2. NOMENCLATURE

• Biomimicry Taxonomy: A three-level, hierarchical classification mechanism developed to organize AskNature's knowledge base

• Class: a function-level classification category of the Biomimicry Taxonomy

• Class support: the number of reference documents in AskNature's database for a specific Biomimicry Taxonomy class

• Class weight: calculated as the maximum support divided by the specific class's support

• Corpus: a collection of biological strategies in natural-language format

• Part of speech: a linguistic category of words, for example, verbs, adjectives, nouns

• Reference document: a biological strategy from AskNature's database used to train the proposed algorithm

• Sample document: any biological strategy described in natural-language format

• Sample score for a class: the number of k-nearest neighbor (k-NN) reference documents with reference classification to a specific class

• Weighted score: the product of the sample score and the class weight

3. RELATED RESEARCH

This section provides an overview of related research with the emphasis on the integration of large numbers of biological strategies in the systematic BID process. The authors provide more detailed descriptions in Vandevenne et al. (Reference Vandevenne, Verhaegen, Dewulf and Duflou2011), where each contribution is positioned in the following four BID process steps: problem formulation, solution search, filter and analysis of alternatives, and knowledge transfer.

A manual, iterative bioinspiration search (Lenau et al., Reference Lenau, Dentel, Ingvarsdóttir and Guðlaugsson2010), starts from a functional keyword search and, from the obtained results, extracts new biological keywords for future searches. In this way, biological search words, initially not known to be relevant to the problem, are identified. A contribution aimed at automating the identification of biologically relevant search words is proposed by Chiu and Shu (Reference Chiu and Shu2007) and Shu (Reference Shu2010). The method aims at bridging the terminology gap between the engineering and biological domain by means of a systematic, semiautomatic search method that requires the design problem to be expressed in functional keywords and then generates biological meaningful bridge verbs and text passages containing them.

There are three model-based approaches that require the manual instantiation of detailed models for each biological phenomenon to be integrated in a structured knowledge base. Such a methodology has currently been reported for structure–behavior–function (SBF) models (Vattam et al., Reference Vattam, Wiltgen, Helms, Goel and Yen2010; Goel et al., Reference Goel, Vattam, Wiltgen and Helms2012), for functional basis models (Nagel et al., Reference Nagel, Nagel, Stone and McAdams2010; Nagel & Stone, Reference Nagel and Stone2012), and for state change, action, part, phenomenon, input, organ, and effect (SAPPhIRE) models of causality (Chakrabarti et al., Reference Chakrabarti, Sarkar, Leelavathamma and Nataraju2005; Sartori et al., Reference Sartori, Pal and Chakrabarti2010). These three model-based approaches have been recently extended in the following ways. In order to scale the SBF approach, Biologue, a social citation cataloguing system is developed (Vattam & Goel, Reference Vattam and Goel2011) to involve more people in the process of manual creation of SBF models. The functional basis approach is extended with an Engineering to Biology Thesaurus (Cheong et al., Reference Cheong, Chiu, Shu, Stone and McAdams2011; Nagel & Stone, Reference Nagel and Stone2012), a lookup table that translates the functional basis terms into biological corresponding terms. Finally, the SAPPhIRE approach is extended by an ontology aimed at providing extra stimuli during bioinspired ideation. The ontology consists of manually derived, biological and engineering term clusters for each of the SAPPhIRE model constructs (Srinivasan et al., Reference Srinivasan, Chakrabarti and Lindemann2012).

The following two contributions require positioning each biological strategy into a classification scheme. First, AskNature places biological strategies in a functional, hierarchical taxonomy called the Biomimicry Taxonomy. The designer looking for bioinspiration needs to formulate his or her design problem in this taxonomy. Second, BioTRIZ (Vincent et al., Reference Vincent, Bogatyreva, Bogatyrev, Bowyer and Pahl2006) aims at integrating biological knowledge in the TRIZ methodology (Altshuller, Reference Altshuller1984) by positioning biological strategies in the BioTRIZ contradiction matrix. To identify bioinspiration, the problem needs to be formulated into a classical TRIZ contradiction, which is then reformulated into a BioTRIZ contradiction. This BioTRIZ contradiction then leads the designer to inventive principles learned from the manual analysis of 2500 contradictions in 500 biological phenomena. In order to illustrate the resources integrated into the above systems, for each contribution, the number of reported biological sources is given in Table 1.

Table 1. Overview of existing database sizes and content

Note: SBF, Structure–behavior–function; SAPPhIRE, state change, action, part, phenomenon, input, organ, and effect.

All of the above methodologies struggle in one way or another with scalably leveraging large numbers of biological resources in natural-language format. Both the iterative bioinspiration search and the bridge verbs-based search entail extensive manual result filtering. All model-based approaches are difficult to scale because they require a detailed analysis of both the engineering and the biological systems to express them on a common abstraction level. Their thesaurus or ontology extensions raise questions about the completeness of the relatively short biological word lists, which, in turn, makes it difficult to estimate how much of the biological inspiration in natural-language texts can be retrieved when scaling these methodologies to large biological corpora; and validation is typically reported on a small number of handpicked cases. Crowd sourcing, recently reported for the SBF model-based approach (Vattam & Goel, Reference Vattam and Goel2011), in theory, could tackle the scalability of any BID ideation system. However, the successful creation of a large number of SBF models by an online community has not yet been reported. Third, the two classification-based approaches require the classification of each new biological strategy. This manual task translates to the positioning of the strategies into the Biomimicry Taxonomy, for the AskNature approach, or to the manual identification of the relevant contradiction, for the BioTRIZ approach. These are again tasks that demand interactive work proportional to the size of the biological databases. Like the SBF model-based approach, AskNature relies on an online community to aid in database expansion. The authors have observed the AskNature database to grow with about 100 strategies a year over the last 3 years. Although this number seems small compared to the total number of biological strategies that are being documented by humans, AskNature's database is the only initiative known to the authors that shows any significant and consistent database growth. The research presented in this paper details an approach to automatically populate the Biomimicry Taxonomy and hence scale AskNature's bioinspiration approach.

4. AskNature AND THE BIOMIMICRY TAXONOMY

AskNature, developed by the Biomimicry 3.8 Institute (http://www.biomimicryinstitute.org), is an online tool that provides support for designers during concept generation in the early stages of the BID process. To use AskNature, a designer is required to formulate his problem into AskNature's Biomimicry Taxonomy, a functional, three-level hierarchical classification mechanism developed to organize his knowledge base. A small excerpt from the Biomimicry Taxonomy is presented in Table 2. The first level is the group level, which consists of eight categories, for example, “move or stay put.” Each group-level category is divided into subcategories named subgroups. The two subgroups for the “move or stay put” group are, for example, “attach” and “move.” Subgroups are further divided into functions. For example, “attach temporarily” and “attach permanently” are two function categories for the subgroup “attach.” In this way, the biological strategy of the octopus using suckers to attach itself is, for example, classified across the three levels as: “move or stay put” / “attach” / “attach temporarily.” Once the person looking for inspiration from nature has formulated his or her problem into the Biomimicry Taxonomy, AskNature returns a list of biological strategies that are previously manually classified into the chosen Biomimicry Taxonomy class.

Table 2. Excerpt from the Biomimicry Taxonomy

A taxonomy is generally an information structure used to classify instances. In a typical taxonomy, instances can be positioned into mutually exclusive, unambiguous categories, for example, the taxonomy of Linnaeus (Reference Linnaeus1767). However, the Biomimicry Taxonomy allows the classification of a single strategy into multiple categories. Take, for example, the harlequin beetle: an organism that uses its strong, large mandibles to escape from the trees in which it is born by chewing through wood. This strategy is currently classified in the following two categories: “break down” / “physically break down” / “biotic materials” and “move or stay put” / “move” / “in solids.” Although one can argue about the word choice of taxonomy in the Biomimicry Taxonomy, the fact that biological strategies can be positioned into more than one category does not impede the functioning of the bioinspiration tool. In contradiction, this property allows the retrieval of a single biological strategy document for different relevant problem formulations or desired functions.

5. AUTOMATED CLASSIFICATION INTO THE BIOMIMICRY TAXONOMY

On September 18, 2012, the AskNature database contained 1531 unique strategy descriptions, with a total of 2826 classifications into the Biomimicry Taxonomy. Forty-seven percent of all strategies are classified into only one third-level Biomimicry Taxonomy class, 32% are classified twice, 13% three times, and 8% more than three times; the average number of classifications per unique biological strategy is 1.82. For the proposed automated classification approach outlined in this section, the 1531 reference documents, with their 2826 classifications, form the reference corpus. Before detailing the processing steps of the automated classification algorithm, in order to understand the choices that were made, a number of challenges of the classification task are the following:

• • Challenge 1: The total number of reference documents (1531) and the total number of their classifications (2826) are low compared to the total number of Biomimicry Taxonomy classes (159). This makes training examples a scarce resource.

• • Challenge 2: The number of reference documents per class, further referred to as class support, is not evenly distributed over these classes. Figure 1 shows the reference document support per Biomimicry Taxonomy class. Such an uneven distribution of reference documents encumbers training an automated classifier because some classes will not have enough reference documents for trustworthy decision making. There are 18 classes that, for the moment, have no reference strategies assigned to them.

• • Challenge 3: The reference documents are short, with on average 435 words per document. Fewer terms in the reference documents entail fewer links to the sample documents. In contrast, a well-written short description of a biological strategy should contain a relatively large number of relevant document features for automated classification into the Biomimicry Taxonomy.

• • Challenge 4: The Biomimicry Taxonomy classes are not mutually exclusive, as explained in Section 4. This introduces ambiguity in the classification task, which has on average 1.82 correct classifications per biological strategy.

• • Challenge 5: Not all possible classifications for each unique strategy are exhaustively identified in the reference corpus. Although this has no consequences for the correct functioning of the bioinspiration tool, when reference documents are used as sample documents during the testing of the classification algorithm, this is likely to lead to performance underestimation.

Fig. 1. Distribution of the number of available reference documents (or support) per class; the class names of the 10 best-supported classes are provided in Table 4.

The answers to these challenges are detailed throughout the description of the proposed system, its testing, and discussion. Figure 2 provides an overview of the proposed approach. A number of preprocessing steps transform the reference and sample corpus into document-term matrices. The reference corpus contains the strategies of the AskNature database, and the sample corpus can contain any number of biological strategies in natural-language format. The document-term matrices of these corpora are their representations in the vector space model (Salton et al., Reference Salton, Wong and Yang1975). This algebraic model represents documents as vectors, where each dimension corresponds to a unique corpus word or feature and each feature value corresponds to the importance of the word in the document. In this vector space representation, mathematical operations between reference and sample document vectors enable the calculation of interdocument distances. For the illustrated algorithm, k-NN classification (Cover & Hart, Reference Cover and Hart1967) identifies a target classification for each sample document. The sections below detail the different steps of the proposed Biomimicry Taxonomy classification system.

Fig. 2. The algorithm for automated classification into the Biomimicry Taxonomy.

5.3. k-NN classification into the Biomimicry Taxonomy

The k-NN is a machine learning approach that takes the k-nearest neighboring reference documents to a sample document, and, to compose a hypothesis, performs a majority vote on those reference document classifications (Cover & Hart, Reference Cover and Hart1967). The k-NN is a type of instance-based learning, meaning that instead of making a generalized classification model, a classification hypothesis is directly constructed from the reference corpus documents. For the specific classification task at hand, the lack of model building is an especially interesting property. As the reference database slowly grows, that is, strategies are currently manually added by a rate of one every couple of days, new documents can be added to the reference set without the need to recalculate the model. Training should be seen as merely storing the new feature vectors with their reference classifications. For the proposed system, the decision of which k reference documents are the nearest neighbors of a sample document is based on the well-known cosine similarity measure between their document vector representations (Baeza-Yates & Ribeiro-Neto, Reference Baeza-Yates and Ribeiro-Neto2011). The more similar reference and sample documents are, the smaller the angle is between their feature vectors, and the more their cosine similarity approaches 1. Completely dissimilar document vectors are orthogonal; hence their cosine similarity is 0. The parameter k, which represents the number of reference documents that take part in the majority vote, currently is set to 50. However, the value of this parameter should be reevaluated when the reference corpus support increases significantly. The process of k-NN classification is illustrated in Appendix D with pseudocode.

Performing a majority vote on the classifications of the k-nearest neighboring reference documents results in an ordered list of candidate Biomimicry Taxonomy classes, as illustrated for an example biological strategy document with the title “Fibers reinforce nests: wasps” in Table 3. The first two columns represent, respectively, the target classes and the number of reference documents in the k-nearest neighbors associated to these classes (depicted as “sample score”). In order to prevent the bias in corpus support, illustrated by Figure 1 and column 3 in Table 3, from weighing on the classification results, the majority vote is counterweighted accordingly with the weights shown in the fourth column. These are calculated for each class as the maximum support divided by the class's support. Given the class support distribution illustrated by Figure 1, it can be easily seen that such weighting currently prevents the classes with very low support from taking part in the majority vote. For example, in the most extreme case, a class with only one reference document would receive a weight equal to the number of reference documents in the best-supported class and, hence, would dominate voting results. In the ideal case, when all target classes contain a sufficiently high and equal number of reference documents, weighting will be omitted and all classes will take part in the majority vote. However, considering the current status of the reference corpus (Challenges 1 and 2), for the time being, classification is limited to the top 10 Biomimicry Taxonomy classes. Applying this class filter to the full list of candidate classifications linked to the example document's k nearest neighbour reference documents results in the short list of candidate classes shown in Table 3. The highest weighted score is taken as classification for the sample document. In this example, 15 out of the 50 nearest neighbouring reference documents have the classification make / physically assemble / structure, which agrees with the reference classification of the example biological strategy.

Table 3. Example of an ordered list of candidate Biomimicry Taxonomy classes

7. DISCUSSION AND FUTURE WORK

There are three arguments that motivate the choice of scaling AskNature. First, the approach of AskNature is the only publicly available bioinspiration system with a continuously growing database that, at current, qualifies as an initial training set for scaling an existing BID ideation approach. Second, using the tool requires very little training, which adds to its appeal. By scaling AskNature's approach, the authors make no argument about the optimality of the specific adopted categorization. AskNature has updated the Biomimicry Taxonomy iteratively, and there is no indication that this process is final. Third, no research exists that identifies the most optimal tool for systematic BID. Therefore, the authors do not claim to have scaled the most optimal systematic BID approach, but to be the first to demonstrate the feasibility of scaling any BID approach, that is, AskNature, the most popular one. The tests in this paper are performed on the Biomimicry Taxonomy as observed on September 18, 2012, and the applied classification algorithm is independent of possible future changes to this taxonomy.

For testing the proposed approach, the classifications provided by AskNature are used as a golden key for the validation tests (see motivation in the previous section). As this database is collaboratively expanded, which is a human operation, it is possible that some error is introduced in the training and test sets. During the interactions with this reference database while testing, the authors have not encountered misclassifications of AskNature strategy documents. Therefore, in the context of illustrating the feasibility of the proposed approach, the reference classifications are taken as a golden key. However, as part of future work, checking the reference database against human raters would be valuable to verify the high level of trust AskNature's reference classifications currently enjoy. Besides quantifying the fraction of potential misclassifications, such validation could also quantify the observation that for some reference documents not all relevant reference classes are exhaustively identified in the reference corpus (see Section 5, Challenge 5).

Four typical steps of the systematic BID process are problem formulation, solution search, filter and analysis of alternatives, and knowledge transfer (Vandevenne et al., Reference Vandevenne, Verhaegen, Dewulf and Duflou2011). By scaling AskNature, the authors have addressed an important bottleneck in the second step: solution search. However, choosing AskNature in the search phase does not exclude other approaches from being integrated in the scaled systematic BID process to come to a hybrid approach. Because all mentioned related research approaches described in Section 3 have a functional component in their problem formulation, in the first step, restating the problem with other guidelines than those of AskNature can inspire us to look at the problem differently and result in new entry points into the Biomimicry Taxonomy. Furthermore, the three model-based systematic BID approaches can support both the analysis of a small number of alternative strategies (Step 3) and the formulation of knowledge transfer (Step 4). When composing any of these models for the low number of alternative biological strategies, the designer is forced to look at the system from different perspectives (e.g., function, behavior, and structure). This helps to form a more complete understanding of biological systems under focus. If there are problems understanding the retrieved biological solutions, the authors envisage a need to contact domain-specific biologists in these two final steps. Instead of asking biologists to identify what is interesting in nature to the design problem at hand (search phase), their role is shifted to the analysis and knowledge transfer phases. Besides a wider search, not limited to the personal knowledge of one or a couple of biologists, this allows to us contact the most relevant biologist(s) for the specific design problem.

The more reference documents will support the classes of the Biomimicry Taxonomy, the more relevant biological strategies will be able to take part in the majority vote. Therefore, taking into account the current scarcity of training examples (Challenge 1), it is reasonably expected that the performance of the system will increase further when the reference corpus grows. In addition, the randomly drawn sample documents from the reference corpus are much shorter documents (on average 435 words per document) than typical biological papers or online articles (thousands of words) that require automated classification. Because shorter documents are likely to have fewer interesting features to form relevant interdocument links, this makes the sample corpus extra challenging. Hence, the reported performance of the outlined approach should be interpreted as a proof of concept that illustrates the potential of the approach. Because any classifier is bound to make at least some misclassifications, the authors envisage a user interface element allowing users to signal the system that a certain strategy is positioned incorrectly. When sufficient human raters agree about such a classification adjustment, the misclassification of the sample document can be corrected and the document can be even moved to the reference corpus.

Every step in the classification procedure runs completely autonomously (hence, the proposed classification approach scales for leveraging large biological databases). Nevertheless, there are two notes to be made in the context of scalability. First, the reference corpus is composed manually by AskNature's online community, with steady but relatively slow growth. In order to realize sufficient support for all classes of the Biomimicry Taxonomy, the community should be steered toward adding strategies for those categories where necessary, instead of directing attention to categories that, from the point of building a reference corpus, are not a priority. Second, by anticipating a small user interface element allowing people looking for inspiration to signal a strategy's potential misclassification, again a light form of crowd sourcing is proposed to further boost the system's precision. However, this adds to the current state of the art because the required task is in the order of a single mouse click for a relatively small number of biological strategies, and once the reference corpus is adequately built, this will be the only manual interaction in support of the scaled ideation system.

One way of obtaining a sample corpus is training a webcrawler to collect biological strategies from the Internet (Vandevenne, Caicedo, et al., Reference Vandevenne, Caicedo, Verhaegen, Dewulf and Duflou2012). Such strategy documents can be fed to the presented classifier as sample corpus, which was previously done with an early version of the classification algorithm (Vandevenne, Verhaegen, et al., Reference Vandevenne, Verhaegen, Dewulf and Duflou2012). In these tests, precision was boosted to 90% by introducing a cutoff score on the sample document classification results. This, however, entailed that most sample documents did not have a target classification that met the cutoff score and were discarded. To avoid such loss of sample documents, the presented algorithm in Section 5 does not integrate a cutoff score. However, in the context of a webcrawling corpus, that in theory could become very large, adopting a cutoff score to boost precision can be argued for.

Although Section 6 presented tests that indicate it is feasible to scale AskNature's database expansion with automated classification, doing so will introduce a new challenge: avoiding information overload. When a designer selects a Biomimicry Taxonomy functional class to retrieve biological strategies relevant to his or her problem, presenting very large lists of strategies should be avoided. Therefore, the authors anticipate the need for a hierarchical representation of the results based on similarities between the strategy descriptions. To implement this user interface functionality, the suitability of standard document clustering techniques will be explored, to compose an extra information level for assisting the designer with the efficient identification of relevant strategies. In addition, when biological strategies in the form of academic papers are returned to the person looking for bioinspiration, the size of these documents (thousands of words per strategy) necessitates highlighting specific paragraphs, sentences, or words. To facilitate analogical transfer, efforts should be made to assist in the identification of terms relevant for structural mapping (Gentner & Markman, Reference Gentner and Markman1997) between the source and target domains. Highlighting the most important words that caused the classification of the retrieved biological strategies into the chosen functional Biomimicry Taxonomy class could be a good start. Furthermore, to accentuate words representing causally related functions, a recently developed extraction algorithm (Cheong & Shu, Reference Cheong and Shu2009, Reference Cheong and Shu2012) could be integrated to further support the abstraction of biological strategies to facilitate cross-domain knowledge transfer.