In the age of digital media, information flows faster and wider than ever before, making it increasingly challenging for media professionals to manage the overwhelming amount of data being disclosed. Verifying the accuracy of information is a complex task, and identifying contradictions and discrepancies in real-time is often difficult due to time gaps.

Moreover, with the rise of social media platforms, personal computers, mobile devices, and wearable technologies, we are living in an era where voice, images, sounds, and text can be digitized and transmitted more rapidly than ever before. As a result, media professionals must be equipped with the necessary tools and techniques to monitor and sift through this vast amount of digital information to ensure that they provide accurate and reliable news to their audiences.

The proliferation of digital media has not only made it easier to access and distribute information, but it has also created an environment ripe for misinformation. With the abundance of false and contradictory data, coupled with the changing opinions of certain entities, it's become increasingly difficult to monitor, detect, and verify the accuracy of information we receive.

While a few dedicated professionals work to address this issue, the sheer volume and complexity of data require automation. Natural Language Processing (NLP), the process of interpreting human language, has proven to be an effective tool in automating tasks previously performed by humans, such as fact-checking, entailment, recognition, and contradiction detection.

However, NLP still faces various challenges, such as the sheer volume of data, the structure of language (as it must be machine-interpretable), the meanings of words, and the relationships between them. To address these challenges, many researchers employ word embeddings in NLP, which use Distributional Semantic Models (DSMs) to extract lexical semantics from words, capturing the essence of words and their relationships to better understand their meaning.

What is a Contradiction?

A contradiction occurs when two sentences cannot both be true at the same time. People's interests may change over time, so it is normal to find contradictions in present and past speeches. These belief changes can be strategic, aiming to please, manipulate, or deceive others.  

Political candidates changing their stances on controversial issues is a perfect example of acceptable contradiction. Therefore, collecting samples of confirmed cases of contradictions is challenging, not only because of a lack of resources but also due to their subjective nature.

Word embeddings and machine learning can help resolve this complex task. Considering data limitations (collecting contradictory statements), this dissertation aims to improve the efficiency of automatic contradiction detection by utilizing modern NLP and machine learning techniques, namely word embeddings and transfer learning.

What is Contradiction Detection?

A contradiction is an inconsistency between two or more statements, propositions, or beliefs. In the context of NLP tasks, contradictions refer to incompatibilities between descriptions of the same event, where two or more sentences cannot be true at the same time.

In NLP, texts are often represented as typed dependency graphs that are generated by Stanford parsers. These graphs capture the relationships between words in a sentence and can be used to identify mismatches between different sentence structures.  

To detect contradictions, NLP models often rely on annotating Recognizing Textual Entailment (RTE) datasets that contain examples of contradictory statements. By comparing the graphs generated from different sentences and identifying mismatches, NLP models can identify and flag potential contradictions.

Features Of Contradiction Detection

These seven features can help identify contradiction patterns:

Polarity Features

When there is strong alignment between the text and hypothesis, the presence or absence of linguistic markers associated with negative polarity can often serve as a reliable indicator of contradiction. These features capture the polarity difference between the two statements.

Generally, words are considered negated if they have a negation dependency on the graph or are explicit linguistic markers of negation (e.g., simple negation (not), downward-monotone quantifiers (no, few), or restricting prepositions).

A polarity difference may exist if one word has negative connotations and the other is not. To confirm this, check that the words are not antonyms and do not contain unaligned prepositions or any further context that suggests they refer to different things.  

Number, Date, and Time Features

The numeric features identify mismatches between numbers, dates, and times. Numbers appear as ranges, and the date and time expressions are normalized. It is considered a mismatch when aligned numbers are incompatible, and surrounding words match well, indicating they refer to the same entity.

Antonymy Features

It is elementary to identify a contradiction when the antonyms are aligned. When determining whether antonyms create a contradiction in a context, the polarity of the context plays a role.

Structural Features

In syntactic structures, the subject of one sentence overlaps the object of the next. These features will help to determine whether the syntactic structures of the text and hypothesis contradict one another. For each aligned verb, compare the subjects and objects. There is a contradiction if the issue in the text overlaps with the object in the hypothesis.

Factivity Features

Sometimes, a verb phrase may appear contradictory to the context in which it appears. Some factivity patterns depend on the use of negative words.  

Modality Features

Text and hypothesis follow one of six modalities ((not) possible, (not) actual, (not) necessary) based on the presence of predefined modality markers such as can and maybe. This feature occurs if the text/hypothesis modality pair gives rise to a contradiction. Patterns of modal reasoning (based on the presence of modality markers such as "can" and "maybe").

Relational Features

Relational features provide accurate information, though they are difficult to extend for broad coverage. Almost all RTE data comes from information extraction.

Conclusion

In conclusion, the proliferation of digital media has made it increasingly challenging to verify the accuracy of information, leading to a rise in misinformation. Natural Language Processing (NLP) has proven to be an effective tool in automating tasks such as contradiction detection, which helps media professionals ensure the accuracy and reliability of news. Despite facing challenges such as the sheer volume of data and the structure of language, researchers employ word embeddings in NLP to better understand the meanings of words and their relationships.  

Additionally, seven features, including polarity, number, date, and time, antonymy, structural, factivity, modality, and relational features, can help identify contradiction patterns. By utilizing modern NLP and machine learning techniques, such as word embeddings and transfer learning, this dissertation aims to improve the efficiency of automatic contradiction detection, providing a valuable contribution to the field of NLP and media.

‍