Are Large Language Models Fit For Guided Reading?

LLMs must demonstrate that they have the potential for an in-depth understanding of a given input text for them to be deployed as reading guides. One indicator of input text’s comprehension is the ability to generate meaningful questions and answers from the input text. Formally, given a passage $P$, the model should be able to generate a set of questions $\{q_{1},\cdots,q_{n}\}$ and their respective answers $\{q_{1},\cdots,q_{n}\}$. Both the set of questions and answers should be retrieved from the passage $P$. Perhaps one of the greatest powers of LLMs such as ChatGPT is the ability to respond to questions posed to it on-the-fly. However, it is unclear to what extent they can connect different aspect of input stories to generate both low cognitive questions and questions that require inference to be answered (i.e., high cognitive demand questions). Moreover,how will it exploit the vast amount of knowledge it acquired during training to boost its question asking ability? Further, will it be able to generate answers from the input text without being "confused" by its internal knowledge. We are interested in evaluating the ability of LLMs to ask meaningful questions that can be solely answered from the input story. We are also interested to evaluate how accurate LLMs are in answering the questions when solely relying on its understanding of the input text. To do this, we prompt a given LLM to generate questions based on an input story (see fig 1). The generated questions are automatically evaluated by comparing their semantic similarity to a set of baseline questions.

To evaluate the performance of a given LLM on question generation, we exploit the popularly used metrics in question generation task which include ROUGE-L [24] and BERTScore [25] and compare the semantic similarity of questions generated with the baseline questions. The similarity between LLM generated questions and reference questions is evaluated by concatenating the generated questions into one sentence and compare it with similarly concatenated reference questions ( see [10] which uses a similar approach).

To test the student’s understanding of a given text being read, questions must be generated that cover nearly all the sections of the story. We are interested in evaluating the ability of ChatGPT and Bard to generate questions that are not biased towards a given section of the text. Concretely, we are seeking to quantify the variation of the questions being generated by LLMs. We hypothesize that the more diverse the questions are, the more exhaustive the questions cover the different topics in the input text. This will give us an idea on suitability of LLMs to generate questions that cover the whole content being read. In machine learning, several entropy reliant techniques have been proposed to evaluate diversity of dataset. Research in [26] proposes Inception Score (IS) to evaluate the synthetic samples generated by a Generative Adversarial model(GAN),$(G)$[27]. The intuition behind IS is that the conditional label distribution $P(y\mid x)$ of images generated by GAN is projected to have low label entropy i.e., the generated images belong to few classes while entropy across the images should be high that is the marginal distribution $\int p(y\mid x=G(z))dz$ should have high entropy. To capture this intuition, they suggest the metric in equation 1.

$$exp(\mathbb{E}_{x}{KL}(p(z\mid x)\parallel p(y)))$$

(1)

Another popular metric that provides a measure of diversity on synthetically generated samples is the Frechet Inception Distance (FID) score [28]. This is a metric that considers location and ordering of the data along the function space. Taking the activations of the penultimate layer of a given model to represent features of a given dataset $x$ and considering only the first two moments i.e., the mean and covariance, the FID assumes that the coding units of the model $f(x)$ follow a multi-dimensional Gaussian, therefore have maximum entropy distribution for a given mean and covariance. If the model $f(x)$,generates embeddings with mean and covariance (m,C) given the synthetic data $p(.)$ and $(m_{w},C_{w})$ for real data $p_{w}(.)$, then FID is defined as:

$$d^{2}((m,C),(m_{w},C_{w})=||m-m_{w}||_{2}^{2}+Tr(C+c_{w}-2(CC_{w})^{1/2})$$

(2)

Other metrics that have been proposed to evaluate diversity, include precision and recall metrics [29] [30]. One major problem with these metrics is that they assume existence of reference samples where the generated samples can be compared to [30]. In our case, we seek to evaluate the diversity of the questions generated without comparing to any reference questions. To achieve this, we use Vendi score (VS), a metric proposed for diversity evaluation in the absence of reference data. VS is defined as

$$VS(X)=exp(-\sum_{i}^{s}\lambda_{i}log\lambda_{i})$$

(3)

Here $X=\{x_{i},\cdots,x_{n}\}$ is the input data whose diversity is to be evaluated. $\lambda_{i},i=\{1,\cdots,n\}$ are the eigenvalues of a positive semidefinite matrix $K/n$ whose entries are $K_{ij}=k(x_{i},x_{j})$ where $k$ is a positive semidefinite similarity function with $k(x,x)=1$ for all $x$. This metric which is like effective rank [31] seeks to quantify the geometric transformation induced by performing a linear mapping of a vector $x$ from a vector space $R^{n}$ to $R^{m}$ by a matrix $A$ i.e $Ax$. Normally, the number of dimensions retained by a linear transformation $Ax$ is captured by the rank of the matrix $A$. The rank however is silent on the shape induced by the transformation. Effective rank introduced by [31] seeks to capture the shape that results due to the linear mapping. Effective rank therefore can be used to capture the spread of data hence ideal to measure diversity. To compute the diversity of questions generated by the two LLMs, we execute the following steps:

1. 1.

   Prompt the LLM to generate a set of questions $Q$ given an input text.

2. 2.

   Replicate the set $Q_{1}=\{q_{1},\cdots,q_{n}\}$ to get a copy of the questions $Q_{2}=\{q_{1},\cdots,q_{n}\}$. Designate $Q_{1}$ as the reference set and $Q_{2}$ as the candidate set.

3. 3.

   Pass the set $Q_{1}$ and $Q_{2}$ through the BERTScore ³³3https://github.com/Tiiiger/bert_scoreto extract the cosine similarity matrix $K$.

4. 4.

   Use the VS package⁴⁴4https://github.com/vertaix/Vendi-Score to extract the VS score diversity value.

5. 5.

   Compare the VS score diversity value to human generated diversity score

On top of generating questions that cover the whole content, it is desirable for LLMs to generate wide range of questions from low to high cognitive challenge questions. This will make students to answer questions that address all the cognitive domains. Low cognitive question mostly requires short answers which require affirmation or not (e.g. "Did the queen have golden hair ?"). Conversely high cognitive challenge questions require explanations, evaluations and some speculations on the extension of text [11](e.g., "Why did the king’s advisors fail to find the right wife for him ?"). The purpose of low cognitive challenge questions is to evaluate the basic understanding of the text by the students and ease them into the study interaction [11]. However, they have the potential of promoting over-dominance of teachers . On the other hand, high cognitive questions foster greater engagement of the students, generate inferential responses and promote better comprehension of the reading material [11]. In [11], the two categories of questions are differentiated by exploiting the syntactic structure of the question. The high cognitive challenge questions are signalled by wh-word i.e., questions that use words such as what, why, how, or when. These questions are also composed of a continuum of high to low challenge questions. Specifically, wh-pronoun and wh-determinat questions that start with who, whom, whoever, what, which, whose, whichever and whatever require low challenge literal responses. However, the wh-adverb questions such how, why, where, and when are more challenging since they require more abstract and inferential responses involving explanation of causation and evaluation (e.g., “Why was the king sad?”; “How did the king’s daughter receive the news of her marriage?”). The low cognitive challenge questions are generally non wh-word questions. It has been suggested in [21][32] that teachers should generally seek to ask high challenge questions as opposed to low challenge questions whenever possible. In our case we seek to establish the types of questions preferred by the two LLMs based on their level of challenge. To evaluate this, we adopt three categories of questions i.e., confirmative, explicit and implicit. Explicit are non-confirmative questions that require low challenge literal responses i.e., where answers can be retrieved from the text without much inference while implicit are questions that require inferential responses. To evaluate the type of questions generated, we employ the following steps:

1. 1.

   Given a paragraph of text we prompt the LLM to generate questions based on the text.

2. 2.

   We provide the questions generated to human evaluators and ask them to read the text and answer the questions.

3. 3.

   Human evaluators then annotate each question whether it is confirmative , explicit and implicit.

4. 4.

   For each question, we select the most popular annotation i.e., annotation where most evaluators agree.

5. 5.

   We compute percentages of each category.

Based on the responses to the teacher’s questions, the teacher can detect students’ weaknesses and their demands [33]. While it was difficult to design an evaluation on LLMs that can uncover students’ weaknesses based on their responses, we resorted to evaluate their ability to recommend part of text where the student needs to re-read based on the responses provided to the questions. Basically, we evaluate the ability of a LLM to detect part of the text that the student did not understand. This we believe plays some part in diagnosing student’s need. To perform our evaluation, we execute the following steps:

1. 1.

   We pass a three-paragraph text to a large language model and prompt it to generate questions from the text.

2. 2.

   We annotate the generated questions based on the paragraph in which they were extracted from i.e $<p_{i},q_{i}>$ where $p_{i}$ is the paragraph $i=1,2,3$ and $q_{i}$ are the questions $i=1,\cdots,n$, linked to paragraph $p_{i}$.

3. 3.

   We prompt the large language model to answer all the questions generated. We then deliberately alter a set of answers belonging to questions from a given paragraph to be wrong. All the answers from other two paragraphs remain as generated by the LLM.

4. 4.

   We then prompt LLM to evaluate all the answers( for all the questions) generated in the previous step.

5. 5.

   We prompt LLMs to suggest the section of the text that the student did not comprehend based on the responses in the previous step.

6. 6.

   We compute the BERTScore between the recommended text and the paragraph which had all question altered to be wrong.

We use the FairytaleQA dataset [9] to perform evaluation. The dataset contains 278 annotated storybooks which is composed of 232 storybooks for training a given question generation model, 23 books for testing and 23 for validation. Each book contains multiple paragraphs. Each paragraph of a book is annotated with several educational question-answer pairs. The questions have different types of annotation with our main concern being annotations linked to the difficulty attached to answering the questions. The question difficulty is annotated as either implicit (high challenge questions) or explicit (low challenge questions). The dataset covers storybooks for children between kindergarten and eighth grade.

For question generation, we compare the performance of ChatGPT and Bard with selected models that were specifically developed for questions generations and were evaluated in FairytaleQA dataset. The models include:
QAG [9]: The model relies on semantic role labelling to identify entities and events. The identified entities and events are then used to generate questions. We use the results reported in [10]. The questions used in evaluation include top 10/5/3/2/1 generated questions by QAG, labeled here as (top 10), QAG (top 5), QAG (top 3), QAG (top 1) respectively.
E2E: This model is proposed in [10] where one BART-large model is trained on FairytaleQA dataset to generate questions.
BERT based model: [10]. This model adapts and trains BERT to learn question type distribution. Using the question type distribution as a control signal, they then train BART summarization model using FairytaleQA dataset question-answer annotation to learn the event-centric summary generation task.

Human evaluators used in the study were recruited through post-graduate social media page in Kenya. An advert was posted on the page where 109 Msc. and PhD students responded positively indicating their willingness to participate. Out of these we selected 36 students who are doing research in various topics in NLP. Each student was paid $30 once the annotation task was completed. We conducted one day online training via zoom on the annotation process.