Qnarre - Quick Narrative Analyzer
Seeking out contradictions and inconsistencies in deceitful text with ever-adapting, fluidly learning, yet still reproducible systems is exactly what machine learning could have been invented for.
Version 0.2.0
Validate Complex Legal Schemes
Judicial "pruning" of "garden variety fraud" Civil RICO cases does not work: it's time for Congress to act. - DOJ Office Of Justice Programs.
"The latest available data from the federal courts show that civil filings under the Racketeer Influenced and Corrupt Organizations Act (RICO) jumped significantly during FY 2018. As of the end of fiscal year 2018, the government reported a total of 1,405 civil lawsuits filed in federal courts under RICO. In fiscal year 2017, just 693 suits were filed. - TRAC, Syracuse University or trac@syr.edu”
Detect Direct Contradictions
Consistency is simple, contradictions are limitless. Direct textual contradictions should be detected quickly.
Representation"The feature extractor transfers the text input into a feature vector. When using word embeddings, it is possible for words with similar meaning to have a similar representation. Classifiers can then selectively extrapolate the results to sentences and paragraphs.”
Infer Indirect Contradictions
False conjectures lead to indirect contradictions. Weighing the cumulative "credibility" of conjectures allows us to infer indirect contradictions.
Preparation"Sentiment analysis is the process of detecting positive or negative sentiment in text. All utterances are uttered in context. Analyzing subjective sentiment without context is hard. For bipolar text, context is learned by training or fine-tuning on both 'poles'."
"Conflict Graph" Of Contradictions
Connecting all the contradictions in a domain results in its perhaps intended "conflict graph". Credibility weights of edges gives us the quantitative means for analysis.
Relationships"An inferred conflict graph is most effectively modeled by an ER model (or entity-relationship model). Granularity of such a model is based on learned and inferred representations allowing for later refinements and even versioning."
Infer Implicit Context
Natural languages rely extensively on the "understood" or implicit context of sentences. Inferring elements of this context (from "parts of speech" statistical analysis) allows us to "fill in" the variables of the lexical scopes of our sentences.
Contexts"Part-of-speech (POS) tagging categorizes words in text depending on the definition of the word and its context. Generated ER models augment such inference mechanisms to create necessary 'contexts' on the fly."
Construct Dependency Graphs
Compilers successfully analyze dependency graphs of variables for optimization. Natural language text is already optimized, and we need to reverse engineer it to arrive to an explicit, or "precise", formulation of sentences.
Synthesis"Inferred contexts of sentences can be expanded to paragraphs and even documents using efficient and proven RDBMS 'tricks' and on-demand synthesized SQL."
Train With 'Annotated Reality'
"Parts of speech" decomposition of sentences allows us to greatly expand our corpus with labeled samples. Targeted explicit contradictions can be introduced to guide the "attention" mechanism of our subsequent training sessions.
Reports"Iterations of the training, inference and ER synthesis loops create the same dynamic that gave us the success of deep learning neural networks, with the added benefits of quantitative versioning.”