1.
Latest Techniques of Entity
Matching in NLP
Avinash Pathak
Expert Data Scientist
TomTom

2.
Agenda
What is entity matching?
Why entity matching is important?
History of Entity matching
Entity matching models
How to measure success of Entity Matching
Entity Embed – Open-source tool for Entity Matching

3.
What is entity matching?
Entity Matching refers to the problem of determining whether two different data representations
refer to the same real-world entity.
Example
Use Cases
Id Title Album Composer Songwriter
1 Me and Mrs Jones Call me irresponsible Michael Bublé Michael Bublé
2 Me and Mrs Jones[Remix] Call me irresponsible Michael Bublé Michael Bublé
Song Matching •Address Matching
•Social Profile
Matching (Facebook,
Twitter)
•Clothes Matching
(for that matter any
item in retail)
•Clothes Matching
(for that matter any
item in retail)

4.
Things get little complicated
• Use of words in Colloquial fashion
• additional information
• Un-normalization
• Unstructured data
• Missing Data
• Dirty Data
• Un-availability of Supervised data
• Scale
Id Title Album Composer Songwriter
1 Me and Mrs Jones Call me irresponsible Michael Bublé
2 Me and Mrs Jones[Remix] Michael Bublé
3 Blowin’ in the Wind The Freewheelin’ Bob Dylam Bob Dylan
4 Blowing in the Wind Bob Dylan

5.
Problem of Scale, hence blocking
• An exhaustive pairwise comparison grows quadratically with the number of records, which is unaffordable for
datasets of even moderate size. As a result, virtually every entity matching task on large datasets requires
blocking, a step that effectively reduces the number of record pairs to be considered for matching without
potentially ruling out true matches.
• A successful application of blocking to an entity matching task should fulfil the following four desiderata
• First, blocking, ideally, should not leave out any true matches (i.e., high recall), since only the candidate record
pairs generated by blocking will be further examined in the downstream matching step.
• Second, the number of candidate pairs should be small so that the cost of applying a usually computationally-
expensive matching algorithm is controlled.
• Third, human effort should not be overspent during the whole blocking process;
• Finally, the blocking algorithm should be scalable enough to handle millions of records.

6.
Without
Blocking
Blocking (non-expensive
comparisons)
kC2*M Rigorous
computationally expensive
Comparisons
nC2 Rigorous
computationally expensive
Comparisons
With
Blocking
10c2 = 10 * 9 / 2 = 45 expensive comparisons without blocking
For k=2, 5 clusters of size 2
kC2*M = 1 * 5 comparisons. K – cluster size, M – number of clusters
For k=3, 3 clusters of size 3 and 1 with size 1
It becomes 2 * 3 + 1. 7 comparisons

7.
History of EM
• Pattern Matching/Fuzzy Matching
• Deep Learning Based Blocking
• Self-supervised blocking

8.
Pattern Matching/Fuzzy Matching
Example
Very specific solution
Needs incremental updates if new examples come in
Cross attributes mismatch is present and would need enormous efforts to cover
that
Id Title Album Composer Songwriter
1 Me and Mrs Jones Call me irresponsible Michael Bublé Michael Bublé
2 Me and Mrs
Jones[Remix]
Call me irresponsible Michael Bublé Michael Bublé

9.
EM - Deep Learning Based Blocking
Autoblock
•Token embedding: A word-embedding model transforms each
token to a token embedding
•Attribute embedding: For each attribute value of a tuple, an
attention-based neural network encoder converts the input sequence
of token embeddings to an attribute embedding
•Tuple signature: Multiple signature functions combine the attribute
embeddings of each tuple and produce multiple tuple signatures (one
per signature function)
•Model training: Equipped with the positive label set, the model is
trained with an objective that maximizes the differences of the cosine
similarities between the tuple signatures of matched pairs and
between unmatched pairs
•Fast NN search: The learned model is applied to compute the
signatures for all tuples, and an LSH family for cosine similarity is used
to retrieve the nearest neighbors for each tuple to generate candidate
pairs for blocking
Architecture of Autoblock

10.
Self-Supervised Blocking for entity matching - DeepBlock
• Encoder-Decoder for self-supervision
• Take a tuple 𝑡, feed it into a neural network (NN) to output a compact embedding vector u𝑡 ,
such that if we feed u𝑡 into a second NN, we can recover the original tuple 𝑡 (or a good
approximation of 𝑡). If this happens, u𝑡 can be viewed as a good compact summary of tuple 𝑡, and
can be used as the tuple embedding of 𝑡. The above two NNs are called encoder and decoder,
respectively.
• We can create u𝑡 in various ways. For example, one can feed in original tuple t with few attributes
missing and recover an original tuple with all the attributes (this would help our model match two
entities even when one of the entities has missing attributes)
encoder decoder entity
entity ut

11.
After Blocking – The computationally expensive
comparisons
• Edit distance between strings
• Embedding distance between entities
• Phonetic similarity
• Length
• Jaro Winkler
• Sequence Matcher
• Jaccard Similarity
• Entity specific Features
Reference: Location Matching Kaggle Competition https://www.kaggle.com/code/icfstat/lightgbm-feature-engineering-training-0-888-pv-lb

12.
How to measure success of Entity Matching?
For Model verification in isolation – Precision, Recall
Business Metrics – Business specific
Can we know if we did the perfect entity matching?
 One use-case, let’s say you are using entity matching for social media profile duplication. There is no perfect way of
knowing if all the duplicate/redundant profiles are identified
Human in the loop
 Sample the population and know if entity matching has worked for your cases
 Be mindful of testing with various size of samples and various types of mixtures of populations
Note: You can also check these benchmarks defined by Machamp (https://github.com/megagonlabs/machamp)

13.
Applications of Entity Matching
Song Matching
Address Matching
Social Profile Matching (Facebook, Twitter)
Clothes Matching (for that matter any item in retail)
Matrimony, Dating sites profile matching

14.
Entity Embed
Installation
pip install entity-embed
Preparing the data
Data needs to be a list of dict objects which must contain ‘id’ and ‘cluster’
[{'id': 0,
'cluster': 0,
'title': '001-Berimbou',
'artist': 'Astrud Gilberto',
'album': 'Look to the Rainbow (2008)'},
{'id': 1,
'cluster': 0,
'title': 'Berimbau',
'artist': 'Astrud Gilberto',
'album': 'Look to the Rainbow (1966)'},
{'id': 2,
'cluster': 1,
'title': '4 - Marooned - Pink Floyd',
'artist': '',
'album': 'The Division Bell'}]

15.
Entity Embed
Defining the fields
We need to define how record fields will be numericalized and encoded by Entity Embed’s deep neural network
field_config_dict = {
'title': {
'field_type': "MULTITOKEN",
'tokenizer': "entity_embed.default_tokenizer",
'alphabet': DEFAULT_ALPHABET,
'max_str_len': None, # compute
},
'title_semantic': {
'key': 'title',
'field_type': "SEMANTIC_MULTITOKEN",
'tokenizer': "entity_embed.default_tokenizer",
'vocab': "fasttext.en.300d",
}
}

16.
Entity Embed
Building the model
Under the hood, Entity Embed uses pytorch-lightning, so we need to create a datamodule object:
from entity_embed import DeduplicationDataModule
datamodule = DeduplicationDataModule(
train_record_dict=train_record_dict,
valid_record_dict=valid_record_dict,
test_record_dict=test_record_dict,
cluster_field="cluster",
record_numericalizer=record_numericalizer,
batch_size=32,
eval_batch_size=64,
random_seed=42,
)

17.
Entity Embed
Training the model
We must choose the K of the Approximate Nearest Neighbors, i.e., the top K neighbors our model will use
to find duplicates in the embedding space.
from entity_embed import EntityEmbed
model = EntityEmbed(
record_numericalizer,
ann_k=100,
)

18.
Entity Embed
Finding candidate pairs
When running in production, you only have access to the trained model object and the production record_dict
(without the true clusters filled, of course). You can get the embedding vectors of a production record_dict using
the predict method:
vector_dict = model.predict(
record_dict=test_record_dict,
batch_size=64
)
But what you usually want instead is the ANN pairs. You can get them with the predict_pairs method:
found_pair_set = model.predict_pairs(
record_dict=test_record_dict,
batch_size=64,
ann_k=100,
sim_threshold=0.3,
)

19.
References
1) Deep Learning for Blocking in Entity Matching
2) Autoblock
3) Entity Embed
4) Location Matching