considerable) and 4 (high), the frequency of triggering spots was missing about half the time.

These distributions of missing information are clearly not random and reflect systematic choices made by forecasters in translating the avalanche danger (referent) to a danger description (symbol) through a thought process unknown to forecast users. It appears that the cases where information is missing are those where values would likely describe the middle ground rather than the extremes. Since this middle ground may in practice be where the interpretation of avalanche forecasts is more difficult for a recreationalist, and given that avalanche danger definitions include all three factors at all levels of avalanche danger, it is important to consider further the likely influence of missing information on users.

Irrespective of whether factors are described in a forecast, it is also important that the factors included are used consistently. In general, we found this to be the case, and the description of the elements characterizing avalanche danger changed significantly from one danger level to the next. As shown in Sect.  (Table ), dry-snow and wet-snow avalanche conditions were described differently: natural avalanches are essentially always mentioned in danger descriptions describing wet-snow or gliding avalanches, regardless of danger level, while in dry-snow conditions they are primarily mentioned at danger level 3 (considerable) or higher. Differences also exist regarding the largest expected avalanche size: these tended to be larger for wet-snow than for dry-snow avalanche conditions. For instance, for cases when avalanche size was described, size 3 avalanches were the most frequently expected avalanche size at 2 (moderate) in wet-snow conditions (55 %) and at 3 (considerable) in dry-snow conditions (50 %). These findings vary from the definitions given by the EADS, which does not distinguish between avalanche sizes expected in dry- or wet-snow conditions. They do, however, correspond well with a study exploring a large data set of avalanche occurrence data in the region of Davos (eastern Swiss Alps), which showed that the largest observed avalanche was larger and that the number of natural avalanches was higher for wet-snow avalanches compared to dry-snow avalanches on days with the same forecast danger level . Although this means that the description of the forecast corresponds to observations, it also highlights an inconsistency in the application of the danger levels allowing more natural avalanches at a larger size in wet-snow conditions than dry-snow conditions. This may also be one explanation for variations in the use of the danger levels in Switzerland compared to, for instance, its neighbors in Italy .

We analyzed the text describing the expected avalanche conditions in almost 6000 danger descriptions, written relying on a catalogue of phrases, published in the public avalanche forecast in Switzerland. We focused specifically on three factors described in the textual danger description – the type of trigger required to release an avalanche, the frequency of potential triggering spots, and the expected largest avalanche size – and their relation with the issued danger level. To conduct this analysis, the Swiss avalanche forecasters assigned categories to individual terms used in the danger description. Although the agreement in the labeling was rather high – 50 % of the terms were assigned to the same class by all eight annotators – not all terms could be assigned to a specific class by some forecasters.

When we linked the factors used in danger descriptions to avalanche danger we found that, especially for avalanche danger level 2 (moderate), only 21 % of descriptions used all three factors, and 30 % of descriptions mentioned only a single factor. Furthermore, avalanche size classes are used differently to describe dry-snow and wet-snow or gliding avalanche conditions. The results highlight the demand to review and harmonize the use of terms to describe the trigger required to release an avalanche, the frequency of potential triggering spots, and the expected largest avalanche size and their relation to the danger level. Since our approach is data-driven, it provides very clear pointers as to terms which are used inconsistently or not at all by forecasters and thus gives a valuable framework for forecasting services in reviewing the quality and consistency of written forecasts. However, we focused exclusively on the perspective of Swiss forecasters working in German. Our results cannot be directly transferred to other forecasting services and languages, and the analysis was greatly simplified by use of the sentence catalogue used to write Swiss avalanche danger descriptions.

Future work should also explore the perspective of the user of the avalanche forecast. Are danger descriptions the best possible way to communicate important pieces of information including avalanche size or the occurrence of natural avalanches? Do users interpret this information in similar ways to forecasters?