We present first results of an analysis of a corpus of linguistic descriptions that were collected in controlled experiments. This corpus and its analysis add to the body of knowledge on formal models for spatial language, language interpretation and generation. The experiments are grounded in qualitative formalisms (RCC and Intersection Models, IM) that have a long standing tradition as means to bridge formal and linguistic descriptions of space and spatial relations. Our experiments address dynamically changing spatial relations (movement patterns/geographic events). By keeping the formal spatial characterizations identical across experiments but changing the semantics (that is, we used movement patterns across seven different geographic domains such as a hurricane in relation to a peninsula, plus two geometric figure domains) we contribute to disentangling spatial and domain specific aspects of spatial (event) language. We briefly discuss here two aspects: First, we hand examine the corpus by selecting participants that show the same conceptual behavior as identified through RCC/IM; second, we analyze the domain specific sub-corpora to address similarities and dissimilarities between individual domains.
Formal models of spatial language play an import role in several disciplines addressing
questions of (natural) language processing, natural language generation, the automatic description
of spatial scenes, or the design of unifying frameworks for multimodal information systems
and processing [
We have developed an experimental paradigm that allows us to evaluate the influences of
domain semantics on the conceptualization of movement patterns as well as how movement
patterns are linguistically described. Here we focus on the linguistic descriptions. Our framework
is based on a topologically defined conceptual neighborhood graph [
With respect to the spatial language we find very diverse ways of conveying spatial information. We do believe that this diversity is fostered by the fact that our research is addressing geographic events / spatio-temporal information (rather than static spatial relations). Especially in the hurricane example we find the following strategies: relative reference frames focusing on the end relations of the geometric characteristics of figure and ground; qualitative distancebased descriptions; negation of what the path does not do; absolute reference; (experiment) context specific descriptions; explicit topological descriptions; intrinsic reference induced by the movement. Interestingly, the explicit spatial descriptions in the lake scenario seem to be less varied, indicating a potential difference between scaling and translation movement patterns that are indistinguishable from a topological perspective.
In both scenarios we also find descriptions that are encoding the spatial event in terms of domain specific language (to different degrees). While, for example, a statement such as “no hit”, “weak hit”, “no landfall” are still rather explicit, a statement such as “weak hurricane” relies heavily on background knowledge of a scenario and is open to interpretation. In case of the lake, the descriptions are much less varied, again, and in most cases refer to a flood not happening.
These findings led us to explore differences between the sub-corpora (the nine different
scenarios). First, some domain-corpus properties can be extracted using AntCont [
From the tag clouds in Figure 2, we can see that top frequent words are mostly related to domain specific semantics. For example, city, desert, and island are referring to objects illustrated in each scenario. It is not surprising that participants make use of the domain semantics for reference to objects in the scenario, as it is a direct and succinct way to describe an object and distinguish it from surroundings. However, for the topological change depicted in different icons, participants would have to use more complicated descriptions such as verb phrases and prepositional phrases. This is the reason that spatial language terms such as on, middle, at, outside, left, right, through, and ended also appear prominent in the tag cloud. Our analytical question is: given scenarios where only domain semantic is different, how different will the descriptions be?
In the next step, we used the Stanford POS tagger [
Across frequently appears in translation scenarios but not in scaling ones.
In sum, POS-tagging offers possibilities to examine linguistic usages by lexical categories. Examining the nine corpora, frequently appearing nouns are highly domain specific; a few verbs and adjectives are domain specific and a general difference in translation vs. scaling can be found; prepositions are least domain specific, only the word “across” is found to be differentiable between translation scenarios vs. scaling scenarios.
The last analysis step here involves topic modeling [
Table 1 shows the keywords that identify each topic model. Unsurprisingly, domain specific nouns are distributed across topic models. These topic models can be used to evaluate the probability of one document (descriptions created by one participant) being associated with a specific topic model (ideally catching the domain). Assigning the most probable topic ID to a document allows for using topic models for document classification. To evaluate the correlation between topics and domain semantics further, we use the already built topic models to classify each document. The results and evaluations are shown in Table 2.
Topic ID 0 1 2 3 4 5 6 7 8 Out of 20 documents from each Domain domain, we evaluated the proportion DGeesoemrtetry_translation 116 104 00 10 11 00 21 01 20 of documents being classified into Lake 0 0 19 0 0 0 1 0 0 the same topic model, which ideally CHGaeunorrnmioceantnrye_scaling 030 004 000 131 3016 1014 011 0110 100 (sthhoisulcdorcreosrpreosnpdoenndcetowotrhkeeddeoxmcaeipnt Oil 5 0 1 3 0 0 11 0 0 for tornado and cannon, where most SThoirpnado 00 00 00 00 03 00 12 105 019 of both are assigned to Topic ID 7). The bold numbers in Table 2 shows the topic model (see also Table 1) that most documents from a domain are assigned to. As shown in Table 2, it is reasonable to match each topic ID to one domain semantic and the matching proportion (sum of diagonal cells divided by total) is 70.56%. Cross-examining the domain semantic with keywords from corresponding topic models (see Table 1) we find that a large proportion of documents are classified correctly.
However, the above matching of topic models and domain semantics may be skewed by the high volume of domain specific nouns. Hence, as a comparison, we removed all the domain specific nouns from all corora and rebuilt the topic models.
Two observations are important: In the first part of this paper we showed an analysis by hand that allows for relating a qualitative formal description of a movement pattern to a linguistic description. The linguistic descriptions are varied and participants used manifold strategies to characterize formally identical movement patterns. However, we seem to be able to clearly reveal domain specific differences, especially if we look into whether or not domain semantics is present. In the second part of this paper we tried to use this insight and compared the documents from each domain (one document with all linguistic descriptions per participant, 20 documents in each domain). We found that figure and ground (moving entity and reference entity) are the dominating linguistic features used and that these nouns allow for classifying documents largely correctly. However, once we remove these obvious, domain specific features, classification and identification of documents becomes very inaccurate despite the differences we found in the first part.
There could be a number of reasons for this. Instead of comparing all documents of a particular domain, which contains linguistic descriptions of several, topologically distinguishable paths, we may need a finer granularity for the analysis. For example, we could extract all DC descriptions from all domains and focus only on these. Likewise, we could extract all descriptions for movement patterns that could be labeled across in the translation scenarios and expand-and-retreat in the scaling scenarios. We could perform this analysis for all topologically equivalent movement patterns that we used to design our experiments.
It also could be that the topic modeling approach we used needs refinement. Topic models make use of terms and co-occurrences with documents to discover topics. It is an effective method for knowledge discovery from large corpora without predefined knowledge. However, we are specifically looking for spatial language usage in this study. In order to reduce the influence of domain specific nouns, we use a crude method which is removing the domain specific nouns. Integrating predefined knowledge (in our case, specific target language and contexts) into topic models would allow an analysis to focus on certain term usages, which would enhance the capability of topic modeling.
To sum up, we presented a first exploratory analysis of a corpus that is the result of the conceptualization of movement patterns in different semantic domains. The unique aspect of our experiments is that grounding the design in qualitative spatial representation and reasoning frameworks allows for keeping the spatial information identical across domains only changing the semantic (domain specific) context. We are hopeful that this corpus can contribute to a better understanding of the relation between formal/computation models and spatial language across different domains.