Educational Testing Service

Michael Flor

Beata Beigman Klebanov

Swapna Somasundaran

Michael Heilman

Martin S Chodorow

Brian W Riordan

Xianyang Chen

Derrick Higgins

Paul Douglas Deane

Chee Wee Leong

Isaac I Bejar

Mo Zhang

Yigal Attali

Yoko Futagi

Gary Feng

Jiangang Hao

Robert A Pugh

Hillary Molloy

Binod Gyawali

Nathan Lederer

Michael E Wagner

Yoav Bergner

Daniel Blanchard

F Jay Breyer

Laura McCulla

Kenneth Steimel

Aoife Cahill

Data is received that includes a passage of text. The passage of text can be tokenized so that features can be extracted from the resulting tokens. One or more machine learning models can detect one or more metaphors within the passage of text using the extracted features. The at least one machine learning model can be trained, for example, by interleaving data relating to metaphors with data relating to one or more auxiliary tasks associated with related figurative language constructs. Data can then be provided that identifies the detected metaphors. Related apparatus, systems, techniques and articles are also described.

Machine learning-based metaphor detection

Quality of a narrative is characterized by receiving data that includes a narrative text. This narrative text is then tokenized and events are extracted from the tokenized words. The extraction can use, in parallel, two or more different extraction techniques. The extracted events are then extracted so that a waveform can be generated based on the aggregated extracted events that characterizes a plurality of emotional arcs within the narrative text. Subsequently, a plurality of waveform elements are extracted from the waveform. The narrative quality (or other quality) of the narrative text is then scored based on the extracted plurality of waveform elements and using a machine learning model trained to correlate emotional arc waveforms with narrative quality scores. Related apparatus, systems, techniques and articles are also described.

Machine learning-based automated narrative text scoring including emotion arc characterization

A computer-implemented technique for characterizing lexical concreteness in narrative includes receiving data encapsulating narrative text having a plurality of words. Thereafter, the function words can be removed from the narrative text to result in only content words. A concreteness score can then be assigned to each content word by polling a database to identify matching words and to use concreteness scores associated with such matching words as specified by the database. Data can then be provided which characterizes the assigned concreteness scores. Related apparatus, systems, techniques and articles are also described.

Lexical concreteness in narrative text

Systems and methods are provided for processing a group of essays to develop a classifier that detects nonsensical computer-generated essays. A data structure associated with a group of essays is accessed, wherein the group of essays includes nonsensical computer-generated essays and good-faith essays. Both the nonsensical computer-generated essays and the good-faith essays are assigned feature values. The distribution of feature values between the nonsensical computer-generated essays and the good-faith essays is measured. A classifier that detects whether an essay is a nonsensical computer-generated essay is developed, wherein the classifier is developed using the distribution of feature values.

Developing an e-rater advisory to detect babel-generated essays

Systems and methods are disclosed for automatically scoring a constructed response using a neural network. In embodiments, a constructed response received by a processing system may be processed to divide the constructed response into multiple series of word tokens, wherein each word token includes a sequence of characters. The constructed response may be further processed to correct one or more spelling errors. The word tokens may be encoded to generate representation vectors for the constructed response. A set of nonlinear operations may be applied to the plurality of representation vectors in a neural network to generate a single vector output. A set of predetermined network weights may be applied to the vector output of the neural network to generate a scalar output for scoring the constructed response.

Systems and methods for neural content scoring

Computer-implemented systems and methods are described herein for automatically generating questions from text. Text including one or more sentences is received. A sentence, comprising a predicate and one or more arguments associated with the predicate, is parsed from the text. Semantic role labels are assigned to the one or more arguments associated with the predicate. One or more questions are automatically generated relating to the predicate based on the assigned semantic role labels. Each answer to the generated questions is one of the one or more arguments associated with the predicate.

Systems and methods for automatic generation of questions from text

Systems and methods are provided for processing a response to essay prompts that request a narrative response. A data structure associated with a narrative essay is accessed. The essay is analyzed to generate an organization subscore, where the organization subscore is generated using a graph metric by identifying content words in each sentence of the essay and populating a data structure with links between related content words in neighboring sentences, wherein the organization subscore is determined based on the links. The essay is analyzed to generate a development subscore, where the development subscore is generated using a transition metric by accessing a transition cue data store and identifying transition words in the essay, wherein the development subscore is based on a number of words in the essay that match words in the transition cue data store. A narrative quality metric is determined based on the organization subscore and the development subscore.

Platform for administering and evaluating narrative essay examinations

Methods and systems for scoring written text based on use of idiomatic expressions, including reading pre-selected idiomatic expressions in a canonical form into memory, expanding idiomatic expressions from the canonical form, reading a written response into the memory, pre-processing the written response, searching the pre-processed written response for idiomatic expressions, and assigning a score to the written response. The score may be based at least in part on the number of idiomatic expressions in the written response. Corresponding apparatuses, systems, and methods are also disclosed.

Systems and methods for automatic detection of idiomatic expressions in written responses

Systems and methods are provided for a computer-implemented method for identifying pairs of cohesive words within a text. A supervised model is trained to detect cohesive words within a text to be scored. Training the supervised model includes identifying a plurality of pairs of candidate cohesive words in a training essay and an order associated with the pairs of candidate cohesive words based on an order of words in the training essay. The pairs of candidate cohesive words are filtered to form a set of evaluation pairs. The evaluation pairs are provided via a graphical user interface based on the order associated with the pairs of candidate cohesive words. An indication of cohesion or no cohesion is received for the evaluation pairs via the graphical user interface. The supervised model is trained based on the evaluation pairs and the received indications.

Computer-implemented systems and methods for generating a supervised model for lexical cohesion detection

A computer-implemented method of training an assessment model for assessing constructed texts expressing opinions on subjects includes accessing a plurality of training texts, which are constructed texts. The training texts are analyzed with the processing system to derive values of a plurality of linguistic features of an assessment model. At least one of the plurality of linguistic features relates to sentiment and at least one of the plurality of linguistic feature relates to specificity. The assessment model is trained with the processing system based on the values of the plurality of linguistic features. Based on the training, a weight for each of the plurality of linguistic features is determined. The assessment model is calibrated to include the weights for at least some of the plurality of linguistic features such that the assessment model is configured to generate assessment measures for constructed texts expressing opinions on subjects.

Systems and methods for automatically assessing constructed recommendations based on sentiment and specificity measures

Systems and methods are provided for scoring a constructed response generated by a user and providing information on the user's writing behavior. A constructed response and associated electronic process log are received. The constructed response is processed to generate first feature values representative of aspects of the constructed response. The electronic process log is processed to generate second feature values related to the user's actions in generating the constructed response. A score for the constructed response is generated using the processing system by applying a computer scoring model to the first and second feature values. A rule of a rule engine that is satisfied is identified, the rule being satisfied when one or more feature values of the second feature values meet a condition associated with the rule. Information on the user's actions in generating the constructed response is provided based on the satisfied rule.

Generating scores and feedback for writing assessment and instruction using electronic process logs

Systems and methods are provided for determining a susceptibility of a computer-implemented automated scoring engine to gaming strategies. A plurality of responses to a prompt are provided to a computer-implemented automated scoring engine to receive a first set of scores. A first transformation is performed on each of the plurality of responses to generate a first set of transformed responses. The first set of transformed responses is provided to the computer-implemented automatic scoring engine to receive a second set of scores, and a gaming susceptibility metric is determined based on the first set of scores and the second set of scores.

Systems and methods for evaluating susceptibility of automated scoring engines to gaming behavior

Computer-implemented systems and methods are provided for determining a document's complexity. For example, a computer performing the complexity analysis can receive a document. The computer can determine the content words within the document and determine an association measure for each group of content words. An association profile can be created for the document using the association measures. The computer can use the association profile to determine the complexity of the document. The complexity of the document may correspond to the document's suitable reading level or, if the document is an essay, an essay score.

Systems and methods for determining text complexity

Systems and methods described herein utilize supervised machine learning to generate a figure-of-speech prediction model for classify content words in a running text as either being figurative (e.g., as a metaphor, simile, etc.) or non-figurative (i.e., literal). The prediction model may extract and analyze any number of features in making its prediction, including a topic model feature, unigram feature, part-of-speech feature, concreteness feature, concreteness difference feature, literal context feature, non-literal context feature, and off-topic feature, each of which are described in detail herein. Since uses of figure of speech in writings may signal content sophistication, the figure-of-speech prediction model allows scoring engines to further take into consideration a text's use of figure of speech when generating a score.

Systems and methods for constructed response scoring using metaphor detection

Systems and methods are provided for identifying one or more target words of a corpus that have a lexical relationship to a plurality of provided cue words. The cue words and statistical lexical information derived from a corpus of documents are analyzed to determine candidate words that have a lexical association with the cue words. The statistical information includes numerical values indicative of probabilities of word pairs appearing together as adjacent words in a well-formed text or appearing together within a paragraph of a well-formed text. For each candidate word, a statistical association score between the candidate word and each of the cue words is determined. An aggregate score for each of the candidate words is determined based on the statistical association scores. One or more of the candidate words are selected to be the one or more target words based on the aggregate scores.

Growing community of inventors

Lawrenceville, NJ, United States of America

Michael Flor