In the inclusion, automated QA might become an effective tool allowing, for example, to ask questions about the interaction between neurotypical and atypical people anonymously and get reliable information immediately. However, the controllability of such systems is challenging. Before the integration of QA in the inclusion, a research is required to prevent the generation of misleading and false answers, and verify that a system is safe and does not misrepresent or alter the information. Although the problem of data misrepresentation is not new, the approach presented in the paper is novel, because it highlights a particular NLP application in the field of social policy and healthcare. The study focuses on extractive and generative QA models based on BERT and GPT-2 pre-trained Transformers, fine-tuned on a Russian dataset for the inclusion of people with autism spectrum disorder.
AI-powered question answering systems might find their practical application in the medical and social domain. Question answering (QA) systems take questions in natural language as input and provide (for example, by text generation or data extraction) corresponding answers as outputs. In the healthcare field, automated QA might benefit both patients and medical practitioners by providing immediate access to required extracts from medical knowledge bases. Closed-domain QA can be used as an additional source of information for volunteers or members of a social institution by providing immediate access to the internal information of a certain organization. Based on a rich and reliable database, QA systems can be used as an additional educational source in the processes of gamification and digitalization at schools or higher education institutions.
The idea of the paper came after the first trial of building an informational question answering system. The system aims to give information about inclusive education in the Russian language. The project supports the inclusion of people with autism spectrum disorder (ASD). In the inclusion, automated QA might become an efficient tool. Limited knowledge of the inclusive education process and lack of awareness about the people with special needs raise anxiety among both neurologically typical members of the inclusion and members with developmental characteristics. The information awareness would help to dispel misconceptions and prevent conflicts in classes.
AI-powered QA is a way to provide information fast and playfully. Children and young adults are not likely to read and analyze extensive texts to find the needed information. The ability to ask any question in a free form would not require a high concentration and save a lot of time, making the inclusion more comfortable. Moreover, members of the inclusion would have an opportunity to ask frequent and uncomfortable questions anonymously. For example, if a student needs a tip for communication with a classmate with ASD and is too shy to ask a friend or teacher, or there is no teacher or tutor around, the student will have a chance to ask a QA bot and get reliable information immediately.
However, the integration of QA systems into inclusive organizations requires confidence that the built applications are safe. Safe applications involve language models that do not generate false information or mislead. Such models should be bias-resistant. They should interact with a user in a friendly way generating coherent and understandable texts, although they should not entertain a user.
One of the challenges of neural approaches towards natural language processing is their
controllability. High scores of perplexity imply coherent text generation but do not exclude the
generation of misleading or false responses. Thus, the outputs of uncontrollable models might be
generic or factually incorrect, whereas, for neural conversation models, semantic control ensuring is
essential [
The paper aims to highlight the linguistic features of question answering systems’ responses and analyze their strengths and weaknesses from the users’ perspective. The study will lead to a broader understanding of the capabilities of the practical efficiency of AI-powered QA. The research focuses on the underlying causes of dialogue system errors and will contribute to the further development of conversational AI.
As a research method, it was chosen to build two question answering systems using two different
approaches. The first approach is extractive. This approach is widespread in the reading
comprehension task, one of the problems of natural language understanding (NLU) [
For the implementation of two approaches, self-attention Transformer network architecture models
were applied. The generative approach was implemented with the Transformer decoder based model
GPT-2 [
One of the issues of modern NLP is that most of the models are evaluated on the English data. However, the English language is rather weakly inflected. That is not typical for most of the IndoEuropean languages. Thus, high model evaluation scores might be reached without taking into consideration the facts about linguistic features of other languages. The Russian language, for example, is fusional. That means that the morphological features are crucial for the understanding of the meaning of a sentence. Spans, which represent the answers in extractive QA, are direct citations of the text. Thus, if the wording of the question is not equal to the wording of the context, the rules of conjugation and declension might be broken.
Although the problem of data misrepresentation is not new, the approach presented in the paper is novel, because it highlights a particular NLP application in the field of social policy and healthcare. The development of two QA models and their analysis presented in the paper should shed light on the problems of building social-oriented conversational AI systems. That might help to predict possible issues and solve them before they happen.
The study focuses on building a conversational AI (ConvAI) system. According to Gao et al. [
The tasks of ConvAI vary, although there are common fundamental tasks that form the basis of the
research field. One of the foundational problems of conversational AI is task completion. While
solving this type of problem, the dialogue agent should be capable of recognizing the user’s needs.
After the task recognition, the agent should be able to accomplish it and give an appropriate response
in the natural language if necessary. The range of tasks varies from the restaurant and hotel
reservations to the meeting scheduling and business planning [
Another foundational task is social chat. Social chatbots are designed for human-AI
communication, which imitates everyday human interaction. The development of such systems may
have the goal of modeling human conversations to pass the Turing test [
The current study focuses on question answering systems. Question answering is another
foundational ConvAI task [
The spectrum of QA-world represents such systems as Knowledge-Based QA agents, or KB-QA,
text-QA, and Machine Reading Comprehension (MRC) models. Question answering systems that use
natural language as a part of their interface are more convenient to use than similar systems not based
on NLP algorithms. For example, KB-QA agents are often compared to SQL-like systems. KB-QA
are considered to be more user-friendly than their predecessors due to their interactiveness [
Neural MRC is another important QA related model. The task of MRC is to generate an answer to
a user’s question posed on a given text. The task aims to evaluate the machine capability of natural
language understanding. Theoretically, the ability of a machine to make some conclusions after the
reading, for example, to answer text-related questions might lead to a breakthrough in human-AI
interaction. MRC might have a broader practical application. For example, MRC algorithms can be
integrated into search engines allowing them to give short answers to a user’s query instead of
providing an unstructured list of possible web-pages with relevant information [
One of the examples of reading comprehension datasets is Stanford Question Answering Dataset
(SQuAD) [
SQuAD v2.0 [
The models built for the experiments were trained on a custom question answering dataset. The dataset was collected by the author of the paper. It is available online (see Online Resources). The dataset is called Autism Spectrum Disorder Question Answering (ASD QA). ASD QA is based on the data from the informational websites about autism spectrum disorder and Asperger syndrome in children and adults, inclusion and support of people with Asperger syndrome and ASD, their health, and communication with neurologically typical people. ASD QA is a long-term project. For the year 2021, it has the status of active, which means that the dataset is in the process of collection and development.
The data for the ASD QA was collected from the informational website about ASD and Asperger syndrome http://aspergers.ru/ with the agreement of the website administration. The data from the website represent a collection of articles and texts of related genres (blog entries, messages to readers, etc.). The texts were created by neurologically typical people and people with Asperger’s syndrome or ASD, created in Russian or translated into Russian from foreign languages. The authors are native or fluent speakers of the Russian language.
According to the website categories, the publications from the informational source cover the following topics: basic information about Asperger’s syndrome and ASD, diagnostics of Asperger’s syndrome and ASD, symptomatic of Asperger’s syndrome and ASD, problems of people with Asperger’s syndrome and ASD, social skills and communication issues of people Asperger’s syndrome and ASD, recommendations for parents of children with Asperger’s syndrome and ASD, education, and training, work and employment, relationships, love and family, discussions about ASD, myths and facts about ASD, etc.
Figure 1 presents a topical data distribution in the ASD QA dataset as at May 2021. The topics were extracted from the website http://aspergers.ru/ which served as a source for the ASD QA dataset. Each article on the website has one or several tags indicating its topics. After we had extracted those tags we built a bar chart showing the number of articles covering each topic. One article could cover several topics.
The data was collected with an HTML parser built with Beautiful Soup 4 [
After the data was collected, it was important to structurize it. Insofar as the dataset was being designed for the question answering models training and evaluation, it was decided to develop it like a reading comprehension one. In contrast with traditional question answering datasets, which contain only sets of QA-pairs, the format of reading comprehension datasets also implies the presence of reading passages. Reading passages are sets of sentences or paragraphs, which an MRC model should learn to “understand” or answer the questions about the information contained in each passage.
Another important aspect is the question acquisition. The reading passages were split into sentences separated by periods, ellipses, question or exclamation marks. We strove to ask one or several questions to each sentence, but some of the text pieces (for example, some introductory remarks or personal reflections) did not contain significant information, so we had to ignore them. We have asked 2-3 questions on average to each sentence containing significant information, using different types of questions. We have chosen the type of a question based on the structure of its possible answer (excerpt from a reading passage). For example, we have asked closed questions to sentences containing affirmative or negative constructions, and we have asked open questions to sentences containing factual information. This was done manually because the ASD QA dataset is being designed for ”safety-first” systems which require the best available training data.
Figure 2 presents an ASD QA dataset sample. The dataset structure was inspired by SQuAD v2.0
[
During the ASD QA development, the dataset was provided with 5% of unanswerable questions on the principle of SQuAD v2.0. Unanswerable questions in the ASD QA dataset are deliberately irrelevant, which means that there are no answers to these questions in the reading passages, and also there are no answers in the dataset at all. Among such questions, there are ones that aim to set an entertaining tone in a human-AI dialogue. For example, some questions ask a system to tell a joke or a fairy tale, some are about artificial intelligence misconceptions, some contain complaints about boredom, etc. Presumably, users can ask such questions for entertainment purposes. However, the systems, for training and evaluation of which the ASD QA dataset is developed, should avoid such questions. These systems aim to consult and give accurate information. They do not have an aim to entertain a user.
For the unanswerable questions, the system includes a label ”is unanswerable”. The JSON object containing the ASD QA data includes the label with a Boolean for each QA-pair. Thus, if a question has a piece of information in a corresponding reading passage, the label ”is unanswerable” is False. Otherwise, the label is True. For example, in Figure 3 two QA-pairs are presented. The question of the first pair is translated from Russian into English as “Is autism a deviation?”. This question has an answer in a corresponding reading passage, which is marked as a “context” in the dataset. The label “is unanswerable” is False. Labels “answer start” and “answer end” mark the answer span, serial numbers of the first and last characters position of answers in the passage.
The question of the second pair is translated from Russian into English as “Tell me the news?”. This question has no answer in the dataset reading passages, it is added in the dataset to complicate the task. The label “is unanswerable” for this question is True. The values of “answer start” and “answer end” are both 0. Despite the fact that the question is unanswerable and irrelevant, the dataset is provided with a plausible answer, which is translated from Russian into English as “I cannot answer this question”. This makes the dataset also suitable for the training of generative QA models. Such models instead of answering irrelevant questions can learn to generate this phrase.
Table 1 presents the ASD QA data statistics in the context of the paper research. For the
implementation of the experiments, the dataset was split with the “train_test_split” method from the
Scikit-learn library [
The extractive approach, which is closely related to machine reading comprehension (MRC), was
implemented using pre-trained Transformer Bidirectional Encoder Representations from
Transformers (BERT) [
In the current research, the generative approach was implemented with a Generative Pre-trained
Transformer (GPT-2) [
Among GPT-2 zero-shot learning achievements are solving question answering tasks and MRC, summarization, and translation without fine-tuning, and others. All this is achieved only by pretraining the model for traditional language modeling. Figure 4 represents the concept of traditional language modeling. Unidirectional arrows in Figure 4 show unidirectional GPT-2 processing. The question mark illustrates the model task to complete a given sequence. The sequence “Today I will” is input data, or prefix, which a model should continue. The sequence “go to school” is a model output.
Both BERT based and GPT-2 based models were trained with a Russian dataset for the inclusion of people with autism spectrum disorder (ASD) (see Online Resources), although, for the generative model training, some changes were required. The dataset includes a special label indicating whether a question has an answer in a corresponding reading passage. If the value of this label is True, the answer presented in the dataset is special (see Figure 2). It is translated from Russian into English as “I cannot answer this question”. This dataset feature was provided for generative models training, so they could learn to answer irrelevant questions politely.
The original version of the dataset is designed for MRC, so it had to be changed for the generative GPT-2 based model training. Firstly, all the answers and questions were extracted from the original dataset. Pairs of questions and answers, or QA-pairs, were located sequentially, separated by an empty row. All the QA-pairs were randomly shuffled. Secondly, the spans metadata was removed. Thirdly, the reading passages were not removed for the model failsafe. That was intended for cases when a possible answer to a user’s question was contained in reading passages but absent in the training QApairs. Finally, the meta-information on answerable and unanswerable questions was removed, but the answer” I cannot answer this question” was saved for each unanswerable question.
Transfer Learning techniques were used to fine-tune the models for the experiments. Transfer Learning allows using the knowledge gained while solving one general task to solve another similar one. The model is first trained on a large amount of data. Then, the pre-trained model is trained on the target dataset to solve a downstream problem. There are different Transfer Learning techniques. In this study, a fine-tuning strategy is used. The network trains end-to-end on a new custom dataset to adjust and adapt for the downstream task. 5.1.
For the question answering evaluation, F1-Score was used as proposed in [
+
In question answering, true positive answers are the tokens shared between the correct (gold)
tokens and all the predicted tokens. False positives are the predicted tokens absent in the correct
(gold) answers, and false negatives are the tokens from the correct (gold) answer absent in the
predicted ones. With this correction, the formula is the following as presented in the SQuAD
evaluation script [
+ 2 ℎ ℎ = ℎ = − ℎ
) ℎ + ( − ℎ ) (2) (3) (4) (5) (6) 5.2.
The model training was performed in Google Colaboratory with the Tesla T4 GPU. The code was
implemented in Python [
F-Score 0.55 0.63
After the linguistic analysis, the author of the paper defines four criteria of the models’ outputs evaluation. The criteria were determined according to the language levels that the author of the paper found essential for the analysis. The analysis focused on the evaluation of the QA models’ safety for their further integration into inclusive education. The criteria and language levels are the following: syntax level, morphology level, grammar correctness, lexical diversity.
On the syntax level, it was found that the extractive BERT based model can give full, syntactically correct sentences, but only if it copes with a user’s question. If the model cannot correctly recognize a user’s question, it would output a single word or a single letter, the first token from a corresponding context, with a high probability. For example, during the research, prepositions were very frequent in the model outputs. The generative GPT-2 based model, in turn, tends to give complete answers more often. However, its outputs contain frequent syntactic violations. That is inappropriate and is yet to be improved.
On the morphology level, the BERT based model did not show significant violations due to the extraction properties, although it truncated words in cases when it did not cope with a question. The GPT-2 based model could generate new words or word forms, which is worse because it might create unexisting lexical units. Grammar mistakes in the extractive model could only be caused by the presence of unknown words (due to the size of the training vocabulary) in the dataset. In the generative model, grammar mistakes were more frequent.
The extractive model did not make lexical repetitions extracting single answers. That makes the model clear and informative. However, this model cannot generate unique utterances. The generative model, in turn, could generate lexically diverse unique sentences. However, it also could create words that do not exist and repeat lexical constructions.
According to the conclusion of the study, extractive question answering is more reliable than generative question answering. The QA chatbot systems integration into inclusive education requires high alertness to its outputs. Thus, generative systems can be unsafe, as they might turn a tool for the information support or consultations into a toy, which is inappropriate.
Nevertheless, the capabilities of generative systems allow them to generate unique answers without grammar mistakes, lexical repetitions, and syntactic violations while maintaining factual accuracy. That makes them efficient. Although the score and errors point are yet far from optimal solution, the solutions presented in the paper provide future directions for improvement. For example, we can build models based on the extractive approach to extract accurate information containing the answer to the user’s question and use generative algorithms as part of the natural language interface to arrange the answer.