Machine Comprehension vs Question Answering: Navigating the Crucial Differences in NLP

As we delve into the realm of Natural Language Processing (NLP), two key concepts often come up: machine comprehension and question answering. While closely related, these two processes are distinct and serve different purposes. Understanding the differences is crucial for developing effective NLP systems. This article explores the nuances between machine comprehension and question answering, providing a comprehensive breakdown of each concept.

What is Machine Comprehension?

Definition: Machine comprehension refers to the ability of a system to read and understand a given text or passage and then answer questions about it. It involves a deeper understanding and reasoning over the content, requiring the system to extract meaning, context, and implications from the text.

Focus: The primary focus is on reading comprehension, where the system must analyze the content to derive inferred meaning and relationships. Common tasks include summarization, inference, and understanding complex relationships within the text.

Examples: Tasks involving machine comprehension might include answering questions that require synthesizing information from multiple sentences or inferring unstated information based on the content. These tasks often necessitate a deeper level of understanding and reasoning.

What is Question Answering?

Definition: Question answering is a broader term that encompasses any system capable of providing answers to questions posed by users. These questions can originate from various sources, including structured databases, text documents, or even conversational contexts.

Focus: The focus can be on retrieving specific information or generating responses based on user queries, making it a more diverse and versatile approach. Question answering may not necessarily require understanding a full passage of text, as long as the system can directly retrieve or generate the required information.

Examples: This can include answering factual questions such as "Tom Cruise's son" or "Fastest man in the world." More complex tasks may require the system to scan through a text or database to find the relevant information.

Differences in Depth of Understanding and Scope

Depth of Understanding: Machine comprehension typically requires a deeper level of understanding of context and implications. It involves analyzing the text in a more nuanced and comprehensive manner to extract the full meaning and reasoning behind the content.

Scope: Question answering encompasses a broader scope, as it includes not only machine comprehension but also other approaches such as information retrieval and direct fact extraction from knowledge bases. Machine comprehension is often a subset of question answering, as it specifically deals with answering questions based on reading comprehension of texts.

In Practice: Integrating Both Capabilities

Many modern NLP systems integrate both machine comprehension and question answering capabilities to provide more accurate and contextually relevant answers. This integration allows for a holistic approach to NLP, combining the advantages of both techniques to handle a wide range of tasks.

Engineering Flavours of Question Answering

Factoid Question Answering: This flavour of question answering involves directly retrieving factual answers from a curated knowledge base or text corpus. For example, using a system that extracts answers from Wikipedia or a knowledge base like Google’s Knowledge Graph. These systems provide pre-engineered facts and are widely used today.

NLP Flavours of Question Answering: This approach leverages NLP techniques to perform reading comprehension, allowing the machine to read and understand a text passage before answering questions. This method aims to avoid the need for pre-engineering and loading facts into a knowledge base, as seen in the Google’s SQuAD and other recent advancements in the field.

Recent Advances in NLP

SOTA Results: Recent advancements in state-of-the-art (SOTA) results have shown promising outcomes in improving reading comprehension. Google's latest attempt, the Squad dataset, exemplifies this progress, showcasing evolving techniques in machine comprehension.

Engineering Flavours: One approach to answering factoid questions is to query a knowledge base or curate facts from sources. For example, the Google Knowledge Graph, originally known as Freebase, contains a vast repository of facts stored in a triple format (subject - predicate - object). This allows for the storage and linking of any fact to form a larger graph, as exemplified by WikiData's BlazeGraph triple store.

For curious readers, understanding the architecture behind these systems can provide insights into how they operate. Reaching out to the architect behind the Google Knowledge Graph, Daniel Kinzler, can offer a deeper understanding of the technology and its applications.

Understanding the differences between machine comprehension and question answering is crucial for developing effective NLP systems. As NLP continues to evolve, the integration of both techniques will likely further enhance the capabilities of modern systems, providing more accurate and contextually relevant answers to users.