A Road-map Towards Explainable Question Answering A Solution for Information Pollution

TL;DR

This paper advocates for developing explainable question answering systems to combat information pollution by providing transparency, validation, and interpretability of answers, addressing core challenges and conceptual questions.

Contribution

It offers a comprehensive overview of the core concepts, challenges, and key questions in developing explainable question answering systems for better information validation.

Findings

Highlights the need for transparency in QA systems

Identifies key challenges in explanation representation and evaluation

Proposes a conceptual framework for XQA development

Abstract

The increasing rate of information pollution on the Web requires novel solutions to tackle that. Question Answering (QA) interfaces are simplified and user-friendly interfaces to access information on the Web. However, similar to other AI applications, they are black boxes which do not manifest the details of the learning or reasoning steps for augmenting an answer. The Explainable Question Answering (XQA) system can alleviate the pain of information pollution where it provides transparency to the underlying computational model and exposes an interface enabling the end-user to access and validate provenance, validity, context, circulation, interpretation, and feedbacks of information. This position paper sheds light on the core concepts, expectations, and challenges in favor of the following questions (i) What is an XQA system?, (ii) Why do we need XQA?, (iii) When do we need XQA? (iv) How to represent the explanations? (iv) How to evaluate XQA systems?

Full text

Index terms— Explainability, Explainable Question Answering, XQA, Knowledge Graph, Corpus, Evidence, Source Checking, Fact Checking, Reasoning, Context, Provenance, Information Pollution.

1 Introduction

Question Answering (QA) applications are a subcategory of Artificial Intelligence (AI) applications where for a given question, an adequate answer(s) is provided to the end-user regardless of concerns related to the structure and semantics of the underlying data. The spectrum of QA implementations varies from statistical approaches [10, 9], deep learning models [16, 10] to simple rule-based (i.e., template-based) approaches [13, 8]. Also, the underlying data sets in which the answer is exploited might range from Knowledge Graphs (KG) holding a solid semantics as well as structure to unstructured corpora (free text) or consolidation of both. Apart from the implementation details and the background data, roughly speaking, the research community introduced the following categories of QA systems:

• •

Ad-hoc QA: advocates simple and short questions and typically relies on one single KG or Corpus.

• •

Hybrid QA: requires federating knowledge from heterogeneous sources [2].

• •

Complex QA: deals with complex questions which are long, and ambiguous. Typically, to answer such questions, it is required to exploit answers from a hybrid of KGs and textual content [1].

• •

Visualized QA: answers texual questions from images [6].

• •

Pipeline-based QA: provides automatic integration of the state-of-the-art QA implementations [12, 11].

A missing point in all types of QA systems is that in case of either success or failure, they are silent to the question of why? Why have been a particular answer chosen? Why were the rest of the candidates disregarded? Why did the QA system fail to answer? whether it is the fault of the model, quality of data, or lack of data? The truth is that the existing QA systems similar to other AI applications are a black box (see Figure 1) meaning they do not provide any supporting fact (explanation) about the represented answer with respect to the trustworthiness rate to the source of information, the confidence/reliability rate to the chosen answer, and the chain of reasoning or learning steps led to predict the final answer. For example, Figure 1 shows that the user sends the question ‘what is the side effect of antibiotics?’ to the QA system. If the answer is represented in a way similar to the interface of Google, then the end-user might have a mixed feeling as to whether s/he can rely on this answer or how and why such an answer is chosen among numerous candidates?

The rising challenges regarding the credibility, reliability, and validity of the state-of-the-art QA systems are of high importance, especially on critical domains such as life-science involved with human life. The Explainable Question Answering (XQA) systems are an emerging area which tries to address the shortcomings of the existing QA systems. The recent article [17] published a data set containing pairs of question/answer along with the supporting facts of the corpus where an inference mechanism over them led to the answer. Figure 2 is an example taken from the original article [17]. The assumption behind this data set is that the questions require multi-hops to conclude the answer, which is not the case all the time. Besides, this kind of representations might not be an ideal form for XQA; for example, whether representing solely the supporting facts is sufficient? how reliable are the supporting facts? Who published them? And how credible is the publisher? And furthermore, regarding the interface, is not the end-user overwhelmed if s/he wants to go through all the supporting facts? Is not there a more user-friendly approach for representation?

The XQA similar to all applications of Explainable AI (XAI) is expected to be transparent, accountable and fair [7]. If QA is biased (bad QA), it will come up with discriminating information which is biased based on race, gender, age, ethnicity, religion, social or political rank of publisher and targeted user [3]. [4] raises six fundamental competency questions regarding XAI as follows:

Why did the AI system do that? 2. 2.

Why did not the AI system do something else? 3. 3.

When did the AI system succeed? 4. 4.

When did the AI system fail? 5. 5.

When does the AI system give enough confidence in the decision that you can trust? 6. 6.

How can the AI system correct an error?

In the area of XQA, we adopt these questions; however, we apply sufficient modifications as follows:

Why did the QA system choose this answer? 2. 2.

Why did not the QA system answer something else? 3. 3.

When did the QA system succeed? 4. 4.

When did the QA system fail? 5. 5.

When does the QA system give enough confidence in the answer that you can trust? 6. 6.

How can the QA system correct an error?

This visionary paper introduces the core concepts, expectations and challenges in favor of the questions (i) What is an Explainable Question Answering (XQA) system?, (ii) Why do we need XQA?, (iii) When do we need XQA? (iv) How to represent the explanations? (iv) How to evaluate XQA systems? In the following sections, we address each question respectively.

2 What is XQA?

To answer the question of What is XQA?, we feature two layers i.e., model and interface for XQA similar to XAI [4]. Figure 3 shows our envisioned plan for XQA where at the end, the end user confidently conclude that he can/cannot trust to the answer. In the following, we present a formal definition of XQA.

Definition 1** (Explainable Question Answering)**

XQA is a system relying on an explainable computational model for exploiting the answer and then utilizes an explainable interface to represent the answer(s) along with the explanation(s) to the end-user.

This definition highlights two major components of XQA as (i) explainable computational model and (ii) explainable interface. In the following we discuss these two components in more details:

Explainable Computational Model.

Whatever computational model employed in XQA system, (e.g., learning-based model, schema-driven approach, reasoning approach, heuristic approach, rule-based approach, or a mixture of various models) it has to explain all intermediate and final choices meaning the rationale behind the decisions should be transparent, fair, and accountable [7]. The responsible QA system distinguishes misinformation, disinformation, mal-information, and true facts [15]. Furthermore, it cares about the untrustworthiness and trustworthiness of data publisher, information representation, updated or outdated information, accurate or inaccurate information, and also the interpretations that the answer might raise. Whereas, the fair QA system is not biased based on the certain characteristics of the data publisher, or the targeted end user (e.g., region, race, social or political rank). Finally, the transparency of QA systems refers to the availability and accessibility to the reasons behind the decisions of the QA system in each step upon the request of involving individuals (e.g., end user, developer, data publisher, policymakers).

Explainable Interface.

The explainable interface introduced in [4] contains two layers (i) a cognitive layer and (ii) an explanation layer. The cognitive layer represents the implications learned from the computational model in an explainable form (abstractive or summarized representation), and then the explanation layer is responsible for delivering them to the end user in an interactive mode. We introduce several fundamental features which the future generation of XQA have to launch. We extensively elaborate on our view about the interface in Section 5.

3 Why do we need XQA?

We showcase the importance of having XQA using the two following arguments.

Information Disorder Era.

The growth rate of mis-, dis-, mal- information on the Web is getting dramatically worsened [14]. Still, the existing search engines fail to identify misinformation even where it is highly crucial [5]. It is expected from the information retrieval systems (either keyword-based search engines or QA systems) to identify mis-, dis-, mal- information from reliable and trustworthy information.

Human Subject Area.

Having XQA for areas being subjected to lives particularly human subject is highly important. For example, bio-medical and life-science domains require to discriminate between the hypothetical facts, resulting facts, methodological facts, or goal-oriented facts. Thus XQA has to infer the answer of informational question based on the context of the question as to whether it is asking about resulting facts, or hypothetical facts, etc.

4 When do we need XQA?

Typically in the domains that the user wants to make a decision upon the given answer, XQA matters since it enables the end user to make a decision with trust. There are domains that traditional QA does not hurt. For example, if the end user is looking for the ‘nearby Italian restaurant’, QA systems suffice. On the contrary, in the domain of health, having the explanations is demanding otherwise the health care providers can not entirely rely on the answers disposed by the system.

5 How to represent the explanation?

Before, presenting the features of the interface; we illustrate the life cycle of information on the Web (see Figure 4). Each piece of information has a source who initially published it, then After reposting the original piece, it might be framed in a context or manipulated. Also, concerning its circulation on social media or the Web, it might be annotated or commented on by the crowd.

We feature the explainable QA interface with respect to its life cycle in Figure 5. The explainable interface should enable the end user to

• •

access context

• •

find the provenance of information

• •

do fact-checking

• •

do source-checking

• •

check the credibility of the source

• •

detect manipulated information

• •

report mis-, dis-, mal- information

• •

access annotations (feedbacks) of the crowd

• •

see the circulation history

6 How to evaluate XQA systems?

The evaluation of the XQA systems has to check the validity and performance of all the features in the computational model and interface. For example, whether the interface disposes a fact-checking feature or source checking feature? If yes, how fair, transparent, and accountable are these features? The research community requires to introduce metrics, criteria, and benchmarks for evaluating these features.

7 Conclusion

In this paper, we discussed the concepts, expectations, and challenges of XQA. The expectation is that the future generation of QA systems (or search engines) rely on computational explainable models and interact with the end-user via the explainable user interface. The explainable computational models are transparent, fair and accountable. Also, the explainable interfaces enable the end-user to interact with features for source-checking, fact-checking and also accessing to context and circulation history. In addition, the explainable interfaces allow the end-user to report mis-, dis-, mal- information.

We are at the beginning of a long-term agenda to mature this vision and furthermore provide standards and solutions. The phenomena of information pollution is a dark side of the Web which will endanger our society, democracy, justice service and health care. We hope that the XQA will be the attention of the research community in the next couple of years.