PaperlayerView exact revision on arXiv

Source-linked AI summary

Textual Explanations for Automated Commentary Driving

Marc Alexander Kühn, Daniel Omeiza, Lars Kunze

arXiv:2304.08178v1cs.CLcs.AIcs.HCcs.LGcs.RO

TL;DR

Deep-learning vehicle controllers can be difficult to audit, and prior evaluation of the end-to-end explanation model covered only one dataset. This work evaluates that pipeline on SAX and modifies its language generator; the modified model improves BDD-X explanation BLEU by 7.7× over the standard model.

  • Problem

    The prior end-to-end model was evaluated on only one dataset with post-hoc free-text explanations, leaving its performance on additional datasets to be assessed.

    Source: §I · p. 1

  • Method

    The authors evaluate the SOTA pipeline on SAX and modify its language generator with PoS prediction and special-token penalties.

    Source: §I · p. 1

  • Results

    On BDD-X, combining PoS prediction and special-token penalties yielded 7.7× higher BLEU for explanation generation than the standard model.

    Source: 1) Quantitative Results · p. 4

  • Takeaways & Limitations

    SAX's real-time commentary and structured annotations provide a setting for learning from more realistic explanations.

    Source: §I · p. 1

    BLEU and METEOR rely mainly on sentence similarity and are limited for semantic comparison; the authors propose human evaluation.

    Source: §V · p. 6

Abstract

from arXiv · show

The provision of natural language explanations for the predictions of deep-learning-based vehicle controllers is critical as it enhances transparency and easy audit. In this work, a state-of-the-art (SOTA) prediction and explanation model is thoroughly evaluated and validated (as a benchmark) on the new Sense--Assess--eXplain (SAX). Additionally, we developed a new explainer model that improved over the baseline architecture in two ways: (i) an integration of part of speech prediction and (ii) an introduction of special token penalties. On the BLEU metric, our explanation generation technique outperformed SOTA by a factor of 7.7 when applied on the BDD-X dataset. The description generation technique is also improved by a factor of 1.3. Hence, our work contributes to the realisation of future explainable autonomous vehicles.

I. INTRODUCTION

The paper motivates evaluating an existing driving explanation model beyond its single prior dataset and introduces evaluation on SAX alongside two language-generator modifications. Prior work spans interpretable attention-based controllers and generated natural-language commentary.

I. INTRODUCTION

  • The prior end-to-end model had been tested on one dataset with post-hoc free-text explanations, motivating evaluation on additional datasets.The paper notes that black-box deep networks provide inadequate insight for justifying use in high-stakes scenarios.Source: §I · p. 1
  • The authors evaluate the SOTA pipeline on SAX and modify the language generator with PoS prediction and special-token penalties.The penalties target sentence structure through the loss function.Source: §I · p. 1

A. Explainable AI for Self-Driving Vehicles

  • Prior approaches include visual-attention maps with causal filtering, attention-based language generation, and tree-based commentary for selected driving actions.Image-captioning models have also been applied to driving data to generate action descriptions and explanations,.Source: A. Explainable AI for Self-Driving… · pp. 1–2

B. Berkeley DeepDrive eXplanation Dataset

BDD-X pairs video and sensor data with separate free-text action descriptions and explanations. SAX adds structured visual annotations and real-time driver commentary to support more realistic, standardized explanation data.

B. Berkeley DeepDrive eXplanation Dataset

  • BDD-X contains over 77 hours across 6,984 videos, with timestamped sensor data and separate post-hoc free-text action descriptions and explanations.Its action-description and explanation vocabularies contain 906 and 1,668 words, respectively.Source: B. Berkeley DeepDrive eXplanation… · p. 2

C. SAX Explanation Driving Dataset

  • SAX provides 9.5 hours of London driving with real-time audio commentary and camera, GPS, and CAN-bus data.Its explanation subset visually annotates agents by type, action, position, and influence on the driver's action.Source: C. SAX Explanation Driving Dataset · p. 2
  • SAX annotations structure influential agents into standardized ground-truth explanations; the dataset contains 491 sequences and a 34-word vocabulary.Most action sequences are about three seconds long.Source: C. SAX Explanation Driving Dataset · p. 2

III. PREDICTION AND EXPLANATION MODEL

The model combines visual features, an attention-based driving controller, and a language generator to produce action descriptions and explanations. The modified generator adds per-word part-of-speech prediction and incorporates its loss into training.

III. PREDICTION AND EXPLANATION MODEL

  • The pipeline encodes video frames, predicts acceleration and course with an attention-based LSTM controller, then generates per-word text using spatial and temporal attention.The language generator uses the controller's spatial attention alongside temporal attention across video frames.Source: §III · p. 2, Figure 2 · p. 2

A. Integration of Part of Speech Prediction

  • An added decoder layer predicts a part-of-speech tag for each word, and those predictions feed the subsequent word-prediction layer.The PoS decoder uses the LSTM hidden state and temporal context vector.Source: A. Integration of Part of Speech… · p. 3
  • PoS supervision adds a cross-entropy loss based on the ground-truth tags and generated tag outputs to the language-generator loss.The overall loss weights the PoS loss by λpos and the original generator loss by (1 − λpos).Source: A. Integration of Part of Speech… · p. 3

B. Integration of Special Token Penalties

The special-token penalties target the raw caption format's sentence boundaries and padding. They add loss penalties for generated NULL tokens and for incorrect counts of required structural tokens.

B. Integration of Special Token Penalties

  • Captions use START, sep, and END as structural tokens, with NULL tokens padding sentences to the maximum length.The separator divides the action description from its explanation.Source: B. Integration of Special Token… · p. 3
  • Each generated NULL token incurs a penalty intended to encourage word prediction rather than NULL when the model is uncertain near sentence end.The authors also state that this should further align generated and ground-truth sentence lengths.Source: B. Integration of Special Token… · p. 3
  • The loss also penalizes deviations from exactly one START, sep, and END token per generated caption.A γother penalty is added once for each respective token whose count differs from one.Source: B. Integration of Special Token… · p. 3

A. Training Procedure and Evaluation Set-Up

The evaluation uses METEOR and BLEU to score description and explanation text separately on BDD-X and SAX. Scores show a strong description advantage on BDD-X, unlike the smaller differences on SAX.

A. Training Procedure and Evaluation Set-Up

1) Quantitative Results:

  • On BDD-X, description METEOR is 4.4× higher and description BLEU is 44.9× higher than explanation scores.On SAX, explanation scores are slightly higher: 41.11% vs. 29.57% for BLEU and 68.30% vs. 62.82% for METEOR.Source: 1) Quantitative Results · p. 4
  • The SAX-trained generator scores higher overall than the BDD-X-trained generator, especially on explanations.The evaluation reports scores separately for each sentence's description and explanation parts.Source: 1) Quantitative Results · p. 4

2) Qualitative Analysis:

Qualitative results show different weaknesses across datasets: BDD-X generations often miss words or contain repetition and grammar problems, while SAX outputs more often preserve grammar but miss semantics. The discussion relates these patterns to annotation structure and dataset size.

2) Qualitative Analysis:

  • BDD-X produced no completely correct sentence; common errors were repetition and missing words, while bad-scoring samples also had grammar and meaning errors.Good-scoring samples captured semantics, but ground-truth sentences were longer and more complex; unclear frames also appeared among bad samples.Source: 2) Qualitative Analysis · p. 4
  • On SAX, descriptions in good-scoring samples were mostly correct, but explanations were never perfect; bad sentences often had valid grammar but wrong semantics.The discussion suggests SAX's smaller dataset may contribute to semantic errors.Source: 2) Qualitative Analysis · p. 4, 3) Discussion · p. 4

3) Discussion:

  • BDD-X's free-text annotations vary in syntax and length, which the authors suggest may make generation especially difficult for explanations.They also suggest that explaining causes is semantically harder than describing a scenario.Source: 3) Discussion · p. 4
  • SAX's more structured annotations reduce vocabulary and sentence-length variety; the authors suggest this, and its real-time commentary, may aid learning and scoring.They also propose that generating the right number of NULL tokens may be easier with uniform SAX samples.Source: 3) Discussion · p. 4

C. Evaluation of Language Generator Modifications

The language-generator modifications improve BDD-X scores, particularly for explanations, but have mixed effects on SAX. The best-performing modification depends on the dataset and whether the output is a description or explanation.

1) Quantitative Results:

  • On BDD-X, combining PoS prediction with special token penalties yields explanation METEOR 1.5× better and BLEU 7.7× higher than the standard model.Absolute explanation scores remain significantly lower than description scores.Source: 1) Quantitative Results · p. 4
  • On BDD-X, token penalties alone improve description METEOR by 1.6% and BLEU by 5.1%, equivalent to 1.3× better BLEU performance.Token penalties alone produce the best descriptions on BDD-X.Source: 1) Quantitative Results · p. 4
  • On SAX, the combined model improves description METEOR by 10.9% and BLEU by 5.8%, while PoS prediction alone modestly improves explanation METEOR by 0.8% and BLEU by 0.7%.Both variants with special token penalties score below the standard model on SAX explanations.Source: 1) Quantitative Results · p. 4
  • Changing λpos from 0.15 to 0.3 makes minimal differences: 0.2–0.5% for METEOR and 0.1–0.7% for BLEU.For description generation, γnull = 4 scores 47.23% METEOR and 20.38% BLEU, compared with 44.15% and 17.09% at γnull = 12.Source: 1) Quantitative Results · pp. 4–5

2) Qualitative Analysis:

Qualitative examples show that high-scoring outputs can capture scene meaning, while errors vary by dataset and sentence part. The illustrated BDD-X and SAX examples contrast the generated texts with their ground-truth annotations.

2) Qualitative Analysis:

  • In BDD-X examples, good-scoring samples capture semantics and nearly match the ground truth, while bad-scoring samples can have incomplete explanations or descriptions with wrong meaning.Ground-truth explanations have complex sentence structures.Source: 2) Qualitative Analysis · p. 5
  • For SAX, a good-scoring PoS sample generates both action description and explanation correctly, while a bad sample has grammatical descriptions but often wrong explanation semantics.Some bad explanations also repeat words.Source: 2) Qualitative Analysis · p. 5
  • In the BDD-X example, the combined model generates a nearly perfect sentence, while the token-penalty-only model performs worst qualitatively.The standard model omits the last word of each sentence part and substitutes “driving forward” for “driving down.”Source: 2) Qualitative Analysis · p. 5, Figure 4 · p. 5
  • In the SAX example, PoS prediction corrects the description's meaning, whereas the combined model produces wrong semantics in both parts and faulty explanation structure.The standard model's explanation is correct, but its description conveys the wrong meaning despite correct grammar.Source: 2) Qualitative Analysis · p. 5, Figure 4 · p. 5

3) Discussion:

The discussion reports that the model modifications behave differently across BDD-X and SAX, while NULL-penalty strength matters more than the PoS-loss weight. Qualitative results also show that sentence structure can be correct even when scene semantics are wrong.

3) Discussion:

  • On BDD-X, the PoS-and-token-penalty combination gave the largest explanation-generation improvement, with token penalties especially influential on BLEU.The authors suggest this may reflect the explanation part’s dependence on predicting the separator and the correct number of ending NULL tokens.Source: 3) Discussion · p. 5
  • On SAX, only description generation improved significantly; token penalties hurt explanation generation, possibly because uniform annotations already support learning syntax and sentence length.The authors suggest additional NULL penalties may prompt incorrect words instead of the correct NULL tokens.Source: 3) Discussion · pp. 5–6
  • Changing the PoS-loss weight had no significant performance effect, whereas NULL-penalty strength did, and an optimal NULL-penalty value exists.Source: 3) Discussion · p. 6
  • Qualitative outputs generally had improved sentence structure, but semantic errors remained, especially for complex captions, and individual sample quality did not always match dataset-level scores.Performance also varied by dataset, probably because ground-truth annotations differed in complexity.Source: 3) Discussion · p. 6

V. CONCLUSION

The paper validates the SOTA pipeline on SAX and reports that its modified pipeline outperforms the SOTA end-to-end model. It also identifies limits in text-based semantic evaluation and in estimating model weights and penalties.

V. CONCLUSION

  • On BDD-X, the modified pipeline improved explanation generation by factors of 7.7 on BLEU and 1.5 on METEOR, and description generation by a factor of 1.3.The modified pipeline outperformed the SOTA end-to-end model.Source: §V · p. 6
  • The SOTA prediction-and-explanation pipeline was validated on SAX, whose structured annotations and real-time driver commentary were associated with higher scores and better sentence grammar than BDD-X.Source: §V · p. 6
  • BLEU and METEOR rely mainly on sentence similarity and are limited for comparing semantics; optimal weights and penalty values also need more rigorous estimation.The authors also identify potential gains from higher-performance feature processing or more sophisticated image representations.Source: §V · p. 6
Loading 2304.08178v1…