A homework grading method based on image recognition

Abstract

The present invention discloses a homework grading method based on image recognition, which relates to the technical field of homework grading. The present invention comprises the following steps: breaking through the surface processing limitations of existing homework grading technologies on cursive characters, integrating cursive deformation into an expression intention analysis framework through the linkage of stroke topological features and semantic compensation, and being able to identify metaphors deliberately emphasized by students, such as "as" with elongated strokes, to avoid misjudging rhetorical structures as word errors; eliminating the fragmentation of long rhetorical sentences by physical paging through forced merging of cross-line syntax trees, ensuring complete parsing of emphasized sentences such as "not only...but also"; and converting abstract grammatical rules into perceptible scenarios through a dynamic visualization mechanism: a timeline animation of tense errors intuitively presents the misalignment of actions and time nodes, and in-depth rhetorical analysis triggered by hand-drawn wavy lines accurately marks the metaphor subject-metaphor mapping relationship, enabling students to understand the essence of the error at a glance.

Description

Technical Field

[0001] The present invention relates to the technical field of homework grading, and in particular to a homework grading method based on image recognition. Background Art

[0002] With the in-depth application of artificial intelligence in the field of education, homework grading systems based on image recognition have become an important auxiliary tool for English teaching. The current mainstream solutions generally adopt an end-to-end deep learning framework: using convolutional neural networks (CNN) to extract handwritten image features, combined with long short-term memory networks (LSTM) for sequence recognition, and finally connecting to natural language processing (NLP) modules to complete grammar checking and simple comment generation. This type of technology is effective for printed homework, but when processing students' actual handwritten homework, cursive characters may cause misjudgment of rhetorical devices.

[0003] To this end, some solutions use the attention mechanism to dynamically focus on the key structural points of cursive letters, such as distinguishing between u and v through stroke endpoint detection, or using the Transformer model to perform semantic correction on the recognition results, for example, completing "Ifeel so h_ppy" to "happy". However, these solutions only stay at the single-word recognition level in terms of optimization, and do not establish the connection between glyph features and the rhetorical function of the sentence. When students write metaphorical sentences in cursive, such as "Her smile is a warm sun", the strokes of "a" and "s" are deformed. The system may misjudge the "warm sun" proper noun and ignore the metaphorical structure due to excessive focus on the continuity of local strokes. Existing models process handwriting as static images and cannot capture traces of expression strategies during the writing process. For example, the tail stroke that students deliberately lengthen to emphasize the metaphor word "ocean" is simplified to an ordinary cursive word.

[0004] Some solutions also introduce stroke trajectory reconstruction algorithms, such as Apple Pencil's handwriting prediction technology, which reverses the writing order from scanned images, or builds rhetoric-sensitive NLP models to add rhetorical tags such as metaphor and parallelism to syntax tree analysis. However, the reconstructed handwriting sequence is difficult to accurately match with NLP rhetorical tags, resulting in a separation of form and meaning. The serial connection of multiple models increases the computational cost exponentially, making it difficult to deploy on mobile devices. Therefore, there is an urgent need for a lightweight algorithm that can achieve end-to-end parsing of rhetorical structures while retaining the original writing characteristics of cursive characters. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides an image recognition-based homework grading method to solve the problems in the prior art caused by the fragmentation of cursive character analysis, lack of cross-line processing and mechanized error feedback, which lead to rhetorical misjudgment and insufficient personalized diagnosis.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] An embodiment of the present invention provides a method for grading homework based on image recognition, which includes:

[0009] Step S1: collecting the original image of the student's handwritten homework and identifying the text area and writing trajectory;

[0010] Step S2, extracting stroke topological features for the connected writing area;

[0011] Step S3: Generate a dynamic semantic weight table based on the stroke topology features and perform semantic compensation in combination with the context word vector;

[0012] Step S4, performing rhetorical structure analysis on the compensated text, marking the starting and ending positions and types of metaphors and parallelisms;

[0013] Step S5: Generate visual annotations, including error explanation animations and beautiful word and sentence tags.

[0014] As a preferred solution of the homework grading method based on image recognition described in the present invention, the stroke topology feature extraction in step S2 includes:

[0015] Break down connected letters into discrete stroke segments;

[0016] Calculate the curvature change sequence of each stroke segment and mark the curvature mutation points;

[0017] Count the number of closed loops formed by the stroke segments;

[0018] When the number of closed loops deviates from the standard glyph by more than the tolerance value, the adjacent stroke space vector analysis is activated:

[0019] Establishing a spatial position coordinate system for stroke segments;

[0020] Calculate the angle and distance vector between adjacent stroke segments.

[0021] As a preferred solution of the homework grading method based on image recognition described in the present invention, the implementation of semantic compensation in step S3 includes:

[0022] Construct a dynamic word vector weight table containing a set of adjacent high-frequency words of the target vocabulary;

[0023] When the confidence level of a single word is lower than the threshold, weight adjustment is performed:

[0024] Get the three words before and after the current word as the semantic context window;

[0025] Calculate the semantic relevance between the candidate word and the context window.

[0026] As a preferred embodiment of the method for grading homework based on image recognition described in the present invention, in step S3, during the weighted compensation process of semantic relevance, when the OCR confidence level for a certain word is lower than a threshold, dynamic compensation is performed on the recognition result and fed back to the weight table, including:

[0027] Current location Both sides The word vectors are aggregated using exponential decay:

[0028] ,

[0029] ,

[0030] in, is the context window radius, is the relative position index, For location word vectors, is the indicator function, is the position weight coefficient, is the decay constant, is the context vector, To sum the intermediate variables;

[0031] Candidate words Calculate cosine similarity with context:

[0032] ,

[0033] in is the candidate word vector, is the semantic relevance;

[0034] Obtain candidate words at time by exponential smoothing Weight:

[0035] ,

[0036] in, and are the old and latest dynamic weights respectively, is the memory decay factor, is the frequency of candidate words, is the frequency smoothing constant, is the current update step;

[0037] The stroke topology features enhance the weight table in a multiplicative manner:

[0038] ,

[0039] in, and To enhance the front and back weight matrices, For Hadamard, is the topological enhancement coefficient, For each The correlation vector composed of

[0040] Then the dynamic weights are linearly fused with the original confidence and compressed by Sigmoid:

[0041] ,

[0042] in, To compensate for the confidence, is the original OCR confidence, is the fusion coefficient, is the Sigmoid function, with the maximum The word form replaces the original recognition result.

[0043] As a preferred embodiment of the method for grading homework based on image recognition according to the present invention, the processing of cross-line rhetorical structures in step S4 includes:

[0044] Locate the first line of rhetorical keywords: Identify the page coordinates of the preset rhetorical keywords in the first line of the sentence and record them as the starting coordinate point ;

[0045] Scan the associated keywords in the next row: Detect semantically associated continued keywords in the preset area at the top of the next row. The preset area satisfies:

[0046] Horizontal deviation constraint: ;

[0047] Vertical extent constraints: ;

[0048] in, is the vertical coordinate of the top of the next row, is the page height;

[0049] Perform syntax tree merging: When a continuation keyword coordinate that satisfies the constraints is detected When the first line and the second line of text are combined into a single syntax tree for rhetorical structure analysis.

[0050] As a preferred solution of the homework grading method based on image recognition described in the present invention, the error explanation animation generation in step S5 includes:

[0051] Temporal errors correspond to timeline animations:

[0052] Extract explicit / implicit time adverbials in sentences;

[0053] The bidirectional time pointer is superimposed at the wrong verb;

[0054] The red pointer points to the correct tense time node;

[0055] Article errors show countable nouns visual model:

[0056] Generate 3D outlines of isolated objects of singular nouns;

[0057] Generate a dot plot of the population distribution of plural nouns.

[0058] As a preferred embodiment of the method for grading homework based on image recognition described in the present invention, the beautiful word and sentence marking includes:

[0059] Detect students' hand-drawn underlines / wavy lines:

[0060] The hand-drawn line was determined by the threshold of curvature variance > 0.8;

[0061] Initiate a deep rhetorical analysis of the sentence covered with hand-drawn lines:

[0062] Identify the subject-object mapping relationship in metaphor structure;

[0063] Detect the parallel structure integrity of parallel sentences;

[0064] A rhetorical type icon is superimposed at the corresponding position of the original image.

[0065] As a preferred embodiment of the method for grading homework based on image recognition described in the present invention, personalized error pattern analysis is also included:

[0066] Record repeated incorrect root words by the same student;

[0067] When a specific error type occurs 3 or more times cumulatively:

[0068] Analyze the linguistic characteristics of incorrect word roots;

[0069] Generate specialized training questions containing contrasting language examples.

[0070] As a preferred solution of the homework correction method based on image recognition described in the present invention, in the process of personalized error pattern analysis, when the same student triggers the same root error multiple times in the correction record, the error type is The propagation probability of is updated in three cycles in sequence, including:

[0071] Assume that the completed job number is , at a length of Record binary sequence in the sliding window , get the recent frequency:

[0072] ,

[0073] in, is the current job number, is the window length, is the relative position within the window, For serial number Error indication, is the position weight, is the decay constant, is the recent frequency;

[0074] For error types The propagation probability uses exponential recursion:

[0075] ,

[0076] in, and Sequential numbers 、 The probability of propagation, is the memory weight factor, is the smooth slope, For Error Cumulative number of occurrences, is the trigger threshold;

[0077] Calculate the upper bound of the error and decide whether to generate special training questions:

[0078] ,

[0079] in, is the normal 95% quantile constant, is the sample size in the window, is the half-width of the confidence interval, is the upper confidence limit, if , immediately push the root Special training tasks as the core.

[0080] As a preferred solution of the homework grading method based on image recognition described in the present invention, the timeline animation generation further includes:

[0081] For sentences without explicit time adverbials:

[0082] Extract time clue words from context paragraphs;

[0083] Build a logical sequence chain of events;

[0084] Inferring implicit temporal reference points based on event chains.

[0085] The beneficial effects of the present invention are as follows: the present invention breaks through the limitations of the existing homework grading technology on the surface processing of cursive characters, and by linking the stroke topology features with semantic compensation, incorporates cursive deformation into the expression intention analysis framework, and can identify metaphors that students deliberately emphasize, such as "as" with elongated strokes, to avoid misjudging rhetorical structures as word errors; by forcibly merging cross-line syntax trees, the present invention eliminates the fragmentation of long rhetorical sentences caused by physical paging, ensuring the complete analysis of emphasized sentences such as "not only...but also".

[0086] The dynamic visualization mechanism of the present invention transforms abstract grammatical rules into perceptible scenarios: the timeline animation of tense errors intuitively presents the misalignment of actions and time nodes, and the deep rhetorical analysis triggered by hand-drawn wavy lines accurately marks the metaphor subject-metaphor mapping relationship, allowing students to understand the essence of the error at a glance; the personalized error propagation model is upgraded from mechanical error correction to learning behavior diagnosis. When it is detected that students are interfered with by their mother tongue, such as those with a French background frequently making -ant suffix errors, comparative language case training questions are automatically pushed, forming a correction-diagnosis-compensation closed loop; implicit tense inference compensates for the lack of time adverbials through event logic chains, avoids misjudgment due to incomplete expression, and improves robustness to real corpus. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0088] Figure 1 This is a flow chart of the homework grading method based on image recognition in Example 1. DETAILED DESCRIPTION

[0089] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0090] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0091] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0092] Example 1, reference Figure 1 This embodiment provides a method for grading homework based on image recognition, comprising the following steps:

[0093] Step S1: collecting the original image of the student's handwritten homework and identifying the text area and writing trajectory;

[0094] Step S2, extracting stroke topological features for the connected writing area;

[0095] The stroke topology feature extraction in step S2 includes:

[0096] Break down connected letters into discrete stroke segments;

[0097] Calculate the curvature change sequence of each stroke segment and mark the curvature mutation points;

[0098] Count the number of closed loops formed by the stroke segments;

[0099] When the number of closed loops deviates from the standard glyph by more than the tolerance value, the adjacent stroke space vector analysis is activated:

[0100] Establishing a spatial position coordinate system for stroke segments;

[0101] Calculate the angle and distance vector between adjacent stroke segments;

[0102] Step S3: Generate a dynamic semantic weight table based on the stroke topology features and perform semantic compensation in combination with the context word vector;

[0103] The implementation of semantic compensation in step S3 includes:

[0104] Construct a dynamic word vector weight table containing a set of adjacent high-frequency words of the target vocabulary;

[0105] When the confidence level of a single word is lower than the threshold, weight adjustment is performed:

[0106] Get the three words before and after the current word as the semantic context window;

[0107] Calculate the semantic relevance between the candidate word and the context window;

[0108] In step S3, during the semantic relevance weighted compensation process, when the OCR confidence level for a word is lower than the threshold, dynamic compensation is performed on the recognition result and fed back to the weight table, including:

[0109] Current location Both sides The word vectors are aggregated using exponential decay:

[0110] ,

[0111] ,

[0112] in, is the context window radius, is the relative position index, For location word vectors, is the indicator function, is the position weight coefficient, is the decay constant, is the context vector, To sum the intermediate variables;

[0113] Candidate words Calculate cosine similarity with context:

[0114] ,

[0115] in is the candidate word vector, is the semantic relevance;

[0116] Obtain candidate words at time by exponential smoothing Weight:

[0117] ,

[0118] in, and are the old and latest dynamic weights respectively, is the memory decay factor, is the frequency of occurrence of candidate words, is the frequency smoothing constant, is the current update step;

[0119] The stroke topology features enhance the weight table in a multiplicative manner:

[0120] ,

[0121] in, and To enhance the front and back weight matrices, For Hadamard, is the topological enhancement coefficient, For each The correlation vector composed of

[0122] Then the dynamic weights are linearly fused with the original confidence and compressed by Sigmoid:

[0123] ,

[0124] in, To compensate for the confidence, is the original OCR confidence, is the fusion coefficient, is the Sigmoid function, with the maximum The word form replaces the original recognition result;

[0125] Specifically, this step uses position decay to perform weighted aggregation on context vectors, making neighboring words contribute more to candidate ranking, while the influence of distant words is suppressed exponentially while retaining the overall context. Dynamic weights use frequency-driven exponential smoothing to maintain historical stability while quickly responding to novel expressions. The connected stroke topology enhancement matrix amplifies the true correlation by using local stroke closure to suppress noise interference.

[0126] Step S4, performing rhetorical structure analysis on the compensated text, marking the starting and ending positions and types of metaphors and parallelisms;

[0127] Processing the interline rhetorical structure in step S4 includes:

[0128] Locate the rhetorical keywords in the first line: Identify the page coordinates of the preset rhetorical keywords (such as "not only") in the first line of the sentence and record them as the starting coordinate point ;

[0129] Scan the next row for related keywords: Detect semantically related continuation keywords (such as "but also") in the preset area at the top of the next row. The preset area meets the following requirements:

[0130] Horizontal deviation constraint: ;

[0131] Vertical extent constraints: ;

[0132] in, is the vertical coordinate of the top of the next row, is the page height;

[0133] Perform syntax tree merging: When a continuation keyword coordinate that satisfies the constraints is detected When , the first line and the second line of text are merged into a single syntax tree for rhetorical structure analysis;

[0134] Step S5, generating visual annotations, including error explanation animations and beautiful word and sentence markings;

[0135] The error explanation animation generation in step S5 includes:

[0136] Temporal errors correspond to timeline animations:

[0137] Extract explicit / implicit time adverbials in sentences;

[0138] The bidirectional time pointer is superimposed at the wrong verb;

[0139] The red pointer points to the correct tense time node;

[0140] Article errors show countable nouns visual model:

[0141] Generate isolated object 3D outlines of singular nouns;

[0142] Generate a dot plot of the group distribution of plural nouns;

[0143] Timeline animation generation also includes:

[0144] For sentences without explicit time adverbials:

[0145] Extract time clue words from context paragraphs;

[0146] Build a logical sequence chain of events;

[0147] Inferring implicit temporal reference points based on chains of events;

[0148] Beautiful word and sentence markers include:

[0149] Detect students' hand-drawn underlines / wavy lines:

[0150] The hand-drawn line was determined by the threshold of curvature variance > 0.8;

[0151] Initiate a deep rhetorical analysis of the sentence covered with hand-drawn lines:

[0152] Identify the subject-object mapping relationship in metaphor structure;

[0153] Detect the parallel structure integrity of parallel sentences;

[0154] Overlay the rhetoric type icon at the corresponding position of the original image;

[0155] This embodiment also includes personalized error pattern analysis:

[0156] Record repeated incorrect root words by the same student;

[0157] When a specific error type occurs 3 or more times cumulatively:

[0158] Analyze the linguistic characteristics of incorrect word roots;

[0159] Generate special training questions containing contrasting language cases;

[0160] During the personalized error pattern analysis, when the same student triggers the same root error multiple times in the correction record, the error type The propagation probability of is updated in three cycles in sequence, including:

[0161] Assume that the completed job number is , at a length of Record binary sequence in the sliding window , get the recent frequency:

[0162] ,

[0163] in, is the current job number, is the window length, is the relative position within the window, For serial number Error indication, is the position weight, is the decay constant, is the recent frequency;

[0164] Error type The propagation probability uses exponential recursion:

[0165] ,

[0166] in, and Sequential numbers 、 The probability of propagation, is the memory weight factor, is the smooth slope, For Error Cumulative number of occurrences, is the trigger threshold;

[0167] Calculate the upper bound of the error and decide whether to generate special training questions:

[0168] ,

[0169] in, is the normal 95% quantile constant, is the sample size in the window, is the half-width of the confidence interval, is the upper confidence limit, if , immediately push the root Special training tasks as the core;

[0170] Specifically, the propagation model captures the error density of students in their most recent assignments through a sliding window and suppresses long-term noise with an exponential weight. As the Sigmoid function increases with the cumulative number of times, it can remain sensitive in the early stages of errors and delay saturation in the frequent error stages. The confidence interval correction gives a conservative upper limit for sparse sample cases to avoid false triggering of training due to occasional errors.

[0171] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A homework grading method based on image recognition, characterized in that: include, Step S1: collecting the original image of the student's handwritten homework and identifying the text area and writing trajectory; Step S2, extracting stroke topological features for the connected writing area; Step S3: Generate a dynamic semantic weight table based on the stroke topology features and perform semantic compensation in combination with the context word vector; Step S4, performing rhetorical structure analysis on the compensated text, marking the starting and ending positions and types of metaphors and parallelisms; Step S5, generating visual annotations, including error explanation animations and beautiful word and sentence markings; The implementation of semantic compensation in step S3 includes: Construct a dynamic word vector weight table containing a set of adjacent high-frequency words of the target vocabulary; When the confidence level of a single word is lower than the threshold, weight adjustment is performed: Get the three words before and after the current word as the semantic context window; Calculate the semantic relevance between the candidate word and the context window; In step S3, during the semantic relevance weighted compensation process, when the OCR confidence level for a word is lower than the threshold, dynamic compensation is performed on the recognition result and fed back to the weight table, including: Current location Both sides The word vectors are aggregated using exponential decay: ， ， in, is the context window radius, is the relative position index, For location word vectors, is the indicator function, is the position weight coefficient, is the decay constant, is the context vector, To sum the intermediate variables; Candidate words Calculate cosine similarity with context: ， in is the candidate word vector, is the semantic relevance; Obtain candidate words at time by exponential smoothing Weight: ， in, and are the old and latest dynamic weights respectively, is the memory decay factor, is the frequency of candidate words, is the frequency smoothing constant, is the current update step; The stroke topology features enhance the weight table in a multiplicative manner: ， in, and To enhance the front and back weight matrices, For Hadamard, is the topological enhancement coefficient, For each The correlation vector composed of Then the dynamic weights are linearly fused with the original confidence and compressed by Sigmoid: ， in, Confidence after compensation, is the original OCR confidence, is the fusion coefficient, is the Sigmoid function, with the maximum The word form replaces the original recognition result.

2. The method for grading homework based on image recognition according to claim 1, characterized in that: The stroke topology feature extraction in step S2 includes: Break down connected letters into discrete stroke segments; Calculate the curvature change sequence of each stroke segment and mark the curvature mutation points; Count the number of closed loops formed by the stroke segments; When the number of closed loops deviates from the standard glyph by more than the tolerance value, the adjacent stroke space vector analysis is activated: Establishing a spatial position coordinate system for stroke segments; Calculate the angle and distance vector between adjacent stroke segments.

3. The method for grading homework based on image recognition according to claim 1, characterized in that: The processing of the cross-line rhetorical structure in step S4 includes: Locate the first line of rhetorical keywords: Identify the page coordinates of the preset rhetorical keywords in the first line of the sentence and record them as the starting coordinate point ; Scan the associated keywords in the next row: Detect semantically associated continued keywords in the preset area at the top of the next row. The preset area satisfies: Horizontal deviation constraint: ; Vertical extent constraints: ; in, is the vertical coordinate of the top of the next row, is the page height; Perform syntax tree merging: When a continuation keyword coordinate that satisfies the constraints is detected When the first line and the second line of text are combined into a single syntax tree for rhetorical structure analysis.

4. The method for grading homework based on image recognition according to claim 1, characterized in that: The error explanation animation generation in step S5 includes: Temporal errors correspond to timeline animations: Extract explicit / implicit time adverbials in sentences; The bidirectional time pointer is superimposed at the wrong verb; The red pointer points to the correct tense time node; Article errors show countable nouns visual model: Generate isolated object 3D outlines of singular nouns; Generate a dot plot of the population distribution of plural nouns.

5. The method for grading homework based on image recognition according to claim 1, characterized in that: The beautiful phrase markers include: Detect students' hand-drawn underlines / wavy lines: The hand-drawn line was determined by the threshold of curvature variance > 0.8; Initiate a deep rhetorical analysis of the sentence covered with hand-drawn lines: Identify the subject-object mapping relationship in metaphor structure; Detect the parallel structure integrity of parallel sentences; A rhetorical type icon is superimposed at the corresponding position of the original image.

6. The method for grading homework based on image recognition according to claim 1, characterized in that: Also includes personalized error pattern analysis: Record repeated incorrect root words by the same student; When a specific error type occurs 3 or more times cumulatively: Analyze the linguistic characteristics of incorrect word roots; Generate specialized training questions containing contrasting language examples.

7. The method for grading homework based on image recognition according to claim 6, characterized in that: In the personalized error pattern analysis process, when the same student triggers the same root error multiple times in the correction record, the error type is analyzed. The propagation probability of is updated in three cycles in sequence, including: Assume that the completed job number is , at a length of Record binary sequence in the sliding window , get the recent frequency: ， in, is the current job number, is the window length, is the relative position within the window, For serial number Error indication, is the position weight, is the decay constant, is the recent frequency; For error types The propagation probability uses exponential recursion: ， in, and Sequential numbers 、 The probability of propagation, is the memory weight factor, is the smooth slope, For Error Cumulative number of occurrences, is the trigger threshold; Calculate the upper bound of the error and decide whether to generate special training questions: ， in, is the normal 95% quantile constant, is the sample size in the window, is the half-width of the confidence interval, is the upper confidence limit, if , immediately push the root Special training tasks with the core.

8. The method for grading homework based on image recognition according to claim 4, characterized in that: The timeline animation generation further includes: For sentences without explicit time adverbials: Extract time clue words from context paragraphs; Build a logical sequence chain of events; Inferring implicit temporal reference points based on event chains.

Images