ibm-research/knowledge_consistency_of_LLMs

What it is:

Each dataset in this delivery is made up of query clusters that test an aspect of the consistency of the LLM knowledge about a particular domain. All the questions in each cluster are meant to be answered either 'yes' or 'no'. When the answers vary within a cluster, the knowledge is said to be inconsistent. When all the questions in a cluster are answered 'no' when the expected answer is 'yes' (or viceversa), the knowledge is said to be 'incomplete' (i.e., maybe the LLM wasn't trained in that particular domain). It is our experience that incomplete clusters are very few (less than 3%) meaning that the LLMs we have tested know about the domains included here (see below for a list of the individual datasets), as opposed to inconsistent clusters, which can be between 6%-20% of the total clusters.

The image below indicates the types of edges the query clusters are supposed to test. It is worth noting that these correspond to common sense axioms about conceptualization, like the fact that subConceptOf is transitive (4) or that subconcepts inherit the properties of their parent concepts (5). These axioms are listed in the accompanying paper (see below)

How it is made:

The questions and clusters are automatically generated from a knowledge graph from seed concepts and properties. In our case, we have used Wikidata, a well known knowledge graph. The result is an RDF/OWL subgraph that can be queried and reasoned over using Semantic Web technology. The figure below summarizes the steps used. The last two steps refer to a possible use case for this dataset, including using in-context learning to improve the performance of the dataset.

Types of query clusters

There are different types of query clusters depending on what aspect of the knowledge graph and its deductive closure they capture:

Edge clusters test a single edge using different questions. For example, to test the edge ('orthopedic pediatric surgeon', IsA, 'orthopedic surgeon), the positive or 'edge_yes' (expected answer is 'yes') cluster is:

  "is 'orthopedic pediatric surgeon' a subconcept of 'orthopedic surgeon' ?",
  "is 'orthopedic pediatric surgeon' a type of 'orthopedic surgeon' ?",
  "is every kind of 'orthopedic pediatric surgeon' also a kind of 'orthopedic surgeon' ?",
  "is 'orthopedic pediatric surgeon' a subcategory of 'orthopedic surgeon' ?"

There are also inverse edge clusters (with questions like "is 'orthopedic surgeon' a subconcept of 'orthopedic pediatric surgeon' ?") and negative or 'edge_no' clusters (with questions like "is 'orthopedic pediatric surgeon' a subconcept of 'dermatologist' ?")

Hierarchy clusters measure the consistency of a given path, including n-hop virtual edges (in graph's the deductive closure). For example, the path ('orthopedic surgeon', 'surgeon', 'medical specialist', 'medical occupation') is tested by the cluster below

  "is 'orthopedic surgeon' a subconcept of 'surgeon' ?",
  "is 'orthopedic surgeon' a type of 'surgeon' ?",
  "is every kind of 'orthopedic surgeon' also a kind of 'surgeon' ?",
  "is 'orthopedic surgeon' a subcategory of 'surgeon' ?",
  "is 'orthopedic surgeon' a subconcept of 'medical specialist' ?",
  "is 'orthopedic surgeon' a type of 'medical specialist' ?",
  "is every kind of 'orthopedic surgeon' also a kind of 'medical specialist' ?",
  "is 'orthopedic surgeon' a subcategory of 'medical specialist' ?",
  "is 'orthopedic surgeon' a subconcept of 'medical_occupation' ?",
  "is 'orthopedic surgeon' a type of 'medical_occupation' ?",
  "is every kind of 'orthopedic surgeon' also a kind of 'medical_occupation' ?",
  "is 'orthopedic surgeon' a subcategory of 'medical_occupation' ?"

Property inheritance clusters test the most basic property of conceptualization. If an orthopedic surgeon is a type of surgeon, we expect that all the properties of surgeons, e.g., having to be board certified, having attended medical school or working on the field of surgery, are inherited by orthopedic surgeons. The example below tests the later:

  "is 'orthopedic surgeon' a subconcept of 'surgeon' ?",
  "is 'orthopedic surgeon' a type of 'surgeon' ?",
  "is every kind of 'orthopedic surgeon' also a kind of 'surgeon' ?",
  "is 'orthopedic surgeon' a subcategory of 'surgeon' ?",
  "is the following statement true? 'orthopedic surgeon works on the field of  surgery' ",
  "is the following statement true? 'surgeon works on the field of  surgery' ",
  "is it accurate to say that  'orthopedic surgeon works on the field of  surgery'? ",
  "is it accurate to say that  'surgeon works on the field of  surgery'? "

List of datasets

To show the versatility of our approach, we have constructed similar datasets in the domains below. We test one property inheritance per dataset. The Wikidata main QNode (the node corresponding to the entities) and PNode (the node corresponding to the property) are indicated in parenthesis.

domain	top concept	WD concept	main property	WD property
Academic Disciplines	"Academic Discipline"	https://www.wikidata.org/wiki/Q11862829	"has use"	https://www.wikidata.org/wiki/Property:P366
Dishes	"Dish"	https://www.wikidata.org/wiki/Q746549	"has parts"	https://www.wikidata.org/wiki/Property:P527
Financial products	"Financial product"	https://www.wikidata.org/wiki/Q15809678	"used by"	https://www.wikidata.org/wiki/Property:P1535
Home appliances	"Home appliance"	https://www.wikidata.org/wiki/Q212920	"has use"	https://www.wikidata.org/wiki/Property:P366
Medical specialties	"Medical specialty"	https://www.wikidata.org/wiki/Q930752	"field of occupation"	https://www.wikidata.org/wiki/Property:P425
Music genres	"Music genre"	https://www.wikidata.org/wiki/Q188451	"practiced by"	https://www.wikidata.org/wiki/Property:P3095
Natural disasters	"Natural disaster"	https://www.wikidata.org/wiki/Q8065	"has cause"	https://www.wikidata.org/wiki/Property:P828
Software	"Software"	https://www.wikidata.org/wiki/Q7397	"studied in"	https://www.wikidata.org/wiki/Property:P7397

The size and configuration of the datasets is listed below

domain	edges_yes	edges_no	edges_in	hierarchies	property hierarchies
Academic Disciplines	52	308	52	30	1
Dishes	197	519	197	62	121
Financial product	112	433	108	40	32
Home appliances	58	261	58	31	13
Medical specialties	122	386	114	55	63
Music genres	490	807	488	212	139
Natural disasters	45	225	44	21	22
Software	80	572	79	114	4

Want to know more?

For background and motivation on this dataset, please check https://arxiv.org/abs/2405.20163 Also to be published in COLM 2024,

@inproceedings{Uceda_2024_1,
  title={Reasoning about concepts with LLMs: Inconsistencies abound},
  author={Rosario Uceda Sosa and Karthikeyan Natesan Ramamurthy and Maria Chang and Moninder Singh},
  booktitle={Proc.\ 1st Conference on Language Modeling (COLM 24)},
  year={2024}
}

Questions? Comments?

Please contact [email protected], [email protected], [email protected] or [email protected]