Automated essay scoring (AES) refers to the process of grading student essays without human interference. An AES system takes as input an essay written for a given prompt, and then assigns a numeric score to the essay reflecting its quality, based on its content, grammar, and organization, etc. Most AES systems are based on a supervised machine learning task (mostly by regression methods or preference ranking) applied to a set of carefully designed features or deep neural network systems which can effectively encode the information required for essay evaluation and learn the complex patterns in the data through non-linear neural layers.

Recent research has found that deep neural networks are vulnerable to adversarial samples and the existence has been shown in domains like image classification and speech recognition. State-of-the-art NLP models can often be fooled by adversaries that apply seemingly innocuous transformations to the input. Question Answering(QA) scoring methods in the form of GRE and TOEFL should especially be resistant to these adversaries since they are part of the student assessment process and responsible for deciding the future of students.

The Essay Scoring systems are formed according to the preconceived bias of responses, so generally when testing everything seems well to do because of the testing on this bias by the researchers. However we took a step against the conformation bias to see how these state-of-the-art models actually react!

The model is based on handcrafted features and regression methods. The features that are extracted can be categorized into four classes namely, Length-based features, Parts-of- Speech (POS), Word overlap with the prompt, Bag of n-grams . After extracting these features, a regression algorithm is used to build a model based on the training data. We have used support vector regression (SVR) as our regression algorithm.Support Vector Regression is a a type of support vector machine that supports both linear and non-linear regression. The problem with regression is to a function that approximates mapping from an input domain to real numbers on the basis of a training sample. The main aim here is to decide a decision boundary at ‘a’ distance from the original hyperplane such that data points closest to the hyperplane or the support vectors are within that boundary line.This model uses NLTK for POS tagging and stemming, aspell or pyspellchecker for spellchecking, and WordNet to extract the synonyms. Correct POS tags are generated using a grammatically correct text which were provided or one can also generate using a pretrained Grammatical Error Correction model. The POS tag sequences not included in the correct POS tags are considered as bad POS. We use python library, scikit-learn for extracting unigram and bigram features and also for implementing the regression based methods.

CountVectorizer is used for tokenizing a collection of text documents and building a vocabulary of known words, and also to encode new documents using that vocabulary. They use CountVectorizer to build two dictionaries, one normal text (unstemmed words and bigrams) and another using stemmed and spell corrected vocab using useful words/ngrams. One can also use TF-IDFVectorizer which takes care of Term Frequency (how often a given word appears in a document) and Inverse Document Frequency (down-scales words that appear a lot across documents) or also HashingVectorizer which is a one-way hash of words to convert them to integer without creating a vocabulary (efficient with large documents and vocabularies). Length-based features are extracted from the essays. This includes, length of the essay, word count, comma count, apostrophe count, punctuation count, characters per word, As listed in the above Table, prompt based features are also generated from the essay by taking an overlap with the question prompt by finding the number of words or synonyms of words from the essay that exist.

This model, takes an essay directly as an input and automatically learns the features from the essay. For this process, we implemented a recurrent neural network based architecture to automatically grade the essays in an end-to-end manner. Recurrent Neural Networks have been one of the most successful machine learning models to solve several tasks especially in NLP domain. They are considered to be more powerful that feed-forward neural networks as they are capable of learning complex patterns from data and sequences.

The RNN structure ideally should encode all information or features required to grade essays. However, since the essays are usually long, with around hundreds of words, the learnt vector representation might not be sufficient for accurate scoring. For this reason, they preserve all the intermediate states of the recurrent layer to keep track of the important bits of information from processing the essay. They continued experiments on long short-term memory units (LSTM) to identify the best choice for this task.

To enhance the coherence and relatedness that can be inferred by a model , two features are kept in mind:

1. To alleviate the inability of the current neural network architecture to model the flow of the essay, coherence of the text and semantic relatedness over the course of the essay.

2. To ease the burden of the recurrent neural network based model.

In order to capture the semantic relationship between points of an essay, the model tries to read the essay using a neural tensor layer. Multiple features of semantic relatedness are aggregated together across the the essay and are used as features for prediction. The semantic relationship between those multiple points is important because it is an indicator of writing flow and textual coherence. The auxiliary features aim to capture the logical and semantic flow of the essay. Secondly, these additional parameters from the external tensor serve as an auxiliary memory of the network which helps in improving the performance of the deep architecture. Overall, the proposed architecture performs sentence modelling and semantic matching in an end-to-end network.

1. Embedding Layer:

The model accepts an essay and the target score as a training instance. The embedding layer is used to represent each essay as a fixed-length sequence,padding all sequences to the maximum length. This results in word embeddings. The parameters of the embedding layer are defined as W e ∈R |V| belongs to N where |V| is the size of the vocabulary and N is the dimensionality of the word embeddings. We have used pretrained Glove embeddings in our model, instead of training an embedding layer.

2. The Long Short-Term Memory (LSTM):

The sequence of word embeddings obtained from the embedding layer is then passed into a LSTM network. At every time step t, LSTM outputs a hidden vector h_t = LSTM(h t_1 , x_i ) (where x i and h_t_1 are the input vectors at time t) that reflects the semantic representation of the essay at position t. To select the final representation of the essay, a temporal mean pool is applied to all LSTM outputs.

3.Neural Tensor Layer:

A tensor layer is used to model the relationship between two LSTM outputs.

4. Fully-Connected Hidden Layer:

All the scalar values that are obtained from the tensor layer are concatenated together to form the neural coherence feature vector. The essay representation which obtained from a mean pooling over all hidden states is then concatenated with the coherence feature vector. This is then given as an input to the fully connected hidden layer.

5. Linear Layer with sigmoid:

This is the final linear regression layer. The output at this layer is the normalised score of the essay.

They develop a Two Stage model mostly based on which makes full use of the advantages of feature-engineered and end-to-end methods which are both vital for the functioning of an AES System. In the first stage, they calculate three scores based on Long Short-Term Memory (LSTM) neural network as follows:

1. Semantic score (Se ): Prompt-Independent and utilized to evaluate essays from deep semantic level.

2. Coherence score (Ce ): Exploited to detect the essays composed of permuted paragraphs.

3. Prompt-relevant score (Pe ): Word Overlap with prompt is considered to discover prompt-irrelevant samples.

In the second stage, they concatenate these three scores with some handcrafted features, and the results are fed through an eXtreme Gradient Boosting model (XGboost) machine learning model for training.

First Stage:

– Semantic Score: LSTM used to map essays into low-dimensional embeddings,which are then fed to a dense output layer to transform the vector into a scalar value using Sigmoid function. They use the last hidden state rather than the average hidden state as it does not perform well comparatively. The objective they choose is the Mean Squared Error (MSE) loss function.

– Coherence Score: Calculated based on LSTM and objective of our coherence model is also built using the MSE. During the training process, they assume that the gold coherence scores of the original essays are equal to the corresponding hand marked scores. For those adversarial samples with some permutation, the gold coherence scores are set as 0.

– Prompt-relevant Score: This is also extracted using the LSTM using the last hidden state hidden state which is fed to a non-linear followed by sigmoid function and objective function based on MSE. Note: for prompt-irrelevant essays, the gold prompt relevant scores are defined as 0. Gold Scores for the training set are normalized to the range [0, 1] which are rescaled during the testing process. All LSTM neural networks are single-layer with hidden size of 1024. Mulilayer and bidirectional LSTMs were also tested for performance. Adam Optimizer is used to update training variables, using hyperparamters β 1 = 0.9, β 2 = 0.999 and = 0.000001.

Second Stage:

Some length-based handcrafted features are extracted. They first see the results using these features and then use features extracted in Feature Extraction using Regression method implemented in previous chapter and compare the results. They concatenate semantic score, coherence score and prompt-relevant score extracted in the first stage with handcrafted features and pass them through XGBoost machine learning model which uses Gradient Boosting Decision Tree (GBDT) and consists of two basic models including the tree model and linear model. Learning rate is as 0.001 and the max depth of tree is equal to 6.

This model consists of four layers: input representation layer, memory addressing layer, memory reading layer, and output layer. Input representation layer is responsible for generating a vector representation for a student response. Memory addressing layer loads selected samples of student responses to memory, and assigns a weight to each memory piece. Afterward memory reading layer gathers the content from memory by taking weighted sum of each memory piece based on the weights calculated from previous layer, and produces a resulting state. Finally the output layer makes the prediction on the basis of the resulting state. Neural networks models are usually featured with multiple computational layers to learn a more abstract representation of the input.This model is extended to have the structure of multiple layers (hops) by stacking memory addressing layer and memory reading layer repeatedly.

Finally, we tried training on the adversarial samples generated by our framework to see if the models are able to pick up some inherent “pattern” from the adversarial samples. The detail of this experimentation can be read from our manuscript. However, the adversarial training did not impact the scoring much. A slight improvement was noticed, but it was not enough to call this experimentation successful.

Through our experiments. we conclude that recent AES systems built mainly with feature extraction techniques and deep neural networks based algorithms fail to recognise the presence of common-sense adversaries in student essays and responses. As these common adversaries are popular among students for ‘bluffing’ during examinations, it is vital for Automated Scoring system developers to think beyond accuracies of their systems and pay attention to complete robustness so that these systems are not vulnerable to any form of adversarial attack.