Back to results
US 12,499,142 B2Granted
Method and apparatus with Arabic information extraction and semantic search
AssigneeElm
InventorsMuhammad Saleh Al-Qurishi, Sarah Ahmed Al-Qaseemi, Riad Souissi
Life of this patent
Estimating the term…
Abstract
An Arabic information extraction apparatus includes one or more processors configured to: receive a query comprising a long query or a short query; extract, using one or more language models, a named entity and a keyword from the query to generate extracted information; classify, using one or more classification models, the query to generate a classified query; convert the classified query and the extracted information into a dense vector representation; and determine and output a similarity match between the dense vector representation and a document vector representation of a knowledge base comprising an Islamic law document.
Claims17
An apparatus for search of Arabic legal cases implemented in hardware or in a combination of hardware and software, the apparatus comprising:
- a memory storing a knowledge base maintaining a plurality of Arabic law documents and a connection layer configured to establish logical relationships between the Arabic law documents and knowledge entities;
- an interactive display device configured to input a query comprising a long query or a short query of Arabic text;
- one or more processors configured to:
- receive the input query;
- extract Arabic legal text data from the received query to generate extracted information, the Arabic legal text data comprising prefixes, stems, and suffixes utilizing a morphological analysis algorithm based on a predefined Arabic-English lexicon, the predefined Arabic-English lexicon comprising morphological compatibility tables for prefix-stem combinations and stem-suffix combinations;
- segment the received query into sentences extracted from Arabic legal texts, wherein each of the sentences does not exceed a predetermined token limit;
- determine semantic importance of the sentences extracted using a deep learning model comprising multiple layers of bidirectional transformer encoders;
- select key sentences of the sentences extracted based on determined semantic relevance scores of the sentences extracted, relative to the Arabic legal text data, to generate a classified query;
- parse the classified query by removing duplicated white spaces, unwanted characters, and normalizing numerals;
- extract, using a language model, a named entity and a keyword from the received query to generate extracted information;
- convert the classified query and the extracted information into a dense vector representation by removing duplicated white spaces, unwanted characters, and normalizing numerals;
- determine a similarity match of the Arabic legal cases between the dense vector representation and a document vector representation of the plurality of Arabic law documents and logically related knowledge entities stored in the knowledge base, the Arabic law documents comprising Arabic legal documents, legislative laws, Fiqh texts, Quran, Hadith texts, royal decrees, Arabic Legal Content (ALC) raw data, and Sharia ruling documents;
- and
- the interactive display device configured to display a Arabic document, or the plurality of Arabic documents, as a result of the similarity match;
- wherein the knowledge base further comprises a database schema configured to store and organize the Arabic law documents into structured tables, and the connection layer configured to establish relationships between different tables in the knowledge base.
12 more elements
A computer-implemented Arabic search method implemented in hardware or in a combination of hardware and software, the method comprising:
- receiving inputting, by an interactive display device, a query comprising Arabic text;
- extracting, using one or more language models, Arabic legal text data from the received query to generate extracted information, the Arabic legal text data comprising prefixes, stems, and suffixes utilizing a morphological analysis algorithm based on a predefined Arabic-English lexicon, the predefined Arabic-English lexicon comprising morphological compatibility tables for prefix-stem combinations and stem-suffix combinations;
- segmenting the received query into sentences extracted from Arabic legal text data, wherein each of the sentences does not exceed a predetermined token limit;
- determining semantic importance of the sentences extracted using a deep learning model comprising multiple layers of bidirectional transformer encoders;
- selecting key sentences of the sentences extracted based on determined semantic relevance scores of the sentences extracted, relative to the Arabic legal text data, to generate a classified query;
- parsing the classified query by removing duplicated white spaces, unwanted characters, and normalizing numerals;
- extracting, using a language model, a named entity and a keyword from the received query to generate extracted information;
- converting the classified query and the extracted information into a dense vector representation by removing duplicated white spaces, unwanted characters, and normalizing numerals;
- determining a similarity match of Arabic legal cases between the dense vector representation and a document vector representation of a plurality of Arabic documents and logically related knowledge entities stored in a knowledge base, the Arabic documents comprising Arabic legal documents, legislative laws, Fiqh texts, Quran, Hadith texts, royal decrees, Arabic Legal Content (ALC) raw data, and Sharia ruling documents, wherein the received query comprises one of a long query of facts of a legal case and a short query of facts of an Arabic legal case, and
- wherein in response to the received query comprising the long query, the long query does not include a punctuation mark;
- and
- displaying, by the interactive display device, an Arabic document, of the plurality of Arabic documents, as a result of the similarity match;
- wherein the knowledge base further comprises a database schema configured to store and organize the Arabic law documents into structured tables, and the connection layer configured to establish relationships between different tables in the knowledge base.
10 more elements
Dependencies are inferred from the claim text.