Truth-O-Meter: Fact-checking content generated by a LLM
Boris Galitsky
ABSTRACT
A text obtained by a Large Language Model (LLM) such as GPT4 usually has issues in terms of incorrectness and hallucinations. We build a fact-checking system 'Truth-O-Meter' which identifies wrong facts, comparing the generation results with the web and other sources of information, and suggests corrections. Text mining and web mining techniques are leveraged to identify correct corresponding sentences; also, the syntactic and semantic generalization procedure adopted to the content improvement task. To handle inconsistent sources while fact-checking, we rely on an argumentation analysis in the form of defeasible logic programming. We compare our fact checking engine with competitive approach based on reinforcement learning on top of LLM or token-based hallucination detection. Our approach is an instance of what we call "Shaped-charge learning architecture" which is intended to combine an efficient LLM with explainable inductive learning. It is observed that LLM content can be substantially improved for factual correctness and meaningfulness.
https://github.com/bgalitsky/Truth-O-...
SPEAKER BIO
Boris Galitsky contributed linguistic and machine learning technologies to Silicon Valley startups as well as companies like eBay and Oracle for over 25 years. Boris’ information extraction and sentiment analysis techniques assisted a number of acquisitions, such as Xoopit by Yahoo, Uptake by Groupon, Loglogic by Tibco and Zvents by eBay. His security-related technologies of document analysis contributed to acquisition of Elastica by Semantec. https://github.com/bgalitsky/relevanc...
As an architect of the Intelligent Bots project at Oracle, Boris developed a discourse analysis technique user for dialogue management and published in the book "Developing Enterprise Chatbots”. He also published a two-volume monograph “AI for CRM”, based on his experience developing Oracle Digital Assistant. Boris is Apache committer to OpenNLP where he created OpenNLP.Similarity component which is a basis for a semantically-enriched search engine and chatbot development.
Galitsky’s exploration and formalization of human seasoning culminated in the book “Computational Autism” broadly used by parents of children with autistic reasoning and rehabilitation personnel. Boris focus on medical domain led to another research monograph, “AI for Health Applications and Management”.
https://www.amazon.com/Books-Boris-Ga...
An Author of 150+ publications, 50+ patents and 6 books, Boris’s focus now is on improving content generation quality.
/ boris-galitsky-342109204
https://www.meetup.com/sf-bay-acm/eve...
0:00 Intro & Background
0:46 The problem: enhancing the factuality of LLMs
2:29 What Truth-O-Meter is doing
4:05 Github page
4:23 The plan
5:33 Language models make mistakes in tail data
6:19 Fact-checking in media is different than fact-checking LLM hallucinations
6:45 "Repairing" generated text by correction of entity values
7:39 Correction of entity values: web mining
8:43 Correction of entity values: what to get where
9:12 Correction of entity values: alignment
9:59 Correcting the wrong factuality
11:32 Raw and True content in health
14:42 Fact-checking: substituting raw entities with true entities
14:57 Fact-checking, correction and extension of the raw content
16:00 Personalized drug recommendation
16:01 Fact checking via web mining
16:18 Coordination between a discourse tree for raw and true texts
17:56 System architecture for GPT2/3
18:53 Characterizing correction scenarios for GPT2/3 in drug recommendation domain
19:43 ChatGPT and GPT4: Now what?
19:50 Question and answer correction
21:11 What needs to be done
22:27 Truth-O-meter improves an LLM by
25:38 Hallucination types
32:23 Handling multiple mutually inconsistent facts obtained from authoritative sources
33:07 We need a defeasible logic program to select the best candidate for truth, given all mined facts
33:30 Applying Natural Language Inference to find which mined facts defeat which other
34:21 Managing the candidate sentences for ground truth
35:03 Forming dialectical tree for search results
35:59 Truth-O-Meter in iterative mode
39:00 Hallucination correction pipeline
42:29 Evaluation of Truth-O-Meter in drug recommendation domain
49:25 Related work in LLM verification
49:26 Token-level Reference-free Hallucination Detection
50:05 Fact-checking can be expressed as Markov Decision Process
50:18 An approach to LLM-Truth-O-Meter based on reinforcement learning
50:20 Retrieval-augmented architecture
50:54 Conclusion
51:20 Q&A
1:03:05 Shaped-charge learning at a glance
1:04:00 Architecture is inspired by the idea of shaped charge projectile