Source-linked AI summary
Unsupervised Commonsense Question Answering with Self-Talk
Vered Shwartz, Peter West, Ronan Le Bras, Chandra Bhagavatula, Yejin Choi
TL;DR
Commonsense language understanding requires implicit background knowledge, but relying on pretrained language models or external knowledge sources has limitations. The paper studies an unsupervised self-talk framework that generates and integrates background knowledge from pretrained language models, substantially improving zero-shot performance on four of six tasks and performing comparably to external-knowledge models. However, human judgments indicate that many helpful clarifications are unhelpful or incorrect.
Problem
Commonsense language understanding requires implicit background knowledge, while current systems rely primarily on pretrained language models or external knowledge sources with shortcomings.
Method
The paper studies an unsupervised framework that generates and integrates background knowledge from pretrained language models without external knowledge or additional supervision.
Results
The method substantially improves a zero-shot LM-score baseline on 4 out of 6 tasks and performs on par with, or sometimes better than, models using external knowledge.
Takeaways & Limitations
Self-talk can provide empirical gains similar to external knowledge resources for multiple-choice commonsense tasks.
Takeaways & Limitations
Among clarifications used by the answer scorer, humans judged about 60% of those yielding correct predictions and 12% of those yielding incorrect predictions factually correct.
Abstract
from arXiv · showhide
Natural language understanding involves reading between the lines with implicit background knowledge. Current systems either rely on pre-trained language models as the sole implicit source of world knowledge, or resort to external knowledge bases (KBs) to incorporate additional relevant knowledge. We propose an unsupervised framework based on self-talk as a novel alternative to multiple-choice commonsense tasks. Inspired by inquiry-based discovery learning (Bruner, 1961), our approach inquires language models with a number of information seeking questions such as "$\textit{what is the definition of ...}$" to discover additional background knowledge. Empirical results demonstrate that the self-talk procedure substantially improves the performance of zero-shot language model baselines on four out of six commonsense benchmarks, and competes with models that obtain knowledge from external KBs. While our approach improves performance on several benchmarks, the self-talk induced knowledge even when leading to correct answers is not always seen as useful by human judges, raising interesting questions about the inner-workings of pre-trained language models for commonsense reasoning.
1 Introduction
Commonsense understanding requires implicit background knowledge, but language models have coverage, precision, and reasoning limitations. The paper proposes self-talk as an unsupervised alternative to external knowledge bases and reports gains on most evaluated tasks.
- Pre-trained language models improve NLU partly through world knowledge learned from massive text, but their knowledge coverage and precision remain limited.Prior work therefore incorporated external knowledge bases to increase high-precision knowledge coverage and make intermediate reasoning steps explicit.
- The paper studies pre-trained language models as an alternative to external knowledge bases for commonsense question answering.
- Self-talk generates clarification questions and answers with language models, then incorporates those clarifications as additional context for answer scoring.The process is framed as inquiry-based discovery learning and does not require external knowledge or additional supervision.
- On 4 out of 6 tasks, self-talk substantially improves a zero-shot LM-score baseline and performs on par with, or sometimes better than, models using external knowledge.
- Human judges often found self-talk clarifications unhelpful or incorrect even when they helped produce correct predictions.The finding motivates research on robust and correct knowledge integration.
2 Tasks
The paper evaluates multiple-choice commonsense question answering across six tasks spanning causal, conceptual, temporal, social, physical, and coreference reasoning.
- Each task instance contains an optional context, an optional question, and several answer choices.
- COPA asks users to choose a plausible cause or result of an event from two alternatives.
- CommonSenseQA asks general questions about ConceptNet concepts, with distractors related to the target concept.
- MC-TACO tests temporal commonsense about event ordering, duration, frequency, and typical time using adversarially selected distractors.
- Social IQa covers social interactions from ATOMIC, including motivations and likely following actions, while PIQA tests physical commonsense involving everyday-object goals.
- WinoGrande is a crowdsourced, adversarially filtered large-scale version of the Winograd Schema Challenge using placeholders instead of pronouns.
3 Models
The models score answer choices with language models, optionally augmenting each choice with clarifications from external resources or LM-generated self-talk. Self-talk creates questions and answers about the context to produce additional knowledge.
- 3.1 LM-only Baseline: The baseline represents each answer choice by combining the context, question, and choice, then predicts the option with the lowest language-model cross-entropy.Task-specific templates determine how the texts are combined.
- 3 Models: With clarifications, each answer choice receives the best score obtained by combining it with any clarification in the clarification list.
- 3.2 Baseline Model with External Knowledge: External-resource clarifications come from ConceptNet relation paths, Google N-gram co-occurrences, and COMeT-generated commonsense dimensions.ConceptNet paths are limited to two edges, while COMeT relations can be selected heuristically for particular tasks.
- 3.3 Self-talk Model: The self-talk model generates clarification questions from task-specific prefixes and samples five questions per prefix conditioned on the context.Questions are limited to six additional tokens and use nucleus sampling with p = 0.2.
- 3.3 Self-talk Model: For each well-formed question, the model generates answers using paired answer prefixes, the context, and the question, retaining well-formed clarifications.It generates ten answers per question, with answer length limited to ten tokens and nucleus sampling p = 0.5.
- 3.3 Self-talk Model: Because the generators were not trained for relevance, coherence, or factual correctness, some generated clarifications may be unhelpful.
4 Results
Self-talk performs competitively with external-knowledge models across commonsense benchmarks, although its benefits vary substantially by task and combining knowledge sources is usually unhelpful.
- Self-talk performs on par or within a few points of external-knowledge models across the evaluated tasks.Table 2 compares the best setup for each model type using development accuracy, with test accuracy reported when available.
- Zero-shot models underperform state-of-the-art supervised models overall but substantially outperform majority baselines on most tasks.WinoGrande is the exception, where zero-shot models only slightly outperform the majority baseline.
- Among knowledge-informed models, COMeT achieves the best performance across tasks.The authors attribute this likely to COMeT’s dynamic context-conditioned predictions and broader coverage than static knowledge sources.
- Relative improvement over the zero-shot baseline varies substantially across tasks, although it is fairly uniform across knowledge sources.Some tasks benefit from added knowledge, while others show no benefit.
- Combining clarifications from all knowledge sources is generally not beneficial, except on MC-TACO.On MC-TACO, the combination adds +7.9 points to development accuracy, reaching 66.7; the authors suggest other resources may add noise.
5 Analysis
The analysis evaluates clarification quality intrinsically and examines both useful and harmful clarifications across tasks and knowledge sources. Although many clarifications that helped the model were relevant and factually correct, humans judged only 40% helpful on average, while harmful clarifications commonly reflected irrelevance, poor form, or insufficient usefulness.
- Evaluation setup: The study evaluates clarification quality intrinsically because end-task performance alone does not establish whether generated knowledge is useful.The analysis focuses on useful and harmful clarifications because preliminary experiments found substantial clarification noise.
- Evaluation setup: Human evaluation measured grammaticality, relevance, factual correctness, and helpfulness for sampled useful clarifications.Workers assessed clarification questions and answers; answers were judged for factual correctness and whether they added helpful information.
- Useful clarifications: Most clarifications were relevant and factually correct, but only 40% on average were considered helpful despite helping the model predict correctly.This indicates a mismatch between model-utilized knowledge and human judgments of helpfulness.
- Useful clarifications: Clarification quality varied by knowledge source: XLNet performed worst across measures, while ConceptNet clarifications were often judged irrelevant and LM-generated clarifications sometimes factually incorrect.ConceptNet’s narrow pairwise relations may limit the kinds of clarifications it provides.
- Useful clarifications: Pre-trained LMs performed particularly well at providing definitions in useful and relevant clarifications.Table 3 categorizes the types of knowledge found in these clarifications.
- Harmful clarifications: Among harmful clarifications, 12.4% supplied alternative correct answers phrased differently from the gold answer.Other harmful cases included irrelevant, ungrammatical or nonsensical, relevant-but-unhelpful, factually incorrect, repeated, wrong-sense, and dataset-error clarifications.
6 Related Work
Related work incorporates external knowledge into neural models through symbolic resources, retrieval, embeddings, rules, and attention-based or scoring mechanisms. This paper instead generates knowledge from pre-trained language models, addressing their ability to supply world knowledge while acknowledging risks of inaccurate extraction and the data demands of supervised explanations.
- External knowledge in neural models: External-knowledge approaches combine a task, neural model, knowledge source, and incorporation method, often targeting commonsense reasoning tasks.Examples include story cloze, machine comprehension, and related question-answering settings.
- External knowledge in neural models: ConceptNet is the dominant external resource, while other approaches use WordNet, corpus retrieval, knowledge-base embeddings, and hand-crafted rules.Knowledge is commonly extracted as relation paths between concepts and entities in the input.
- External knowledge in neural models: External knowledge is incorporated through vector representations and attention, answer scoring and pruning, or auxiliary multitask training.These mechanisms connect symbolic knowledge to neural input representations or answer selection.
- External knowledge in neural models: The paper presents its method as the first to generate knowledge from pre-trained language models and incorporate it as external knowledge into question answering.Concurrent work used separate language models to generate hypotheses and score answers for CommonSenseQA.
- Extracting knowledge from LMs: Pre-trained LMs can encode world knowledge and support knowledge-base completion when relations are converted into natural-language templates.Prior work queried masked or prompted language models for missing facts.
- Extracting knowledge from LMs: Extracted LM knowledge can be inaccurate because models may treat negated facts as likely and over-generalize to incorrect facts.These findings qualify the use of language models as knowledge providers.
- Generating questions and explanations: Question-generation research spans data augmentation, question answering, conversational reading, simplification, and uncertainty reduction.Related systems include supervised question-answering or reinforcement-learning frameworks.
- Generating questions and explanations: Human-explanation approaches add generated explanations as inputs but require collecting specific explanations for each new dataset.This dataset-specific supervision distinguishes them from the paper’s unsupervised approach.
7 Discussion and Conclusion
The paper presents an unsupervised framework that generates and integrates background knowledge from pre-trained language models for multiple-choice commonsense tasks. It reports gains over zero-shot baselines and discusses shortcomings of LM-generated knowledge.
- 7 Discussion and Conclusion: The framework generates and integrates background knowledge from pre-trained language models without external knowledge or additional supervision.It makes one additional reasoning step explicit through generated clarifications.
- 7 Discussion and Conclusion: On most tasks, the framework performs substantially better than the baseline and similarly to a model with external knowledge resources.
- 7 Discussion and Conclusion: LM knowledge providers have insufficient coverage, insufficient precision, and limited reasoning capabilities.
- 7 Discussion and Conclusion: About 60% of clarifications yielding correct predictions and 12% yielding incorrect predictions were judged factually correct by humans.
- 7 Discussion and Conclusion: A preliminary two-hop experiment produced mixed results, motivating dialogue-based clarification generation as a future direction.
B Best Language Model
The study compares language models used across knowledge sources and tasks, finding a general preference for GPT-2, especially larger models, with an exception for COPA.
- B Best Language Model: GPT-2, particularly larger models, is generally preferred across knowledge sources, except that distilled GPT-2 works best for COPA.
- B Best Language Model: A possible explanation for distilled GPT-2 performing best on COPA is that distillation reduces the likelihood of rare words.The passage connects this property to COPA’s simple sentences.
- B Best Language Model: XLNet models perform poorly, perhaps because their training corpus is smaller than GPT-2’s.The passage contrasts 16GB for XLNet with 40GB for GPT-2.
C.1 Useful Clarifications
Most helpful clarifications are grammatical or understandable to humans, although some useful clarifications contain grammar errors.
- C.1 Useful Clarifications: The majority of helpful clarifications are considered grammatical by human judges.
- C.1 Useful Clarifications: XLNet models are slightly worse in terms of grammaticality than other knowledge sources.
- C.1 Useful Clarifications: A clarification with grammar errors can still help the language model determine the correct answer.The example concerns an XLNet-generated clarification for a PIQA instance.
C.2 Harmful Clarifications
Harmful clarifications produce different error patterns across tasks and knowledge sources. Language models often make incorrect or nonsensical statements, while static resources have distinct relevance and word-sense limitations.
- C.2 Harmful Clarifications: Social IQa and CommonSenseQA often receive alternative correct answers among harmful clarifications, unlike WinoGrande, which allows only one correct answer.
- C.2 Harmful Clarifications: MC-TACO clarifications are more irrelevant than average, potentially reflecting limited temporal commonsense or misguided attempts to extract it.
- C.2 Harmful Clarifications: All knowledge sources except ConceptNet make incorrect statements, while language models also tend to produce nonsensical statements, especially XLNet.
- C.2 Harmful Clarifications: ConceptNet tends to generate irrelevant clarifications and is insensitive to word senses because it is a static resource.
- C.2 Harmful Clarifications: Google Ngrams does not show the same word-sense issue, likely because unrelated senses of polysemous terms typically do not co-occur.