Food Knowledge Representation Learning with Adversarial Substitution

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 653–664 November 20–23, 2022. ©2022 Association for Computational Linguistics 653 Food Knowledge Representation Learning with Adversarial Substitution Diya Li∗, Mohammed J. Zaki Computer Science Department, Rensselaer Polytechnic Institute 916lidiya@gmail.com, zaki@cs.rpi.edu Abstract Knowledge graph embedding (KGE) has been well-studied in general domains, but has not been examined for food computing. To fill this gap, we perform knowledge representation learning over a food knowledge graph (KG). We employ a pre-trained language model to encode entities and relations, thus emphasiz- ing contextual information in food KGs. The model is trained on two tasks – predicting a masked entity from a given triple from the KG and predicting the plausibility of a triple. Analysis of food substitutions helps in dietary choices for enabling healthier eating behaviors. Previous work in food substitutions mainly fo- cuses on semantic similarity while ignoring the context. It is also hard to evaluate the sub- stitutions due to the lack of an adequate vali- dation set, and further, the evaluation is sub- jective based on perceived purpose. To tackle this problem, we propose a collection of adver- sarial sample generation strategies for different food substitutions over our learnt KGE. We pro- pose multiple strategies to generate high quality context-aware recipe and ingredient substitu- tions and also provide generalized ingredient substitutions to meet different user needs. The effectiveness and efficiency of the proposed knowledge graph learning method and the fol- lowing attack strategies are verified by exten- sive evaluations on a large-scale food KG. 1 Introduction Structured knowledge furnishes an in-depth under- standing of the world. Knowledge graph embed- ding (KGE) maps entities and relations into vectors while retaining their semantics (Wang et al., 2017; Lin et al., 2018). KGE has been well-studied and applied in general KGs with common ontological knowledge (i.e., WordNet (Miller, 1995), DBpe- dia (Auer et al., 2007), and Freebase (Bollacker et al., 2008)). Only a few works have targeted * Diya Li is the corresponding author. domain-specific KGs (Mohamed et al., 2021; Bon- ner et al., 2021) and to the best of our knowledge, there is no work for KGE in the food domain. Even though previous work (Li and Zaki, 2020) trains recipe embeddings on a large-scale dataset, KG information is utilized only as side information to assist embedding learning and only recipes get represented, and other node types in the food KG, such as ingredients are ignored. To fill this gap, we aim to conduct knowledge representation learning over the entire food KG to get high-dimensional vectors of nodes and relations while capturing their semantic meanings. As for encoding models in KGE, most deep learning-based methods like convolutional neural networks (CNN) (Dettmers et al., 2018), recur- rent neural networks (RNN) (Guo et al., 2019) and graph neural networks (GNN) (Schlichtkrull et al., 2018; Shang et al., 2019) allow a single static em- bedding for …

Content truncated to 500 words