skogler commented on a change in pull request #993: URL: https://github.com/apache/systemds/pull/993#discussion_r457312560
########## File path: scripts/staging/entity-resolution/README.md ########## @@ -0,0 +1,99 @@ +# Entity Resolution + +## Pipeline design and primitives + +We provide two example scripts, `entity-clustering.dml` and `binary-entity-resolution.dml`. These handle reading input +files and writing output files and call functions provided in `primitives/pipeline.dml`. + +### Input files + +The provided scripts can read two types of input files. The token file is mandatory since it contains the row identifiers, +but the embedding file is optional. The actual use of tokens and/or embeddings can be configured via command line parameters +to the scripts. + +##### Token files + +This file type is a CSV file with 3 columns. The first column is the string or integer row identifier, the second is the +string token, and the third is the number of occurences. This simple format is used as a bag-of-words representation. + +##### Embedding files + +This file type is a CSV matrix file with each row containing arbitrary-dimensional embeddings. The order of row identifiers +is assumed to be the same as in the token file. This saves some computation and storage time, but could be changed with +some modifications to the example scripts. + +### Primitives + +While the example scripts may be sufficient for many simple use cases, we aim to provide a toolkit of composable functions +to facilitate more complex tasks. The top-level pipelines are defined as a couple of functions in `primitives/pipeline.dml`. +The goal is that it should be relatively easy to copy one of these pipelines and swap out the primitive functions used +to create a custom pipeline. + +To convert the input token file into a bag-of-words contingency table representation, we provide the functions +`convert_frame_tokens_to_matrix_bow` and `convert_frame_tokens_to_matrix_bow_2` in `primitives/preprocessing.dml`. +The latter is used to compute a compatible contigency table with matching vocabulary for binary entity resolution. + +We provide naive, constant-size blocking and locality-sensitive hashing (LSH) as functions in `primitives/blocking.dml`. + +For entity clustering, we only provide a simple clustering approach which makes all connected components in an adjacency +matrix fully connected. This function is located in `primitives/clustering.dml`. + +To restore an adjacency matrix to a list of pairs, we provide the functions `untable` and `untable_offset` in +`primitives/postprocessing.dml`. + +Finally, `primitives/evaluation.dml` defines some metrics that can be used to evaluate the performance of the entity +resolution pipelines. They are used in the script `eval-entity-resolution.dml`. + +## Testing and Examples + +There is a test data repository that was used to develop these scripts at +[repo](https://github.com/skogler/systemds-amls-project-data). In the examples below, it is assumed that this repo is +cloned as `data` in the SystemDS root folder. The data in that repository is sourced from the Uni Leipzig entity resolution +[benchmark](https://dbs.uni-leipzig.de/research/projects/object_matching/benchmark_datasets_for_entity_resolution). + +### Preprocessing + +Since there is no tokenization functionality in SystemDS yet, we provide a Python preprocessing script in the data repository +that tokenizes the text columns and performs some simple embedding lookup using Glove embeddings. + +The tokens are written as CSV files to enable Bag-of-Words representations as well as matrices with combined embeddings. D +epending on the type of data, one or the other or a combination of both may be better. The SystemDS DML scripts can be +called with different parameters to experiment with this. + +### Entity clustering + +In this case we detect duplicates within one database. As an example, we use the benchmark dataset Affiliations from Uni Leipzig. +For this dataset, embeddings do not work well since the data is mostly just names. Therefore, we encode it as Bag-of-Words vectors +in the example below. This dataset would benefit from more preprocessing, as simply matching words for all the different kinds of +abbreviations does not work particularly well. + +Example command to run on Affiliations dataset: +``` +./bin/systemds ./scripts/algorithms/entity-resolution/entity-clustering.dml -nvargs FX=data/affiliationstrings/affiliationstrings_tokens.csv OUT=data/affiliationstrings/affiliationstrings_res.csv store_mapping=FALSE MX=data/affiliationstrings/affiliationstrings_MX.csv use_embeddings=FALSE XE=data/affiliationstrings/affiliationstrings_embeddings.csv +``` +Evaluation: +``` +./bin/systemds ./scripts/algorithms/entity-resolution/eval-entity-resolution.dml -nvargs FX=data/affiliationstrings/affiliationstrings_res.csv FY=data/affiliationstrings/affiliationstrings_mapping_fixed.csv +``` + +### Binary entity resolution + +In this case we detect duplicate pairs of rows between two databases. As an example, we use the benchmark dataset DBLP-ACM from Uni Leipzig. +Embeddings work really well for this dataset, so the results are quite good with an F1 score of 0.89. + +Example command to run on DBLP-ACM dataset with embeddings: +``` +./bin/systemds ./scripts/algorithms/entity-resolution/binary-entity-resolution.dml -nvargs FY=data/DBLP-ACM/ACM_tokens.csv FX=data/DBLP-ACM/DBLP2_tokens.csv MX=data/DBLP-ACM_MX.csv OUT=data/DBLP-ACM/DBLP-ACM_res.csv XE=data/DBLP-ACM/DBLP2_embeddings.csv YE=data/DBLP-ACM/ACM_embeddings.csv use_embeddings=TRUE +``` +Evaluation: +``` +./bin/systemds ./scripts/algorithms/entity-resolution/eval-entity-resolution.dml -nvargs FX=data/DBLP-ACM/DBLP-ACM_res.csv FY=data/DBLP-ACM/DBLP-ACM_perfectMapping.csv +``` + +## Further Work + +1. Better clustering algorithms. +2. Multi-Probe LSH. +3. Classifier-based matching. +4. Better/built-in tokenization +5. Better/built-in embeddings. Review comment: Will extend a bit in the evening, some quick explanations: 1. Some approaches that could be implemented are described here: https://dbs.uni-leipzig.de/en/publication/title/comparative_evaluation_of_distributed_clustering_schemes_for_multi_source_entity_resolution 3. In the DeepER paper subsection 2.4 they train a classifier such as an SVM to decide if a pair of tuples is duplicate. We just use a similarity threshold. 4/5. The idea is to have something like transformencode but for text, such that we can have different tokenizers and embedding transformers built into SystemDS. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
