NeuroQL: A Neuro-Symbolic Language and Dataset for Inter-Subjective Reasoning

We present a new AI task and baseline solution for Inter-Subjective Reasoning. We define inter-subjective information, to be a mixture of objective and subjective information possibly shared by different parties. Examples may include commodities and their objective properties as reported by IR (Information Retrieval) systems, that need to be cross-referenced with subjective user reviews from an online forum. For an AI system to successfully reason about both, it needs to be able to combine symbolic reasoning of objective facts with the shared consensus found on subjective user reviews. To this end we introduce the NeuroQL dataset and DSL (Domain-specific Language) as a baseline solution for this problem.

We call this type of questions inter-subjective given that they involve a mixture of subjective and objective information that may be shared by different parties. Currently, answers to such queries are treated with a semi-manual two stage process by information systems. In the first step the user needs to pinpoint a single product of interest using a query involving only its objective properties and description (e.g a query involving the title, price, number of reviews etc. as in the example above). Then in the second step, the user needs to scan the reviews of this product either manually or with a similarity search in order to find relevant subjective opinions regarding the quality of the product (e.g. the quality of the bass in the above example). With the advent of deep-learning [24, 17] for natural-language processing [48] and more specifically deep-learning Q&A (Question & Answering) models [46], this second step can now be supplemented with neural retrieval and neural comprehension of reviews.

In this work we investigate the possibility of automating and merging these two stages that involve both symbolic reasoning (over structured factual information) and neuronal reasoning (over unstructured subjective sources). The problem of interfacing symbolic and neuronal reasoning, is a known open problem in AI literature involving neuro-symbolic systems [8, 32]. Our contribution is to propose a new AI task and baseline solution for inter-subjective queries and reasoning, as the one we saw above.

To test (H1) we experiment with a neural translation solution fine-tuned to this domain, aiming to distinguish not only between objective and subjective components, but also between different kinds of sub-query categories (i.e. opinion, title, price, reviews, manufacturing etc.). These sub-query categories have different translations in NeuroQL (as shown in the bottom part of Figure 2) which are highlighted to match their equivalent categories in natural language.

Furthermore, we hypothesize that:

(H2): Symbolic reasoning through unification (covering structured objective facts),

similar to the one found in prolog [9, 11] and datalog [6] systems, can be extended

with neuronal reasoning (for unstructured subjective data) in a disciplined manner,

producing satisfactory answers for inter-subjective queries.

Meaning, that the neuro-symbolic synthesis can be expressed in a concise syntactic and semantic form for NeuroQL users, while still being able to answer inter-subjective queries in a satisfactory way.

1 search (

2 bm25 _match ( asin_ . title == __title_ _,’headphones ’,__80),_

3 __asin_ . price == __price_ ,

4 op_filter ( lambda e: abs__(e[’? price ’] - 30__) < 10__),

5 __asin_ . total_reviews == __total_reviews_ ,

6 op_filter ( lambda e: e[’? total_reviews ’] >= 14000__),

7 asin_ . is_discontinued_by_manufacturer _==’no ’__,_

8 __asin_ . review == __review_ ,

9 neural_match (

10 review_ . text == __review_text_ _,’how is the bass ?’__,__5_

11 ),

12 neural_extract (

13 answers_ , __review_ . text == __review_text_ _,’how is the bass ?’__,__2_

14 )

15 )