Unsupervised Matching of Data and Text

PromptEM: Prompt-tuning for Low-resource Generalized Entity Matching

“…We use all the seven real-world benchmark datasets with different structures from Machamp [50] and one geospatial dataset (GEO-HETER) [2]. The detailed GEO-HETER construction can be found in our online version 1 . The statistics of datasets are summarized in Table 1.…”

“…• Rotom [30] proposes a meta-learning framework that selects and weights the augmented data to better fine-tune the LMs. • TDmatch [1] is an unsupervised approach to match textual and structured data using graph creation and random walk. Furthermore, we build an MLP classifier on top of its embeddings to perform in the supervised setting, called TDmatch*.…”

“…Here, h [CLS] is the embedding of special classification token [CLS], 𝑏 is the bias for the layer, and W is a learnable matrix randomly initialized before fine-tuning. The parameters of M, 𝑏, and W are tuned to maximize 1 |X | 𝑥 ∈X log 𝑝 (𝑦 | 𝑥).…”

“…Nonetheless, these methods only focus on EM tasks in low-resource scenarios but perform poorly on GEM tasks. TDmatch [1] first attempts to match textual content and structured data under an unsupervised setting. However, it has one serious shortcoming: it is not scalable on large-scale datasets, which makes it hard to be used in practical scenarios.…”

“…Nonetheless, the fine-tuning paradigm still requires a non-trivial amount of high-quality labeled examples (e.g., thousands of labels for a typical GEM application [50]). TDmatch [1] performs in an unsupervised learning way via graph creation and random walk. However, two drawbacks restrict its real-world applications as confirmed via experiments in Section 5: (i) it has unstable performance due to the absence of labeled samples; and (ii) random walk is not scalable on large datasets [44], incurring enormous execution time and memory usage (e.g., more than 120 hours and 130 gigabytes for once training of the SEMI-REL).…”