Analysing Siamese Neural Network Architectures for Computing Name Similarity.
Bibliographic record
Abstract
IntroductionA wide assortment of string similarity measures can be used to determine how similar two names are. A diverse set of discriminating and independent features for name similarity are important for classification during record linkage. A Siamese neural network could surpass traditional string similarity measures for the name similarity problem.
 Objectives and ApproachThis research aims to compare a classifier based on the Siamese network architecture with a Random Forest classifier. In addition to comparing overall performance, we seek to answer whether there are any special properties of certain matching name pairs where the complexity of the Siamese network offers particular benefit.
 Our data consists of 25,000 last name pairings, with each pair being two variants of a family name. Name similarity predictions from the Siamese network are compared to a Random Forest model that serves as an ensemble of existing string similarity measures.
 ResultsWe compare the similarity scores yielded by the two methods and discuss the results. We describe the representation of names to each method; name representation is computed formulaically for the traditional measures but is learned by the Siamese network during training. The comparison of different methods is made both in terms of their similarity prediction quality, and the computational cost to generate the predictions.
 As expected, the Siamese network necessitates a significant computational cost to train. Unexpectedly, the ensemble of traditional measures yields almost identical overall classification performance. However, we expect that further analysis of false positives and false negatives will yield some insight into when practitioners should consider one method over the other.
 Conclusions/ImplicationsResults suggest that there may be instances where a Siamese network outperforms other similarity measures, although training a Siamese network comes at a considerable computational cost. It is worth considering this approach to name similarity as an additional similarity feature when performing record linkage tasks.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.017 | 0.006 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.002 | 0.000 |
| Scholarly communication | 0.002 | 0.001 |
| Open science | 0.007 | 0.003 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".