Imitation of Life: A Search Engine for Biologically Inspired Design

Hen Emuna, Nadav Borenstein, Xin Qian, Hyeonsu B. Kang, Joel Chan, Aniket Kittur, Dafna ShahafAAAI 2024 · DOI

TL;DR

Biologically Inspired Design (BID) offers innovative solutions by mimicking nature, but finding relevant biological inspirations is challenging due to limited engineer knowledge and scarce, manually curated resources. This paper introduces BARCODE, a scalable, automated search engine that mines the web for bio-inspirations, significantly improving the accessibility and practicality of BID.

What Problem Does This Paper Address?

Nature is a treasure trove of ingenious solutions to complex engineering problems. Biologically Inspired Design (BID), or biomimicry, leverages these natural strategies to drive innovation. However, the practical application of BID faces significant hurdles. Firstly, engineers and designers often lack the extensive biological knowledge needed to identify relevant natural analogues. Secondly, existing resources like AskNature or Design by Analogy to Nature Engine (DANE) are hand-curated, making them small and costly to scale. This scarcity of readily accessible, comprehensive bio-inspiration data hinders widespread adoption. The paper aims to bridge this gap by creating an automated system that can efficiently search vast amounts of biological information for potential design inspirations.

Methods

BARCODE employs a two-phase approach. The first phase, the "bio-inspiration module," uses a data-programming framework (Snorkel) to process a large corpus of biological documents (Wikipedia in this case). It leverages linguistic labeling functions and external data sources to identify sentences likely to describe a biological solution to a problem. This module generates a "bio-inspiration score" for each sentence.

The second phase, the "query-relevance module," filters out low-scoring sentences and then, given a user's query (representing an engineering challenge), retrieves and ranks the most relevant candidate sentences. To achieve this, BARCODE first extracts candidate phrases from sentences in the form of verb-object pairs, using techniques like Semantic Role Labeling (SRL) and dependency parsing. It then employs a sophisticated matching and ranking system that combines sentence embedding models (SBERT) with Natural Language Inference (NLI) models (DeBERTa). The NLI component is crucial for capturing nuances like causality and common sense, which are often missed by standard embedding models. This combined approach allows BARCODE to identify sentences that not only match keywords but also capture the underlying functional aspects of challenges and strategies.

To address the computational cost of processing large corpora, BARCODE includes an optional "speedup" phase. This phase uses heuristics and labeling functions to pre-filter sentences, identifying those likely to contain a challenge or strategy. This significantly reduces the dataset size without substantially impacting performance.

Datasets

The primary corpus used in this paper is Wikipedia, specifically articles categorized under "Articles with 'species' microformats." This corpus contains 780,949 sentences from 27,640 articles.

For evaluation, the authors also utilized:

  • AskNature's "Biological Strategies" section: Titles from this section were used to formulate queries, representing innovation challenges.
  • Patent data: A dataset of 20 million sentences from the "Claims" section of patents was used to construct a list of known problems for the filtering heuristic.

No new benchmark datasets were introduced, but the paper's methodology and released code aim to facilitate the creation and expansion of such datasets in the future.

Key Insights and Ideas

  • The power of NLI for bio-inspiration: Standard sentence embedding models struggle with semantic nuances like antonyms and causality. The integration of Natural Language Inference (NLI) models, specifically DeBERTa, proved crucial for BARCODE to correctly identify subtle connections, such as how "staying moist" relates to "reducing water loss," a capability that significantly enhances the retrieval of relevant bio-inspirations.
  • Leveraging data programming for scalability: The use of Snorkel and labeling functions allows BARCODE to process vast amounts of unstructured biological text, overcoming the limitations of manual curation and enabling a scalable approach to bio-inspiration discovery.
  • The trade-off between filtering and performance: The optional filtering heuristic drastically reduces the search space (to about 3% of the original data) with minimal impact on performance, especially for top-k results, making the system more efficient for practical use.
  • Bridging the gap between technical queries and biological solutions: The system's ability to handle paraphrased and more technical queries, particularly with the help of the NLI component, suggests it can effectively translate human design needs into biological contexts.

Key Results

  • BARCODE significantly outperforms a baseline search engine (Elasticsearch) in terms of Precision@k and NDCG@k for both challenge and strategy retrieval.
  • On the entire dataset, BARCODE retrieves a strategy in 72% of the top 7 sentences and 69% of the top 15, compared to the baseline's 36% and 29%, respectively.
  • The filtering heuristic reduces the dataset size to 3% of the original, with only a minor impact on performance, especially for the top 7 results.
  • In an ideation experiment with three bio-engineering experts, BARCODE successfully provided relevant inspirations, with 33% of results deemed "highly interesting" and 21% already known to the experts but still valuable.
  • BARCODE also demonstrated robustness to query variations, showing a higher Rank-Biased Overlap (RBO) on filtered data compared to the unfiltered dataset.

Limitations and Future Work

  • Corpus idiosyncrasies: The authors acknowledge that substantial variance in performance across queries might be due to the specific composition of the Wikipedia corpus.
  • Interactive query refinement: Exploring interactive mechanisms for query reformulation and expansion, potentially with the aid of large language models, could further improve user experience.
  • Broader data sources: The future work includes applying BARCODE to more detailed biological data sources, such as scientific papers, to expand its coverage.
  • Generative AI integration: An exciting future direction is integrating BARCODE with generative AI technologies to not only retrieve inspirations but also to generate sketches of solutions by adapting retrieved biological strategies to user problems.

Summary updated

BibTeX
@inproceedings{emuna2024imitation,
  author = {Hen Emuna and Nadav Borenstein and Xin Qian and Hyeonsu B. Kang and Joel Chan and Aniket Kittur and Dafna Shahaf},
  title = {{Imitation of Life: A Search Engine for Biologically Inspired Design}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i1.27805},
}

Loading related content...

Figure