/AM-FM

Primary LanguagePythonApache License 2.0Apache-2.0

The Deep AM-FM Framework

This framework intends to serve as a general evaluation framework for natural language generation tasks, such as machine translation, dialogue system and summarization.

Adequacy Metric

This component aims to assess the semantic aspect of system responses, more specifically, how much source information is preserved by the dialogue generation with reference to human-written responses. The continuous space model is adopted for evaluating adequacy where good word-level or sentence-level embedding techniques are studied to measure the semantic closessness of system responses and human references in the continous vector space.

Fluency Metric

This component aims to assess the syntactic validity of system responses. It tries to compare the system hypotheses against human references in terms of their respective sentence-level normalized log probabilities based on the assumption that sentences that are of similar syntactic validity should share similar perplexity level given by a language model. Hence, in this component, various language model techniques are explored to accurately estimate the sentence-level probability distribution.

Toolkit Requirements

  1. python 3.x
  2. emoji=0.5.4
  3. jsonlines=1.2.0
  4. tensorflow-gpu=1.14.0
  5. tqdm=4.38.0

Examples

Please refer to the example folder for detailed experimental setup and implementation steps of various evaluation tasks.

References

[1] Banchs, R. E., & Li, H. (2011, June). AM-FM: A semantic framework for translation quality assessment. In Proceedings of the 49th Annual Meeting of the ACL: Human Language Technologies: short papers-Volume 2 (pp. 153-158). ACL.

[2] Banchs, R. E., D’Haro, L. F., & Li, H. (2015). Adequacy–fluency metrics: Evaluating MT in the continuous space model framework. IEEE/ACM TASLP, 23(3), 472-482.

[3] D'Haro, L. F., Banchs, R. E., Hori, C., & Li, H. (2019). Automatic evaluation of end-to-end dialog systems with adequacy-fluency metrics. Computer Speech & Language, 55, 200-215.