PaperlayerView exact revision on arXiv

Source-linked AI summary

Pointwise Mutual Information Based Metric and Decoding Strategy for Faithful Generation in Document Grounded Dialogs

Yatin Nandwani, Vineet Kumar, Dinesh Raghu, Sachindra Joshi, Luis A. Lastras

arXiv:2305.12191v2cs.CL

TL;DR

Document-grounded dialogue faithfulness metrics often compare responses with documents but may overlook dialogue history, which matters for context-dependent replies. The paper introduces conditional PMI-based faithfulness scoring and decoding, reporting improved human-judgment correlation on BEGIN and more faithful responses across three datasets.

  • Problem

    Existing automated metrics compare generated responses with the grounding document but are not well aligned with human judgments and can ignore dialogue history needed to interpret context-dependent responses.

    Source: §1 · p. 1

  • Method

    PMI–FAITH scores conditional pointwise mutual information between a response and its document given dialogue history, and PMI–DECODE combines faithfulness with response likelihood during decoding.

    Source: §1 · p. 2

  • Results

    PMI–FAITH improves over existing metrics on BEGIN, while PMI–DECODE generates more faithful responses than standard likelihood decoding on three datasets; its F1 gain over Q2 is 8.7%.

    Source: §1 · p. 2, §5.2 · p. 6, §6 · p. 9

  • Takeaways & Limitations

    The reported results support conditional PMI as a basis for measuring faithfulness and guiding response generation in document-grounded dialogue.

    Source: §1 · p. 2, §6 · p. 9

    PMI–DECODE was observed to be unable to respond to generic chit-chat or pleasantries such as greetings and goodbyes.

    Source: Limitations · p. 9

Abstract

from arXiv · show

A major concern in using deep learning based generative models for document-grounded dialogs is the potential generation of responses that are not \textit{faithful} to the underlying document. Existing automated metrics used for evaluating the faithfulness of response with respect to the grounding document measure the degree of similarity between the generated response and the document's content. However, these automated metrics are far from being well aligned with human judgments. Therefore, to improve the measurement of faithfulness, we propose a new metric that utilizes (Conditional) Point-wise Mutual Information (PMI) between the generated response and the source document, conditioned on the dialogue. PMI quantifies the extent to which the document influences the generated response -- with a higher PMI indicating a more faithful response. We build upon this idea to create a new decoding technique that incorporates PMI into the response generation process to predict more faithful responses. Our experiments on the BEGIN benchmark demonstrate an improved correlation of our metric with human evaluation. We also show that our decoding technique is effective in generating more faithful responses when compared to standard decoding techniques on a set of publicly available document-grounded dialog datasets.

1 Introduction

Document-grounded dialogue agents must remain faithful to their source, yet existing metrics can miss responses whose meaning depends on dialogue history. The paper introduces PMI–FAITH and PMI–DECODE to measure and encourage faithfulness using the document and dialogue context.

1 Introduction

  • Existing faithfulness metrics can miss context-dependent responses because they ignore dialogue history and such replies may lack domain-specific words.Similarity metrics can then be ineffective, while question-answer pairs generated from these replies may capture incomplete information.Source: §1 · pp. 1–2
  • PMI–FAITH measures faithfulness as conditional pointwise mutual information between a response and its document, given dialogue history.It reflects whether the response is more likely with the document than without it.Source: §1 · p. 2
  • PMI–DECODE combines response likelihood and faithfulness in a decoding objective, enabled by PMI–FAITH’s autoregressive factorization.The paper reports better human-judgment correlation for PMI–FAITH on BEGIN and more faithful responses from PMI–DECODE than standard likelihood decoding on three datasets.Source: §1 · p. 2

2 Related Work

Prior faithfulness work includes metrics and benchmarks for document-grounded dialogue, but the paper identifies dialogue history as an overlooked part of evaluation. It positions PMI–FAITH as the first CPMI faithfulness metric for responses conditioned on both a document and dialogue history.

2 Related Work

  • Faithfulness is distinct from hallucination: a response may be non-hallucinatory yet unfaithful if its information cannot be verified from the document.The paper treats faithful responses as a subset of responses that are not hallucinations.Source: §2 · p. 2
  • Existing classifier-based metrics and Q2 evaluate document faithfulness, but Q2 ignores dialogue history and may fail on non-self-contained responses.PMI–FAITH addresses this issue; the paper also notes that its metric requires no training and is agnostic to the underlying data.Source: §2 · p. 2
  • The authors use Dziri et al.’s benchmark for document-grounded dialogues to compare faithfulness metrics.They distinguish it from benchmarks intended to standardize datasets across generation tasks.Source: §2 · pp. 2–3
  • The paper presents its work as the first use of CPMI to evaluate faithfulness against a document and dialogue history in document-grounded response generation.Earlier mutual-information work targeted generic responses, while Ren et al. (2023) used CPMI to evaluate relevance against a reference hypothesis.Source: §2 · p. 3

3 Background

The background formalizes document-grounded response generation as predicting a response from dialogue history and a grounding document. It then defines faithfulness relative to both inputs and describes a score whose higher values indicate greater faithfulness.

3 Background

  • Document-grounded response generation predicts the next response one token at a time from dialogue history and a grounding document.The model assigns a probability distribution over possible responses.Source: §3 · p. 3
  • A response is faithful to a document and dialogue history when those inputs logically entail it.The paper writes this condition as d, h ⊨ r and states that ⊨ represents logical entailment.Source: §3 · p. 3
  • A faithfulness metric takes the response, document, and dialogue history as inputs and returns a score, with higher scores indicating greater faithfulness.The response probability is typically factorized over its tokens.Source: §3 · p. 3

4 Approach

The approach defines PMI–FAITH to measure a response’s document association given dialogue history, then incorporates that faithfulness signal into PMI–DECODE. The decoder combines it with response likelihood and restricts token selection to mitigate grammar interference.

4.1 PMI–FAITH

  • PMI–FAITH uses conditional pointwise mutual information between the response and document given dialogue history to quantify document influence.A positive value indicates the response is more probable given the document and dialogue history than given the dialogue history alone.Source: §4.1 · p. 3

4.2 PMI–DECODE

  • PMI–DECODE selects responses by combining likelihood with a faithfulness score, modifying the standard log-likelihood objective.The objective uses a linear combination of likelihood and faithfulness, with α ∈ [0, 1].Source: §4.2 · p. 4
  • PMI–FAITH’s likelihood-like factorization makes its faithfulness term tractable for token-by-token decoding.The same language model can compute the conditional probabilities needed for CPMI by separately processing the document and dialogue history, or the history alone.Source: §4.2 · p. 4
  • Top-p masking limits CPMI-based token selection to a likelihood-ranked vocabulary subset because CPMI-selected document tokens can interfere with grammar.The subset is the minimum-cardinality set of tokens whose likelihood probabilities sum to p; the method aims to preserve likely, grammatical responses while using CPMI.Source: §4.2 · p. 4

5 Experiments

The experiments evaluate PMI–FAITH on BEGIN and PMI–DECODE on three document-grounded dialog datasets. PMI–FAITH aligns better with human judgments than other faithfulness metrics, while PMI–DECODE generally improves faithfulness over standard decoding, with relevance trade-offs.

5.1 PMI–FAITH: Experimental Setup

  • The BEGIN evaluation uses 11,059 human-annotated responses across CMU–DoG, TopicalChat, and WoW, labeled fully attributable, generic, or not fully attributable.For metric evaluation, fully attributable responses are treated as positive and the other two categories as negative.Source: §5.1 · p. 5
  • PMI–FAITH uses P(r|d,h) and P(r|h); across eight tested BLOOM and GPT-2 language models, its F1 varied by only 0.02 points, so later experiments use BLOOM-560m.The experiments also include UPMI–FAITH, which compares response and document probabilities without conditioning on dialogue history.Source: §5.1 · p. 5

5.2 PMI–FAITH: Experimental Results

  • On BEGIN, PMI–FAITH gains 5.6% in accuracy and 8.7% in F1 over Q2, while PMI–FAITH and UPMI–FAITH outperform other metrics across reported performance measures.Lexical- and semantic-similarity metrics have accuracy below the majority-class classifier’s 76.7%.Source: Table 1 · p. 5, §5.2 · p. 6
  • PMI–FAITH achieves the highest F1 on CMU–DoG and TopicalChat, with absolute gains over other metrics of more than 12% and 9.7%, respectively; FaithCritic leads on WoW.The authors attribute FaithCritic’s WoW overfitting to training data adapted from WoW and note its low performance on the other datasets.Source: §5.2 · p. 6, Table 2 · p. 6
  • PMI–FAITH correlates with human judgments better than the other faithfulness measures reported on BEGIN.The evaluation reports Spearman and Pearson correlations with human annotations and AUROC.Source: §5.2 · p. 6, Table 3 · p. 6

5.3 PMI–DECODE: Experimental Setup

  • PMI–DECODE is compared with likelihood-based beam search and beam sampling on MultiDoc2Dial, TopicalChat, and FaithDial using separately fine-tuned BART-Large models.The decoding baselines use beam size 4; α and top-p settings are selected on the development set.Source: §5.3 · p. 7, Table 4 · p. 7

5.4 PMI–DECODE: Experimental Results

  • For fixed α, increasing top-p raises faithfulness but lowers relevance; increasing α first raises then lowers faithfulness, while relevance decreases as α increases.The highest faithfulness is reported at α = 0.5 and p = 0.6; α = 0.25 and p = 0.6 match the standard variant’s relevance of 0.40 while yielding more faithful responses.Source: §5.4 · p. 8, Table 5 · p. 8
  • PMI–DECODE improves faithfulness over standard decoding, with a 15% improvement over greedy decode on both MD2D and TC; relevance improves on MD2D but slightly deteriorates on TC.The human study compared beam-search responses generated with standard and PMI–DECODE objectives.Source: §5.4 · pp. 8–9

6 Conclusion

PMI–FAITH measures faithfulness with conditional PMI between a response and document given dialogue history, while PMI–DECODE combines likelihood and faithfulness in decoding. The metric correlates better with human judgments on BEGIN, and the decoding objective generates more faithful responses on three datasets.

6 Conclusion

  • On BEGIN, PMI–FAITH correlates better with human judgments than existing metrics.The conclusion reports this benchmark result without giving a correlation value.Source: §6 · p. 9
  • Across three document-grounded dialog datasets, PMI–DECODE generates more faithful responses than the standard likelihood objective, according to automated metrics and a human study.Source: §6 · p. 9
  • PMI–FAITH quantifies response faithfulness using conditional PMI between the response and document given dialogue history.Source: §6 · p. 9
  • PMI–DECODE encourages faithful responses by maximizing both response likelihood and faithfulness.Source: §6 · p. 9

Limitations

PMI–DECODE has limitations in generic conversational settings, computational throughput, and grammaticality. The authors note possible mitigation approaches but leave further work open.

Limitations

  • PMI–DECODE cannot respond to generic chit-chat or pleasantries such as “Hello!” or “Good-bye”.Source: Limitations · p. 9
  • Computing CPMI requires three decoder inputs, which may reduce overall system throughput by half even when computed in parallel.Source: Limitations · p. 10
  • PMI–DECODE sometimes generates grammatically incorrect responses despite the fluency of pretrained language models.Source: Limitations · p. 10
  • Training with CTRL tokens is suggested as a possible way to support both generic and faithful responses, while choosing between them remains open work.Source: Limitations · p. 10

Ethics Statement

The authors report no new ethical concerns beyond those already associated with large language models. They caution that document-grounded decoding may increase or reduce profanity depending on the document.

Ethics Statement

  • The work introduces no new ethical concerns per se beyond those already faced by large language models.Source: Ethics Statement · p. 10
  • The decoding objective operates on top of any trained language model and produces text more faithful to an input document.Source: Ethics Statement · p. 10
  • A document containing profanity may increase the model’s likelihood of generating similar content, whereas a valid document may reduce its inherent likelihood of generating profanity.Source: Ethics Statement · p. 10

A Human Evaluation

The human evaluation presented two systems’ responses to incomplete document-grounded dialogs and asked evaluators to rate faithfulness, relevance, and grammar. The instructions define each dimension separately, and six evaluators with ML/NLP coursework and varied experience participated.

A Human Evaluation

  • Six in-house annotators participated; four were dialog-research experts and two beginners, and each had completed at least one formal ML/NLP course.Source: A Human Evaluation · p. 12
  • The task presented 50 incomplete dialogs with grounding documents and two systems’ next responses in randomized order.Source: A Human Evaluation · p. 12
  • Evaluators judged responses for faithfulness, relevance, and grammar using Agree, Neutral, or Disagree ratings.Source: A Human Evaluation · p. 12
  • Faithfulness was judged by whether the complete response could be inferred from the document, independently of relevance to the dialog.Source: A Human Evaluation · p. 12
  • Relevance was judged by whether a response was apt given the dialogue context and knowledge, without conveying incorrect information.Source: A Human Evaluation · p. 12
  • Grammar was scored independently of the dialog and document, with agreement reserved for responses resembling expert human writing.Source: A Human Evaluation · p. 13

B Faithfulness Metric Normalization

This section reports the dev-set minimum and maximum values used to normalize each metric, along with the faithfulness-classification thresholds selected for highest dev-set F1.

B Faithfulness Metric Normalization

Related papers

Loading 2305.12191v2…