PaperlayerView exact revision on arXiv

Source-linked AI summary

Think about it! Improving defeasible reasoning by first modeling the question scenario

Aman Madaan, Niket Tandon, Dheeraj Rajagopal, Peter Clark, Yiming Yang, Eduard Hovy

arXiv:2110.12349v1cs.AIcs.CL

TL;DR

The paper asks whether neural models can benefit from modeling a question scenario before answering a defeasible query. CURIOUS generates and corrects an inference graph, then encodes it with the query. It achieves state-of-the-art results on three defeasible reasoning datasets.

  • Problem

    The paper asks whether inference graphs can help neural models envision a question scenario before answering a defeasible query.

    Source: §1 · p. 1

  • Method

    CURIOUS generates an inference graph for the query, corrects graph errors, and uses the graph as additional input when answering.

    Source: §1 · p. 2

  • Results

    CURIOUS establishes a new state-of-the-art on three defeasible reasoning datasets.

    Source: §1 · p. 2, §8 · p. 9

  • Takeaways & Limitations

    The results illustrate that guiding a system to model a question scenario before answering can improve performance.

    Source: §1 · p. 2, §8 · p. 9

    The maximum percentage of errors occurred in δ-ATOMIC, in part due to low question quality.

    Source: prev. succ · p. 8

Abstract

from arXiv · show

Defeasible reasoning is the mode of reasoning where conclusions can be overturned by taking into account new evidence. Existing cognitive science literature on defeasible reasoning suggests that a person forms a mental model of the problem scenario before answering questions. Our research goal asks whether neural models can similarly benefit from envisioning the question scenario before answering a defeasible query. Our approach is, given a question, to have a model first create a graph of relevant influences, and then leverage that graph as an additional input when answering the question. Our system, CURIOUS, achieves a new state-of-the-art on three different defeasible reasoning datasets. This result is significant as it illustrates that performance can be improved by guiding a system to "think about" a question and explicitly model the scenario, rather than answering reflexively. Code, data, and pre-trained models are located at https://github.com/madaan/thinkaboutit.

1 Introduction

The paper asks whether modeling a question scenario with an inference graph can help machines answer defeasible queries. CURIOUS generates and corrects such graphs, then encodes them with the query; it reports state-of-the-art results on three datasets.

1 Introduction

  • Defeasible reasoning allows new information to modify a conclusion, such as a pillow weakening the inference that a fallen glass broke.This formulation follows Koons (2017) and Rudinger et al. (2020).Source: §1 · p. 1
  • The research question is whether an inference graph can help a machine model the question scenario before answering a defeasible query.The paper frames the graph as a mental model of the scenario, drawing on cognitive-science ideas.Source: §1 · p. 1
  • CURIOUS combines inference-graph generation, error correction, and graph-augmented question encoding to answer defeasible queries.The authors report that correcting graphs matters because higher-quality graphs are more helpful for downstream QA.Source: §1 · pp. 1–2
  • CURIOUS achieves a new state-of-the-art across three defeasible reasoning datasets.The datasets cover natural language inference, social norms, and commonsense reasoning; judicious encoding accounts for interactions among graph nodes.Source: §1 · p. 2

2 Task

The task tests whether a scenario strengthens or weakens a plausible inference from a premise to a hypothesis. It is framed as binary prediction from premise, hypothesis, and scenario sentences.

2 Task

  • The task concerns defeasible inference, in which additional information can modify a conclusion.This task formulation is the one defined by Rudinger et al. (2020).Source: §2 · p. 2
  • Given premise P, hypothesis H, and scenario S, predict whether S strengthens or weakens the plausible conclusion that H follows from P.The input is x = (P, H, S), and the output label is strengthens or weakens.Source: §2 · p. 2
  • The input components are sentences describing a premise, a hypothesis, and a scenario.The scenario determines whether the plausible inference from premise to hypothesis is strengthened or weakened.Source: §2 · p. 2

3 Approach

CURIOUS models a defeasible query with an inference graph and encodes that graph alongside the query. Its approach combines graph generation and correction with hierarchical mixture-of-experts pooling.

3 Approach

  • Inference graphs represent contextualizers, situations, hypotheses, and mediators as nodes in a model of the question scenario.Situation nodes represent new situations that might overturn an inference; contextualizers set the situation's context.Source: §3 · p. 2

3.1 Overview of CURIOUS

  • CURIOUS generates an inference graph, corrects it, and encodes the resulting graph with the query for classification.The system comprises GENinit, GENcorr, and a graph encoder.Source: §3.1 · pp. 2–3

3.2 Graph generator

  • GENinit generates inference graphs from defeasible queries using a sequence-to-sequence approach trained on WIQA process texts and influence graphs.Its WIQA training input includes premise, situation, and hypothesis; the target is the corresponding influence graph.Source: §3.2 · p. 3

3.3 Graph corrector

  • GENinit-generated graphs undesirably contained repeated nodes; these repetitions violate a graph’s semantic structure and introduce noise, with C+ and C− repeated despite their expected opposite semantics.Source: §3.3 · p. 3
  • GENcorr repairs an initial graph by mapping G′ to a corrected graph G*, then the corrected graph is supplied as query context.The corrector is trained as a T5 sequence-to-sequence model on generated input-output graph pairs.Source: §3.3 · pp. 3–4

3.4 Graph Encoder

  • The graph-augmented task pairs the concatenated query x = (P∥H∥S) with graph G to predict a strengthening or weakening label.The encoder receives both the query and the corrected inference graph.Source: §3.4 · p. 4

3.4.1 Encoding the query and nodes

  • A language model encodes the query and each graph node into dense representations before graph pooling.The query is represented as hx, while the node representations form hV.Source: §3.4.1 · p. 4, §3.4 · p. 4

3.4.2 Graph representations using MoE

  • Hierarchical MoE first pools node representations into a graph representation, then combines that representation with the query for classification.The design can select useful graph nodes and vary how much the classifier relies on the graph versus the query.Source: Figure 3 · p. 3, §3.4.2 · pp. 4–5

4 Experiments

Across three diverse defeasible-inference datasets, CURIOUS improves test accuracy over prior systems; experiments also examine graph correction, graph encoding, and expert-gating behavior. The analyses associate corrected graphs and MoE encoding with stronger task performance and characterize which graph information the model uses.

4.1 Experimental setup

  • Experiments measure test accuracy on δ-ATOMIC, δ-SNLI, and δ-SOCIAL, which cover commonsense reasoning, natural language inference, and social norms.The reported QA checkpoint is selected by development-set accuracy; the authors use p < 0.05 as the significance threshold.Source: §4.1 · p. 5

4.2 Results

  • CURIOUS achieves new state-of-the-art accuracy across δ-ATOMIC, δ-SNLI, and δ-SOCIAL, with consistent gains and about 4 points on δ-SNLI.The gains are reported in Table 2, which marks statistically significant results with an asterisk.Source: §4.2 · p. 5, Table 2 · p. 5

4.3.1 Impact of graph corrector

  • GENcorr adds 2 points across datasets over using GENinit graphs, although this interpretation assumes GENcorr reduces graph noise.The ablation consistently lowers performance when the corrector is removed.Source: §4.3.1 · p. 6, Table 3 · p. 6
  • GENcorr reduces repetitions per graph by approximately 40% (2.11 to 1.25) and the fraction of graphs with repetitions by 25.7%.The two reported measures are repetitions per graph and the percentage of graphs with at least one repetition.Source: §4.3.1 · p. 6, Table 4 · p. 6

4.3.2 Impact of graph encoder

  • The authors hypothesize that GCN message passing amplifies noise in generated graphs, while MoE can select useful nodes; a GCN-MoE hybrid performs about like GCN.STR encoding also helps the end task, supporting graph utility beyond the MoE encoder.Source: §4.3.2 · p. 6
  • MoE graph encoding significantly outperforms GCN on every dataset and STR on δ-SNLI and δ-SOCIAL.STR concatenates node strings; GCN refines node representations before attending to the question.Source: §4.3.2 · p. 6, Table 5 · p. 7

4.3.3 Detailed MoE analysis

  • In MOE-GX, graphs contribute more for strengthens, while the question receives higher weight for weakens across datasets.The authors conjecture that language models favor generating events that happen, making weakener events harder to generate.Source: §4.3.3 · p. 7, Figure 5 · p. 7
  • MOE-V favors situation nodes over contextualizers and mediators; situation-context gate weights are negatively correlated, and entropy is 0.52 (max 1.61) for MOE-V versus 0.08 (max 0.69) for MOE-GX.The authors report that only one of the situation and context node groups is activated at a time.Source: §4.3.3 · pp. 7–8, Figure 6 · p. 7, Figure 7 · p. 8

5 Error analysis

The error analysis examines graph errors in sampled development examples and identifies ignored but otherwise useful graphs as the largest reported category. It also notes low question quality as one contributor to δ-ATOMIC errors.

prev. succ

  • CURIOUS corrected previously failing cases, but also caused some previously successful examples to fail.The analysis manually annotated random development samples to examine graph errors.Source: prev. succ · p. 8
  • Among 50 annotated development examples, 42% had good but ignored graphs, 34% uninformative mediators, 20% repeated nodes, and 4% off-topic nodes.The sample contained 26 weakens and 24 strengthens examples.Source: prev. succ · p. 8
  • δ-ATOMIC had the highest percentage of errors, partly attributed to low question quality.Source: prev. succ · p. 8

6 Explainability

CURIOUS’s graph-based explanations are discussed in terms of plausibility and faithfulness. The authors examine whether its graph representations align with human-interpretable reasoning and with model performance.

Plausibility

  • The explanation discussion distinguishes plausibility—convincing interpretations for people—from faithfulness to the model’s actual reasoning process.The authors frame their approach as taking a step toward addressing both goals.Source: §6 · p. 8
  • MoE gate values indicate which graph components the model leveraged, offering a reasoning chain similar to explanations humans understand.A prior study found that human annotators selectively chose graph parts to explain model decisions and improve defeasible reasoning.Source: §6 · p. 8

Measuring faithfulness w.r.t. graphs

  • 82% of graphs are correct when CURIOUS answers correctly, compared with 42% when it answers incorrectly.The paper measures faithfulness to the reasoning graph through the correlation between model performance and graph correctness.Source: §6 · pp. 8–9

7 Related work

The related work connects CURIOUS to cognitive-science mental models and prior graph-based reasoning methods. It distinguishes its generated event inference graphs and MoE pooling from approaches that retrieve and encode knowledge graphs.

Mental Models

  • Mental models are coherent, constructed representations of the world; CURIOUS uses inference graphs as a machine’s model of the problem scenario.Constructing the graph is likened to asking clarification questions before answering, and to self-talk directed toward chains of influence (Shwartz et al., 2020).Source: §7 · p. 9

Injecting Commonsense Knowledge

Encoding Graph Representations

8 Summary and Conclusion

The paper models a question’s situation with an inference graph and pools graph representations using mixture-of-experts. It reports state-of-the-art results on three defeasible reasoning datasets and presents MoE pooling as an alternative to GCN for reasoning tasks.

8 Summary and Conclusion

  • CURIOUS establishes a new state-of-the-art on three defeasible reasoning datasets.The paper presents its approach as guiding a system to model a question before answering rather than answering reflexively.Source: §8 · p. 9
  • The paper presents MoE-based pooling as a strong alternative to GCN for graph-based reasoning, in both performance and explainability.Its situational model is an inference graph inspired by cognitive-science accounts of mental models.Source: §8 · p. 9
  • The conclusion argues that performance can improve when a system explicitly models the question scenario before answering.The authors contrast this with answering reflexively.Source: §8 · p. 9

A Training graph corrector

The appendix describes graph correction, analyzes MoE gradients, lists training details, and sketches an influence-graph schema.

A Training graph corrector

  • Appending the helps-or-hurts influence label to GENinit’s training input reduces repetitions by 13%.GENinit maps premise, situation, and hypothesis to a graph; because the influence label is unavailable at test time, GENcorr learns to map the initial graph to the target graph.Source: A Training graph corrector · p. 13
  • GENcorr corrects repetitions in generated graphs by learning a graph-to-graph mapping from the initial graph to the target graph.Figure 8 outlines the training-data generation process; an example describes repeated node pairs being replaced with meaningful labels.Source: A Training graph corrector · p. 13, Figure 8 · p. 13, B MoE gradient analysis · p. 14

B MoE gradient analysis

  • With cross-entropy, the gradients for an expert’s mixture probability and correct-class logit are zero at perfect prediction, which may slow expert specialization.The authors leave analysis of alternative loss functions for MoE representation pooling to future work; Figure 10 depicts the weighted expert output feeding a softmax.Source: B MoE gradient analysis · p. 13, Figure 10 · p. 14

C Hyperparameters

  • The experiments use a single Nvidia GeForce RTX 2080 Ti; the MoE gates and experts are single-layer MLPs with expert input size matching output size.The implementation is based on PyTorch and also uses PyTorch Lightning and Huggingface.Source: C Hyperparameters · p. 14

D Schema of an influence graph

E Runtime Analysis

The runtime analysis weighs graph-encoder costs against STR and considers ways to scale MoE parameters. Although both GCN and MoE increase runtime and parameter counts, the authors identify parameter sharing or deeper MoE layers as possible scaling approaches.

E Runtime Analysis

  • Compared with STR encoding, GCN and MoE use about 7% more parameters and have about 2x the runtime; MoE uses one expert per node, so its parameter count scales linearly with node count.Source: E Runtime Analysis · p. 15
  • Correlated behavior among semantically similar nodes suggests their experts could share parameters.Alternatively, MoE with more than two layers can make parameter growth logarithmic in the number of nodes.Source: E Runtime Analysis · p. 15

F Error Analysis Examples

The supplied material describes dev-set error cases, significance tests, a GCN encoder, the scope of the all-results comparison, and a graph-augmented reasoning algorithm. It includes a significance claim but does not provide the underlying result values.

F Error Analysis Examples

  • The dev-set examples cover CURIOUS errors where the baseline answers correctly, including an unused good graph, an off-topic graph from a sense error, and a wrong mediator.The cases are from δ-ATOMIC, δ-SNLI, and δ-SOCIAL, respectively.Source: F Error Analysis Examples · p. 15

G Significance Tests

  • Table 10 reports that results comparing with and without graphs are highly significant across the three datasets and two statistical tests.The tests are the micro-sign test and McNemar’s test.Source: Table 10 · p. 15, G Significance Tests · p. 15

H Description of GCN encoder

  • The GCN encoder initializes node representations, refines them through neighbor-based layers, then uses multiheaded attention to combine node and query representations into a graph representation.The graph representation is concatenated with the question representation for prediction.Source: H Description of GCN encoder · p. 19

I All results

  • The all-results comparison covers 18 combinations of two graph types, three datasets, and three graph-encoding schemes.Table 14 is identified as presenting results for different graph-encoder and graph-type combinations.Source: I All results · p. 19, Table 14 · p. 19

J Graph-augmented defeasible reasoning algorithm

  • The graph-augmented MoE algorithm encodes the query’s graph nodes, combines node representations, combines the query with the graph representation, and returns a prediction.The algorithm takes a language model and a defeasible query with graph as input.Source: J Graph-augmented defeasible reasoning… · pp. 19–20
Loading 2110.12349v1…