Value-directed parsing for data extraction
AssigneeZoho Corporation Private Limited
InventorsJayaraj Poroor, Harish Chandrasekar Reddy, Raghul Selvaraj, Bharath Sivakumar
Life of this patent
Estimating the term…
Abstract
Described are methods and systems for parsing unstructured or semi-structured text to extract named entities, data types defined to include semantic fields. Fields are constrained to sets of potential field values. These sets can overlap, leading to ambiguous parses. For example, the text string “3-4-2023” parsed as a date can yield Mar. 4, 2023 or Apr. 3, 2023. Potential field values in alternative parses are scored and the scores used to select and disambiguate the resultant value parses.
Claims15
A method of parsing a text string of characters, on a computing system that includes at least one processor and memory, to produce a target value parse with semantic fields, including a first semantic field and a second semantic field:
- extracting tokens from the text string using the at least one processor, the tokens including every possible combination of adjacent characters in a direction of the text string;
- applying, by the at least one processor, field constraints to each of the tokens to produce potential field values for the first semantic field and the second semantic field;
- applying, by the at least one processor, scores to each of the potential field values;
- ordering a first set of the potential field values in a first value parse;
- combining the scores applied to the first set of the potential field values to obtain a first parse score for the first value parse;
- ordering a second set of the potential field values in a second value parse;
- and
- combining the scores applied to the second set of the potential field values to obtain a second parse score for the second value parse;
- and
- selecting one of the first value parse and the second value parse responsive to the first parse score and the second parse score, wherein the first parse score equals the second parse score, the method further comprising disambiguating the first value parse from the second value parse responsive to the first parse score and the second parse score;
- wherein the disambiguating comprises:
- assigning a datatype to the first value parse and the second value parse;
- applying a first bias to the scores applied to the first set of the potential field values;
- updating the first parse score responsive to the first bias;
- applying a second bias to the scores applied to the second set of the potential field values;
- updating the second parse score responsive to the second bias;
- wherein the first bias and the second bias favor an order of the first semantic field relative to the second semantic field;
- and
- selecting between the updated first parse score and the updated second parse score to disambiguate between the first value parse and the second value parse of the datatype.
A method of parsing a text string of characters, on a computing system that includes at least one processor and memory, to produce a target value parse with semantic fields, including a first semantic field and a second semantic field:
- extracting tokens from the text string using the at least one processor, the tokens including every possible string of at least one of the characters along a direction of the text string;
- applying, by the at least one processor, field constraints to each of the tokens to produce potential field values for the first semantic field and the second semantic field;
- applying, by the at least one processor, scores to each of the potential field values;
- ordering a first set of the potential field values in a first value parse;
- combining the scores applied to the first set of the potential field values to obtain a first parse score for the first value parse;
- ordering a second set of the potential field values in a second value parse;
- combining the scores applied to the second set of the potential field values to obtain a second parse score for the second value parse;
- and
- selecting one of the first value parse and the second value parse responsive to the first parse score and the second parse score, wherein the first parse score equals the second parse score, the method further comprising disambiguating the first value parse from the second value parse responsive to the first parse score and the second parse score;
- wherein the first value parse is a maximum value parse, and wherein the disambiguating comprises grouping the first and second value parses with additional value parses that lie within a threshold of the maximum value parse.
Dependencies are inferred from the claim text.