NL Texts Research Articles

Generation of Training Examples for Tabular Natural Language Inference

Tabular data is becoming increasingly important in Natural Language Processing (NLP) tasks, such as Tabular Natural Language Inference (TNLI). Given a table and a hypothesis expressed in NL text, the goal is to assess if the former structured data supports or refutes the latter. In this work, we focus on the role played by the annotated data in training the inference model. We introduce a system, Tenet, for the automatic augmentation and generation of training examples for TNLI. Given the tables, existing approaches are either based on human annotators, and thus expensive, or on methods that produce simple examples that lack data variety and complex reasoning. Instead, our approach is built around the intuition that SQL queries are the right tool to achieve variety in the generated examples, both in terms of data variety and reasoning complexity. The first is achieved by evidence-queries that identify cell values over tables according to different data patterns. Once the data for the example is identified, semantic-queries describe the different ways such data can be identified with standard SQL clauses. These rich descriptions are then verbalized as text to create the annotated examples for the TNLI task. The same approach is also extended to create counterfactual examples, i.e., examples where the hypothesis is false, with a method based on injecting errors in the original (clean) table. For all steps, we introduce generic generation algorithms that take as input only the tables. For our experimental study, we use three datasets from the TNLI literature and two crafted by us on more complex tables. Tenet generates human-like examples, which lead to the effective training of several inference models with results comparable to those obtained by training the same models with manually-written examples.

Glossa — A Formal Language as a Mapping Mechanism of NL Sentences into SPN State Machine for Actions/Events Association

Natural Language Understanding (NLU) is an old and really challenging field with a variety of research work published on it. In this paper we present a formal language methodology based on a state machine for efficiently representing natural language events/actions and their associations in well-written documents. The methodology consists of the following steps. We firstly apply Anaphora Resolution (AR) to the pre-processing natural language text. Then we extract the kernel(s) of each sentence. These kernels are formally represented using a formal language, (Glossa) to map the language expressions (kernels) into Stochastic Petri Nets (SPN) graphs. Finally we apply a set of rules to combine the SPN graphs in order to achieve the associations of actions/events in time. Special emphasis of this paper is the mapping of kernels of NL sentences into SPN graphs. Note that this work does not cover all the aspects of the NLU. Examples of SPN graphs of different NL texts, produced by our proposed methodology are given.

On Understanding Understanding. Perception-Based Processing of NL Texts in SCIP Systems, or Meaning Constitution as Visualized Learning

Inspired by information systems theory, semiotic cognitive information processing (SCIP) is grounded in (natural/artificial) system-environment situations. SCIP systems' knowledge-based processing of information makes it cognitive, their sign and symbol generation, manipulation, and understanding capabilities render it semiotic. Based upon structures whose representational status is not a presupposition to, but a result from recursive processing, SCIP algorithms initiate and modify the structures they are operating on to realize (rather than simulate) language understanding by meaning constitution. Thus, the symbolic (de)composition of propositional structures in traditional semantics is complemented by SCIP, which models learning and understanding dynamically by visualizing what is understood in a perception-based, subsymbolic, multiresolutional way of processing natural language discourse. An experimental 2-dim scenario with object locations described relative to a mobile agent's varying positions allows to test the SCIP systems' performance against human natural language understanding in a controlled way. The implementation of the SCIP system-environment testbed is due to my PhD-students, Christoph Flores and Daniel John, whose design and programming proficiencies are thankfully appreciated.

Integrating discourse pragmatics and propositional knowledge for multilingual natural language processing

We present an integrated knowledge representation system for natural language processing (NLP) whose main distinguishing feature is its emphasis on encoding not only the usual propositional structure of the utterances in the input text, but also capturing an entire complex of nonpropositional — discourse, attitudinal, and other pragmatic — meanings that NL texts always carry. The need for discourse pragmatics, together with generic semantic information, is demonstrated in the context of anaphoric and definite noun phrase resolution for accurate machine translation. The major types of requisite pragmatic knowledge are presented, and an extension of a frame-based formalism developed in the context of the TRANSLATOR system is proposed as a first-pass codification of the integrated knowledge base.