This repository contains to run the experiments in our paper, Interpreting Indirect Answers Using Self-Rationalizing Models, where we finetune T5 (Raffel et al., 2019) to classify indirect answers and produce rationales which we evaluate with Leakage-Adjusted Simulatability Hase et al., 2020 and human judgments of plausibility, faithfulness, and quality.
We wrote this paper for the Spring 2021 edition of the University of Amsterdam MSc Artificial Intelligence course, Computational Dialogue Modelling.
A joint process of label prediction and free-form natural language rationale generation has been shown to offer faithful insights into a model's internal reasoning process. In this work, we formulate a text-to-text mixture task to investigate if we can successfully transfer self-rationalization capacity from existing human-provided explanations of natural language inference and question answering examples to the interpretation of indirect answers to polar questions. We show that our setup enables a model to generate faithful and coherent rationales — as measured by simulatability and human judgments of quality — with a minimal loss in predictive power when interpreting indirect answers. We identify evident shortcomings in our model's reasoning capacity through an extensive analysis of the generated rationales. Our results offer a clear direction in which the community can improve the logical prowess of current text-to-text models.
The Circa dataset (Louis et al., 2020) of (question, indirect answer) pairs does not come with human-provided rationales. To teach our T5 model to rationalize (generate a natural language explanation) its predictions on the Circa dataset, we employ a mixture task with data from two datasets that do contain such rationales.
We finetune a pretrained T5-Large model on a mixture of data from the Circa, e-SNLI (Camburu et al., 2018) and Cos-E (Rajani et al., 2019) datasets. During the finetuning process we ask our T5 model to predict one of four labels in a Natural Language Inference style task. We also ask of our T5 model to rationalize some of its predictions on the e-SNLI and Cos-E datasets. Supervised rationale generation is possible for the latter two datasets, since they contain reference rationales.
We then evaluate our finetuned T5 by asking it to both predict and rationalize instances from a held out Circa test set. We measure predictive power with accuracy. We measure rationale quality with the Leakage-Adjusted Simulatibilty metric as well as with human annotations of quality with a Mechanical Turk survey.
We modify to Natural Language Inference setting to match the E-SNLI dataset which supplies the bulk of the explanations from which the T5 model learns to rationalize. We use the relaxed Circa labels, mapped to the NLI setting as follows:
Yes --> entailmentNo --> contradictionIn the middle, neither yes nor no --> neutralYes, subject to some coniditions --> noneOther --> noneN/A --> none
We begin all input sequences with nli keyword to tell T5 that this is an NLI task. If we want T5 to explain its prediction, we put the explain keyword before nli. We use the declarative form of the question (e.g., Do you have any pets becomes You have pets) as the hypothesis, prepended with a hypothesis: keyword. We use the answer (e.g., My cat just turned one year old) as the premise prepended with a premise: keyword.
The Circa Dataset has two settings:
- Matched: where all response scenarios are seen during training.
- Unmatched: where some response scenarios are held out in the validation and test sets.
In the unmatched setting, we add the response scenario to the input sequence prepended with a context: keyword.
Thus, we transform a relaxed Circa Example such as:
Scenario: Y has just travelled from a different city to meet X.
Question: Did you drive straight here?
Answer: I had to stop at my mom's house.
Label: No
Into the following in the matched setting (with the optional explain keyword at the start):
input: nli hypothesis: I did drive straight here. premise: I had to stop at my mom's house.
target: contradiction
And the following in the unmatched setting: (with the optional explain keyword at the start):
input: nli context: Y has just travelled from a different city to meet X. hypothesis: I did drive straight here. premise: I had to stop at my mom's house.
target: contradiction
Since the Circa dataset has no pre-defined splits, we generate three unique splits based on three random seeds. Those are available in circa/data/
We follow the method of Narang et al., 2020, turning e-SNLI instances to input sequences in a similar manner to the circa dataset. The only difference is that there is no context and no none label.
Thus, we tranformed an e-SNLI example such as:
premise: Cardinals lost last night.
hypothesis: The Saint Louis Cardinals always win.
label: contradiction
explanation: you can't lose if you always win.
Into the following when we want T5 to only predict the correct label:
input: nli hypothesis: The Saint Louis Cardinals always win. premise: Cardinals lost last night.
target: contradiction
And the following when we want T5 to both predict and rationalize:
input: explain nli hypothesis: The Saint Louis Cardinals always win. premise: Cardinals lost last night.
target: contradiction. explanation: you can't lose if you always win.
The Cos-E dataset provides common-sense explanations to some instances of the CommonSenseQA dataset (Talmor et al., 2019). Like Narang et al., 2020, we transform instances into an NLI/Question-Answering hybrid format, using a premise keyword for the question and choice keyword before each possible answer.
For example, we transform the following Cos-E instance:
question: Where can you go to use a piano in your neighborhood if you don't have one?
choices: music school, music store, neighbor's house, lunch, drawing room
correct answer: neighbor's house
explanation: This word is the most relevant
Into the following when we want T5 to only predict the correct label:
input: nli premise: Where can you go to use a piano in your neighborhood if you don't have one? choice: music school choice: music store choice: neighbor's house choice: lunch choice: drawing room
target: neighbor's house
And the following when we want T5 to both predict and rationalize:
input: explain nli premise: Where can you go to use a piano in your neighborhood if you don't have one? choice: music school choice: music store choice: neighbor's house choice: lunch choice: drawing room
target: neighbor's house explanation: This word is the most relevant
To reproduce our experiments on Google Cloud TPUs, please follow the instructions detailed in finetune_on_gcp.md. We describe all of our experiments below. We conduct three runs of each experiment with three different seeds.
-
Circa Baseline: this involves finetuning the pretrained T5 model to generate the correct NLI labels for the Circa dataset (without explaining them), thus setting the benchmark for accuracy.
-
Circa Baseline - Hypothesis Only: This involves finetuning the pretrained T5 model to generate the correct NLI labels for the Circa dataset (without explaining them) without access to the hypothesis, thus setting the benchmark for accuracy in this ablated setting.
-
Circa Baseline - Premise Only: This involves finetuning the pretrained T5 model to generate the correct NLI labels for the Circa dataset (without explaining them) without access to the premise, thus setting the benchmark for accuracy in this ablated setting.
-
e-SNLI and Cos-E Zero Shot Transfer to Circa: This involves finetuning the pretrained T5 model on a mixture of E-SNLI and Cos-E data and then performing zero shot evaluation (predictions with explanations) on the Circa Dataset.
In these experiments, we finetune the T5 model on a mixture of five different tasks. The mixing rate for each task is proportional to the size of its dataset:
-
Predicting e-SNLI instances
-
Predicting and rationalizing e-SNLI instances
-
Predicting Cos-E instances
-
Predicting and rationalizing Cos-E instances
-
Predicting Circa instances
We then select the model checkpoint with the highest accuracy on the Circa dataset and ask that model to both predict and rationalize all instances from the held out Circa test set. We train one model per seed in both the matched and unmatched setting.
To train simulators and calculate LAS scores, please follow the instructions detailed in run_LAS.md.
For each seed and each setting (matched and unmatched), we take the finetuned T5 model checkpoint with the highest Circa validation accuracy. We then generate rationales on the training and validation sets as well, so we have rationales for every instance from the Circa dataset.
We then use those predictions to train a DistilBERT simulator model per the method of Hase et al., 2020. Finally, we calculate the LAS scores.
We use test set predictions from the unmatched model with the highest LAS score to generate our Mechanical Turk survey.
The questions are sampled such that we cover all four cases: (LAS leaked vs nonleaked) x (correct vs incorrect classification).
Annotators are shown the original Circa context and conversation, plus an interpretation and explanation generated by our model. In the case where our model makes the wrong prediction, the interpretation is either the gold standard label or the model prediction, but we do not disclose the correctness of this interpretation to the annotator so that they focus on rating the quality of the explanation and whether the explanation supports the given interpretation.
This setup also allows us to study both the faithfulness (whether the explanation supports the prediction) and plausibility (whether the explanation supports the gold standard label) of our model.
For more details of this survey, see Mechanical Turk Survey.
Read our paper, view our presentation, or explore the IPython notebooks in this repository to learn more!
We would like to thank Google for providing free access to Google Cloud v3-8 TPUs via the TPU Research Cloud program. We would also like to thank Dr. Raquel Fernández, Mario Giulianelli and Ece Takmaz for their assistance in conducting this research project.