Data processing: speech signal processing – linguistics – language – Linguistics
Reexamination Certificate
1998-07-23
2002-03-12
Thomas, Joseph (Department: 2644)
Data processing: speech signal processing, linguistics, language
Linguistics
C704S009000, C434S353000
Reexamination Certificate
active
06356864
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
This invention relates to methods that compute the quality of a text that purports to convey information on a topic by computing the amount of relevant subject-matter contained in the sample text and by computing the similarity of that subject-matter information to that contained in at least one standard reference text.
2. Description of Related Art
A need exists, in the educational field, to automate the process of evaluating the quality of essays. No automated process exists for providing quick evaluation of a student's subject matter knowledge and ability to articulate that knowledge in writing. An essay tests the depth of the writer's subject-matter knowledge, and ability to express ideas, but currently no automatic process exists for a quick and through evaluation of the essay.
Currently, the process of grading an essay with regard to quality and quantity of subject matter information and semantic coherence is labor intensive, tedious, and subjective. Whether the essay is written for an in-class project, in-dass test, during a nation-wide admissions test or an assessment test, each person grading an essay spends time analyzing the quality and quantity of subject matter content. Often two graders will compute different scores for the same essay. A need exists for a fair and objective method to grade essays, but the method must also reduce the amount of time spent grading the essay. In addition, this method must be as reliable and accurate as a grade a teacher, professional exam grader, or subject matter expert would assign to the text.
Related automated systems evaluate text with respect to spelling, grammar, sentence length, number of sentences within a paragraph, frequency of words, and total number of words written. These systems do not evaluate the text based on the amount and relevance of subject matter information, and coherence of the writing. Sentences within a paragraph can be grammatically correct, but these sentences can fail to answer the question presented or follow any coherent structure. In addition, these systems do not diagnose and provide information indicating which components of subject matter the essay should or should not contain, and indicate the weak areas in an essay.
Several automated systems use natural language queries for document retrieval. These systems match the words or phrases within the query to documents within a database. The query result lists in descending order the name of documents that contain the greatest number or some other function of the combination of words within the query. Other automated document retrieval systems develop information based on word proximity and importance of the word. If different words are used in query and text, these automated systems cannot evaluate the text based on the amount and relevance of subject matter information, without the addition of intellectually created thesauri, or semantic routines.
Latent Semantic Indexing (LSI) is well known in the art. U.S. Pat. No. 4,839,853, issued on Jun. 13, 1989 to Deerwester et al., and U.S. Pat. No. 5,301,109, issued on Apr. 5, 1994 to Landauer et al, utilize Latent Semantic Indexing to model the underlying correlational structure of the distribution of terms in documents. Computer-based document retrieval and multi-language retrieval systems utilize this process to statistically analyze a text database and an information query and estimate the correlation of topical meaning between words in the query and documents within a database. The '853 and '109 patents present a detailed discussion of the mathematics underlying the LSI techniques. There is no teaching in the art regarding the application of LSI to evaluate the relevant quality and quantity of semantic-content or the coherence of a writing.
SUMMARY OF THE INVENTION
This present invention solves the above and other problems, thereby advancing the useful arts, by providing an automated method for analysis and evaluation of semantic content of text. This method does not utilize conventional expert systems, artificial intelligence, or natural language processing. Instead, this method uses statistical analysis, hereinafter Latent Semantic Analysis (LSA) (based in part on known LSI methods), to analyze a corpus of representative text and thereafter to evaluate a sample text, such as an essay. This method assigns a numerical score or grade to a sample text. This score is based on the quality, quantity and relevance of subject matter content in the sample text as compared to one or more standard references, and in some cases aspects of the coherence of the writing.
This invention presumes there is an underlying latent semantic structure in the usage of words that is not easily evident due to the variety of words in our vocabulary that have similar meanings. The variety of words in our vocabulary gives the writer great flexibility to express concepts. The words the writer chooses depends upon the writer's experience or education. Consequently, two writers may choose different words to describe similar information on a particular subject matter topic. This latent semantic structure can be expressed mathematically. A mathematical relation can be computed to determine the relevant variety of words that express a particular topic.
LSA uses a data matrix comprising words and passages of sufficient length to express a full idea. LSA statistically analyzes this body of information in the data matrix to mathematically determine direct and indirect relationships between the words and the various contexts of the words. The statistical analysis provides a trained set of matrices from which all sample texts can be compared to estimate the quality, quantity, and relevance of subject matter content in the each sample text. Additionally, the coherence of the writing in a sample text is also estimated using LSA.
LSA ignores the order of words within the passage, and defines the meaning of a passage as the average of the vectors of the words it contains. Specifically, LSA computes the sum of each word in each vector representing the passage then computes the average of this value. The meaning of the passage can be similarly computed as the sum of the vectors of the words the passage contains. This definition represents the passage as a single vector. The passage can be a phrase, a paragraph, a portion of a text, or the whole text.
Specifically, this invention uses a set of reference documents to create a data matrix which defines the domain of knowledge. Reference documents, for example, can be authoritative text on the subject-matter, the text used to learn the subject-matter, articles on the subject-matter, or essays written by fellow students.
A data matrix is created from unique terms used in two or more reference documents. Terms are referred to as text objects and are used to form rows, the first dimension, within the data matrix. Text objects are unique words, concepts, or phrases. In the data matrix, each element of the text object vector represents of the number of times the text object is used in two or more reference documents.
A segment vector represents the individual reference documents. The segment vector is used to form columns, the second dimension, within the data matrix. A segment vector, for example, can be an entire reference text, abstract of a reference text, title of a reference text, at least one paragraph of a reference text, at least one sentence of a reference text, or a collection of text objects that convey an idea or summarize a topic. Each reference text is associated with a single segment vector within the data matrix.
The intersection of a row and column, a matrix cell, contains the number of times a particular text object appears in a particular segment. A weighted value is applied to each cell value. The resulting cell is a proportional representation of the importance of the cells original information, for example, rare words are weighted more heavily.
Singular value decomposition is applied to the data matrix to
Foltz Peter William
Kintsch Walter
Laham, II Robert Darrell
Landauer Thomas K.
Rehder Robert Ernest
Fishman Daniel N.
Lathrop & Gage L.C.
Thomas Joseph
University Technology Corporation
LandOfFree
Methods for analysis and evaluation of the semantic content... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Methods for analysis and evaluation of the semantic content..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Methods for analysis and evaluation of the semantic content... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2844407