Breviter is the reverse semantic search for Geolexica.
Ever forgotten a specific word but could describe its meaning? Internet search engines more than often return unrelated entries. The solution is reverse semantic search: given an input of the meaning of the word (search phrase), provide an output with dictionary words that match the meaning.
This often happens especially in technical domains and in standardization vocabulary, where terms can be highly specific and difficult to locate without specialized knowledge.
The key to accurate reverse search lies in the machine's ability to understand semantics. Geolexica employs deep learning approaches in natural language processing (NLP) to enable better comparison of meanings between the search phrases with word definitions. Accuracy will be significantly increased.
The ability to identify entities with similar semantics facilitates ontology discovery in the Semantic Web and in Technical Language Processing (TLP).
This project is funded by NLnet to develop a reverse semantic search engine (including experiments and a report of the experiment results).
The project proposal (including plan and deliverables) can be found on our Google drive.
- The Geolexica ISO/TC 211 concepts database is submoduled under this repo.
- The AI model for English is done.
- A demo (for English) and a bulk test platform are provided.
- We need to deploy the demo on isotc211.geolexica.org
The ISO/TC 211 Geolexica concept dataset (YAML) is submoduled under 'data/'.
To update the database, simply go inside 'data/' and run git pull.
We use Tensorflow.js (the JavaScript version of Tensorflow).
The pre-trained model used is MobileBERT:
- From the paper (2020) MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou for details about the model.
MobileBERT, is a compact version of BERT and so execution of the model is faster with a minimal tradeoff in accuracy.
The model is located in the 'sent' folder under the 'public' folder. To update the model, simply replace the model files in this folder.
The demo application is developed using Next.js.
It loads the Geolexica dataset, its pre-computed vectors and MobileBERT in the browser.
There are two implemented UIs:
- Single search interface
- Bulk search interface
The Yarn package manager is used for package management.
To start the app, first install all the dependencies:
yarn install
Then compute the dataset using the model:
yarn compute
# which is a shorthand for:
# yarn cli compute data -o public/db.json
Then, start the web server:
yarn dev
To access the app, use any browser and go to:
localhost:3000
The single reverse search interface is located at:
localhost:3000/reverse
It uses the Geolexica concepts dataset with the English definitions.
The result is a list of words with the highest similarity scores with a given search phrase.
The bulk search interface is meant for performing experiments and collecting accuracy data for the report.
It is located at:
localhost:3000/testing
This interface requires uploading a file that contains multiple test cases.
The structure of the file is:
- Each test should be on a separate line.
- Each test case has two items, separated by a semicolon
;.- The first part is the search phrase of the test case.
- The second part is the expected resulted term.
For example, if the search phrase is terms not in use and the expected result
is obsolete term, the test case should be written as
terms not in use;obsolete term
Samples of the test cases can be found under the 'public/testcases/' directory.
-
Prepare additional test cases.
-
Collect accuracy statistics of our platform and compare to the existing competitors (Reverse dictionary and OneLook). As there is no API from the competitors, we either need to perform the tests manually or write some scripts to do the tests. To collect the accuracy statistics of our platform, we can use UI #2.
-
Supporting other languages. Currently, there is only English model that is publicly available for the Tensorflow.js format. To achieve this, there are two possible ways that are simple enough. First, there are models for other languages in other formats. There could be transformer tools to transform models of other format into the formats supported by Tensorflow.js. Second, there could be JS/TS reader APIs that load AI models that can be used by Tensorflow.js. This problem is probably faced by many other developers but I haven't studied in depth in this area yet.
Tensorflow.js pre-trained models are available at:
Off-the-shelf Tensorflow models can be found at:
NOTE: The Tensorflow.js model format is specific -- normal Tensorflow models must be converted to this format for them to be useable. Non-Tensorflow formats are not directly usable by Tensorflow.js. More details about the format can be found here.
Link to the original BERT model and paper from Google:
While there is a multilingual version of the model that supports 102 languages in one single model, it is unsure how we can make it usable in the current application.
Didn't have time to study it in details. Probably worth to study how to make it usable in a web environment.
The multilingual BERT model is native to python. Maybe it is simpler if the web can invoke some python codes.