Automatic Methods for Detecting Contradictions, Claims, and Misinformation Sample Clauses

Automatic Methods for Detecting Contradictions, Claims, and Misinformation. ‌ STATE OF THE ART: The task of automatic Contradiction Detection (CD) is relatively unstudied in language processing, in comparison to other tasks such as information and sentiment extraction. It has been formulated primarily as a kind of textual entailment, where a pair of statements asserts contradictory information (Xxxxxxxx, 2008). (Xxxxxxxx et al, 2008) have established a typology of contradictions, which distinguishes easier to detect, overt contradictions (antonymy, negation, and data/number mismatch) from harder-to-detect, implied contradictions arising from factive or modal words, structural and lexical contrasts, and world knowledge. Automatic contradiction detection approaches have focused mostly on detecting contradictions based on negation, antonymy, and numeric mismatches, e.g. (Xxxxxxxx et al, 2008; Xxxxxxxxx et al, 2006; Xxxxxxxx et al, 2010), using supervised machine learning methods. Linguistic features are typically derived from named entity types, syntactic and semantic parsing, temporal information, antonyms, events, paraphrases, and string overlap. Topic similarity and event coreference also play an important role, since genuine contradictions only arise in cases where two texts discuss the same event or entity. (Xxxxxx et al, 2008) studied the harder task of detecting contradictions where world knowledge is required (e.g. the assertion that Xxxxxx is born in Austria does not contradict that he is born in Salzburg). The method uses WordNet as a source of synonymy and meronymy knowledge, but found problems with its low coverage. Other relevant knowledge is hypernyms (renal failure is a kind of kidney disease), so there is no contradiction between documents using one term vs. the other. Xxxxxx et al also established that in most cases seemingly contradictory statements are actually entailments, if world knowledge is present (34% of errors are due to missing knowledge and 49% of errors are due to location ambiguity). The problem is that current CD methods do not draw on sufficiently large world knowledge sources, which leads to low accuracy results. Controversial information may appear in both “traditional” and social media (e.g. whether aluminium causes Alzheimer’s disease). Fact-checking websites (e.g. xxxxxxxxx.xxx and xxxxxxxxxx.xxx) are useful for checking known disputed claims manually. However, the automatic detection of such disputed information has not been studied extensively in NLP. A first step in that direction was made in ...