Recognition of Handwritten Words in School Essays Using Conditional Random Fields

本研究关注儿童在阅读理解测试中手写回答的单词识别。单词识别方法将基于条件随机场（CRFs），这是一种判别方法，不对基础数据进行任何假设，因此已知在序列标记问题上优于隐马尔可夫模式（hmm）。首先使用现有的基于神经网络的算法将学生的反应分割成单词图像。然后将每个词图像过度分割成许多小的片段，这样这些片段的组合就形成了字符图像。片段被标记为具有从CRF模型评估的概率的字符。使用动态规划算法计算代表词典条目的单词图像的总概率，该算法评估片段的最佳组合。使用从阅读文章、测试提示、答案标题和学生回答中衍生出来的词汇来限制探索路径的数量。从手写样本中估计CRF模型的状态和过渡参数。状态参数对应的特征有：位置（按长度归一化）、位置（在起点、中间或终点）、高度、宽度、到原型的距离、高度偏差等。转换参数对应的特征包括：字符对的标签（th, er， qu等），垂直重叠（候选字符图像的像素），高度差，宽度差，长宽比差，重字宽度等。研究测试平台将包括纽约布法罗一所内城学校8年级和5年级对阅读理解提示的打分手写回答。有300个8级的回答和200个5级的回答，每个回答大约100-150个单词。参数估计的训练数据最初将包括150个学生的回答和1000个半页的成人写作。随着研究的进展，这些数据将得到额外的学校数据的补充。复杂文档图像分析、自然语言处理和机器学习的目标导向集成将推动手写识别方法的改进。孩子吗?在文件分析中，手写识别从未被研究过。用于复杂文档的手写识别技术在很大程度上还没有出现。如果成功，全州范围内的考试可以在学年的晚些时候进行，结果可以更快地提供，从而对改善教育产生影响。

This research concerns recognition of words in handwritten responses of children in reading comprehension tests. The approach to word recognition will be based on conditional random fields (CRFs), which are discriminative methods that do not make any assumptions about the underlying data and hence are known to be superior to Hidden Markov Modes (HMMs) for sequence labeling problems.The student response is first segmented into word images using an existing neural network based algorithm. Each word image is then over- segmented into a number of small segments such that the combination of segments forms character images. Segments are labeled as characters with probability evaluated from the CRF model. The total probability of a word image representing an entry from the lexicon is computed using a dynamic programming algorithm which evaluates the optimal combination of segments. A lexicon derived from the reading passage, testing prompt, answer rubric and student responses is used to limit the number of paths to explore.The state and transition parameters of the CRF model are estimated from handwriting samples. State parameters correspond to features such as: position (normalized by length), place (in start, middle or end), height, width, distances to prototype, deviations of height, etc. Transition parameters correspond to features such as: label of character pair (th, er, qu, etc), vertical overlap (of pixels of candidate character images), height difference, width difference, aspect ratio difference, bigram width, etc.The research test-bed will consist of scored handwritten responses to reading comprehension prompts from Grades 8 and 5 of an inner city school in Buffalo, New York. There are 300 Grade 8 responses and 200 Grade 5 responses, with about 100-150 words in each response. Training data for parameter estimation will initially consist of 150 student responses and 1,000 half-page writings of adults. These will be supplemented with additional school data as research progresses.Goal-oriented integration of complex document image analysis, natural language processing and machine learning will drive improved handwriting recognition methods. Children?s handwriting recognition has never before been studied in document analysis. Handwriting recognition technology for complex documents is as yet largely unavailable. Success will allow statewide testing to be done later in the school year with results provided sooner thereby having an impact on improved education.