Machine Reasoning for Multi‑turn Semantic Parsing and Question Answering

Reasoning is a crucial and challenging task in natural language processing, aiming to generate outputs for unseen inputs by manipulating existing knowledge with inference techniques. This paper introduces the latest methods and progress of machine reasoning for multi‑turn semantic analysis and question answering.

Multi‑turn conversational QA and semantic parsing are core problems for voice assistants, chatbots, and search engines. Effective understanding of dialogue history is essential because users often omit entities or intents in follow‑up questions.

The proposed machine reasoning framework leverages three types of knowledge:

Grammar knowledge : guides the generation of each token in the semantic representation, ensuring syntactic correctness.

Context knowledge : records semantic parses of previous turns and reuses them to handle omissions in multi‑turn scenarios.

Data knowledge : retrieves training instances similar to the current input and applies a meta‑learning based inference model to produce a personalized model for the specific example.

For knowledge‑graph based QA, a set of grammar operations (e.g., lookup, compare, count, copy‑history) is defined, each acting as a deduction rule mapping data types to functions.

The system adopts a top‑down sequence‑to‑action model that converts serialized semantic representations into action sequences respecting the grammar, while naturally incorporating context knowledge via a dialog memory that can copy previous semantic sub‑sequences.

To mitigate the tendency of neural generators to produce generic sequences, a meta‑learning inference model is introduced: for any input, similar training samples are retrieved, and the base model f(θ) is fine‑tuned to a task‑specific model f(θ′). Both the base model and the fine‑tuning process are trained jointly.

Experiments on the IBM CSQA multi‑turn complex QA benchmark demonstrate that the proposed approach achieves state‑of‑the‑art performance.

In conclusion, the paper shows that integrating grammar, context, and data knowledge through machine reasoning substantially improves multi‑turn semantic parsing and QA, and it anticipates broader applications of machine reasoning to other inference tasks.

References: [1] Zhou et al., 2019; [2] Duan & Zhou, 2018; [3] Guo et al., NeurIPS 2018; [4] Guo et al., ACL 2019; [5] Sun et al., NLPCC 2019; [6] Hashimoto et al., NeurIPS 2018.