Learning English with Peppa Pig

Attempts to model or simulate the acquisition of spoken language via grounding in the visual modality date to the beginning of this century (Roy and Pentland, 2002) but have gained momentum recently with the revival of neural networks (e.g., Synnaeve et al., 2014; Harwath and Glass, 2015; Harwath et al., 2016; Chrupała et al., 2017; Alishahi et al., 2017; Harwath et al., 2018; Merkx et al., 2019; Havard et al., 2019a; Rouditchenko et al., 2021; Khorrami and Räsänen, 2021; Peng and Harwath, 2022). Current approaches work well enough from an applied point of view, but most are not generalizable to real-life situations that humans or adaptive artificial agents experience. Commonly used training data consist of images or videos paired with spoken descriptions of the scene depicted: However, the type of input that a language learner receives from the environment is much more challenging. Firstly, speech is only loosely coupled with the visual modality in naturalistic settings (Matusevych et al., 2013; Beekhuizen et al., 2013). Speakers often mention concepts that are not present in the immediate perceptual context, or talk about events that are remote in space and/or time (for example past experiences or future plans).

Secondly, in addition to correlations between the visual scenes and the meaning of spoken utterances, there are also correlations with non-semantic aspects of the speech signal, such as the voices of specific speakers, as well as with non-speech ambient sounds. Although it is plausible that such non-semantic correlations can sometimes be useful to the learner in the general endeavor of making sense of the world, for the specific task of learning the semantics of linguistic units they are likely more often an obstacle, as they make it harder to zoom in on the meaning-bearing aspects of the audio signal.

In the current study we make a first step towards simulating the acquisition of language via grounding in perception in a more naturalistic scenario. Our main focus is on learning the meaning of linguistic expressions from spoken utterances grounded in video. We use the well-known children’s cartoon Peppa Pig as a case study. Compared to commonly used video datasets, this dataset has a number of interesting characteristics. The visual modality is very schematic, the language is simple in terms of vocabulary size and syntactic complexity, and analysis of its linguistic features suggests its suitability for beginner learners of English (Kokla, 2021; Scheffler et al., 2021). Crucially, however, most of the speech in the videos consists of naturalistic dialogs between the characters in which they do not only discuss the here and now, but also often use displaced language.¹ Thus, the utterances are only loosely and noisily correlated to the scenes and actions depicted in the videos.

This choice of data thus allows us to directly address the ecological limitations of the current approaches. In addition, the cartoon videos contain comments interjected by the narrator. We use these for evaluating the acquisition of meaning as they are more descriptive and less noisy and allow us to measure performance, while controlling for speaker characteristics.

We implement a simple bi-modal architecture that learns spoken language embeddings from videos, and train it on the Peppa Pig dataset. Our contributions are the following:

• We evaluate model performance in terms of video fragment retrieval and additionally design controlled evaluation protocols inspired by the intermodal preferential looking paradigm (Hirsh-Pasek and Golinkoff, 1996).

• We carry out ablations of model components in order to understand the effects of pre-training for the audio and video encoders, the role of temporal information, and of segmentation strategies while training.

We show that despite the challenges of our naturalistic training data, our model succeeds at learning associations between the form of spoken utterances and their visual semantics. Moreover, even though the model rarely hears words in isolation, it captures aspects of the visual meaning of frequent nouns and verbs. Our ablation studies suggest that temporal information contributes to video modeling (especially for longer segments), and that self-supervised pre-training followed by fine-tuning of the audio encoder is key to the best performance.

Early attempts at simulating grounded language learning focus on interactions between adults and young children while playing with a set of objects from different categories (Roy, 1999, 2002; Gorniak and Roy, 2003; Mukherjee and Roy, 2003). In a representative study from this series, Roy and Pentland (2002) use speech recorded from such interactions paired with different views of the visible objects to identify linguistic units (i.e., words) and visual categories, and to map these two modalities together. A hard-coded visual system extracts object representations from images, and spoken utterances are represented as phoneme probabilities generated by an RNN pre-trained on spectrograms. Their experiments on small-scale data (around 20 words and seven visual categories) show that the model can segment words and map them to visual categories.

The availability of datasets of images associated with spoken captions such as Flickr Audio Captions (Harwath and Glass, 2015), Places (Zhou et al., 2014), and Spoken COCO (Hsu et al., 2019) led to a rapid development of neural models of grounded language learning; see Chrupała (2022) for a comprehensive overview. In contrast to earlier approaches, these models are trained end-to-end directly on large datasets.

Following the architecture proposed in Karpathy et al. (2014), the visual and speech modality are usually encoded using separate pathways, and subsequently mapped into a joint representation space. Visual features are extracted from a pre-trained image classification model that processes the whole or a specific region of an image (however, see Harwath et al. [2018], who train the model end-to-end on images and their spoken captions on the Places dataset). The audio encoder component in most models is either an adaptation of Harwath et al. (2016), which feeds a spectrogram of the speech signal to a convolutional architecture, or a hybrid architecture of convolutional followed by recurrent layers using Mel-Frequency Cepstral Coefficient (MFCC) features from the audio signal as input, as introduced by Chrupała et al. (2017).

The majority of models of speech grounded in images are optimized for and evaluated on image retrieval via spoken caption and vice versa. Additionally, a range of diagnostic analyses have been performed on the hidden representations of these models to study whether they encode the identity and boundaries of subword units such as phonemes and syllables (Alishahi et al., 2017; Harwath and Glass, 2019; Khorrami and Räsänen, 2021) as well as individual words (Chrupała et al., 2017; Havard et al., 2019b). Moreover, in addition to examining form-meaning associations at the utterance level, Harwath and Glass (2017) explicitly learn a lexicon by extracting audio and image segments, clustering each modality separately, and mapping them together by calculating the pairwise similarities of their members in the joint semantic space.

There have also been recent attempts to learn spoken language grounded in video instead of static images. Boggust et al. (2019) sample audio-visual fragments from cooking videos; their grounded model treats video frames as still images ignoring the temporal dimension. Rouditchenko et al. (2021) integrate the temporal information when encoding videos from the Howto100m dataset (Miech et al., 2019), and perform better than previous work in language and video clip retrieval.

Models trained on instructional video datasets often do not generalize well to other domains. Monfort et al. (2021) highlight this limitation and show that training on their larger and more diverse Spoken Moments in Time dataset leads to better generalization. But the point remains that these video datasets contain descriptive speech, thus ensuring that there is a strong correlation between the spoken language and their visual context, a characteristic that is not representative of the experience of learning language in real world. We remedy this limitation by using a video dataset that does not guarantee a direct description of the visual context.

There are many studies on young children learning language by watching videos; see Vanderplank (2010) for a survey. The main takeaway of these studies is that language learning is much more effective in a social, conversational setting than by passively watching videos (Kuhl et al., 2003; Anderson and Pempek, 2005; Robb et al., 2009), but learning does happen in such contexts. Importantly for our goal, techniques such as the intermodal preferential looking paradigm have been developed to systematically test young language learners’ knowledge of words, syntactic structure, and semantic roles (Hirsh-Pasek and Golinkoff, 1996; Bergelson and Swingley, 2012; Noble et al., 2011). Nikolaus and Fourtassi (2021) use this evaluation strategy to test semantic knowledge at word and sentence level in their computational model of word learning from images. We adapt this approach to evaluate how our grounded model associates semantic information to spoken words and utterances from video.

One further aspect of learning spoken language via visual grounding is the fact that grounding is only part of the story. Human children arguably infer substantial amounts of information about language structure and meaning from purely intra-linguistic co-occurrence statistics (e.g., Saffran et al., 1996). A similar mechanism is what allows written language models such as BERT (Devlin et al., 2019) or GPT-3 (Brown et al., 2020) to capture and exhibit relatively sophisticated linguistic knowledge. Loosely similar approaches have started to also make an impact for the spoken modality (e.g., Baevski et al., 2020; Hsu et al., 2021). Here we take a simple pre-training-based approach to integrating this type of self-supervision with learning-via-grounding.

The main focus of this study is on the data and evaluation. We thus keep the components of our architecture simple, and follow established modeling practices whenever possible.

We use the dataset provided by (Papasarantopoulos and Cohen, 2021), which consists of metadata for the set of 209 episodes (seasons 1–5) of the English-language version of Peppa Pig.² The annotations created by Papasarantopoulos and Cohen (2021) feature written transcriptions aligned with the audio as well as segmentation into dialog and narration.³ Dialogs are the parts spoken by the characters, while narrations are comments inserted by the narrator, which are more descriptive in nature. All the narration segments are uttered by the same voice actor. We use the dialogs for training the model, and set aside the narrations for evaluation purposes only. A small portion of the dialog data is also used for validation. Specifically, out of the total 209 episodes, we use dialog from episodes 1–196 for training, and 197–209 for validation. We set aside narrations from episodes 1–104 for validation and 105–209 for testing. We disregard portions of the video that are annotated as neither dialog nor narration: This means our data consists mostly of video clips that contain some speech.⁴Table 1 shows the sizes of the training and validation splits. The vocabulary size of transcriptions corresponding to the training data is 5,580.

Our model is trained to discriminate positive video-audio pairs from negative ones. The positive pairs are those that are temporally coincident in the original video file. In order to generate these training items we need to split the videos into fragments. When preparing training data, we use annotations to separate dialog and narration data, but we do not use alignment with transcriptions for further segmentation, in order to make the setting naturalistic. Processing long segments of video and audio is not tractable on commodity GPU hardware, and we thus segment the data into brief snippets roughly comparable in length to the duration of a short sentence or a phrase. We use the following two segmentation strategies:

Using this approach we simply split sections into fixed-length non-overlapping fragments of 2.3 second duration. This length is close to the mean duration of audio aligned to a single line of subtitles. The number of clips for each dataset split is shown in Table 1.

For evaluation we have a number of different conditions and evaluation metrics described in detail in Section 3.4 and in some of these conditions we use the subtitles to guide segmentation.

The audio encoder portion of the model consists of a small wav2vec2 model (Baevski et al., 2020) pre-trained in a self-supervised fashion, without any supervised fine-tuning.⁶ The wav2vec2 architecture learns audio embeddings by self- supervised learning driven by a contrastive loss applied to quantized latent representations of masked frames, loosely inspired by the BERT approach to language modeling (Devlin et al., 2019).

As a video encoder we use the 18-layer ResNet (2+1)D architecture (Tran et al., 2018), pretrained on the action recognition dataset Kinetics-400 (Kay et al., 2017). The pre-trained model is available via PyTorch.⁷ This architecture implements 3D convolution by decomposing it into a 2D spatial convolution followed by 1D temporal convolution. The output of this module is aggregated using the attention mechanism with the same architecture as for the audio module, linearly projected to the same dimensionality as the audio (512) and L₂ normalized. For our experiments we also use a version of the video encoder without pre-training.

As a baseline to investigate the contribution of temporal information to video modeling, we swap the video ResNet (2+1)D with the 2D ResNet pre-trained on ImageNet, which embeds each video frame separately. These frame embeddings are then attention-pooled as with the standard video encoder.

To further investigate the impact of temporal information while controlling for model architecture, we evaluate model performance in a condition where we randomly scramble the video frames within a clip at test time, thereby removing any useful temporal information.

The most common approach to evaluation for visually grounded models trained on spoken image captions is caption-to-image retrieval (often combined with image-to-caption retrieval); this technique has been carried over from text-based image-caption modeling. With the standard spoken caption datasets this approach is unproblematic because the content of the captions is not correlated with extra-linguistic clues in the speech signal, such as speaker identity (since speakers are randomly assigned to captions) or non-speech environmental sounds. In such an artificial setting, a retrieval metric measures the ability of the model to match spoken utterances to images based on their semantic content. This is not the case for the Peppa Pig dataset: Here we can expect that when a video segment depicts a particular character (e.g., George) then the audio in this segment is more likely to contain utterances spoken by the voice actor playing George. Moreover, some characters might have a tendency to talk about certain topics more often than others, and the model might pick up on these associations instead of paying attention to the actual meaning of the uttered words. Due to these factors, in a naive retrieval setting, a model could obtain a high score by mostly capturing these non-linguistic correlations.

In order to control for these factors, we leverage the narrator speech in the videos. These utterances are always spoken by the same actor, so speaker identity cannot be used as a clue for matching video and audio. Furthermore, the narration segments are akin to video captions in that they tend to describe what is happening in the video and thus their semantic content is more strongly correlated with the content of the video than in the case of the dialog, which is also a desirable feature for the purposes of system evaluation.

For the retrieval evaluation, as for training, we also use both the fixed and the jitter segmentation strategies; however, for most conditions, we only report retrieval for the fixed evaluation data.

We encode each audio clip in a candidate set sampled from the validation (or test) data using the speech encoder part of the model; similarly, we encode each video clip using the video encoder. We then measure cosine similarity between the encodings of the audio clip and all the video clips. If the video clip corresponding to the audio is among the n most similar video clips, we count that as a success. The proportion of successes across all audio clips gives us the retrieval metric known as recall@n. In Section 5 we report recall@N of the complete model on narration test data for values of N between 1 and 10; for the rest of the experiments in this paper we focus on n = 10. We set the candidate set size to 100, and thus the random baseline for the recall@10 is 10%. In order to quantify uncertainty in this evaluation due to the test data we repeat this procedure 500 times with randomly sampled candidate sets and visualize the score distribution.

The absolute value the recall@10 of this metric may be hard to interpret as it depends on the size and content of the candidate set. For this reason, we evaluate model performance using a simpler, controlled scenario, inspired by intermodal preferential looking paradigms in child language acquisition (Hirsh-Pasek and Golinkoff, 1996). The proposed metric can be seen as a multimodal version of the ABX score proposed in Schatz (2016).

We extract clips aligned to a single subtitle line, group them by length, and for each pair of same-length video clips,⁸ we extract the audio from one of them (selected at random)—this is our anchor. The video clip from which the anchor was taken is the positive one, and the other video clip is the negative one. This triplet of stimuli forms a single test item.

While the triplet evaluation gives us a general idea about whether the model has learned a mapping between audio and video at the utterance level, it cannot tell us whether the model has acquired the grounded semantics of individual words.

To address this question, we probe the model’s performance in a more targeted triplet setup, where the model is required to select the correct video from a pair of videos whose corresponding transcripts only differ in one target word. To construct the evaluation set, we search the transcripts of the validation data for phrases with minimal differences with respect to the most commonly occurring nouns, verbs and adjectives. We set the minimum frequency of the target word in our training set to 10, and the minimum phrase duration to 0.3 seconds.⁹ Following Nikolaus and Fourtassi (2021), we pair every such triplet example with a corresponding counter-example to control the evaluation for linguistic biases in the dataset.

Figure 2 shows an example of how two counter- balanced test trials are constructed from audio and video clips. Here, the anchor A_example of the example triplet is the audio of Peppa loves jumping, the positive video P_example is the corresponding video, and the negative video N_example is the video corresponding to George loves jumping. In the counter-example triplet, the anchor A_counterex is the audio of George loves jumping, and the positive and negative videos are flipped.

We measure word accuracies by calculating the triplet accuracy for all triplets that contain a given word (e.g., Peppa in the previous example) either as target or distractor. That is, we take into account all cases where the model needs to use the meaning of the given word for either choosing or rejecting a video. We report word accuracy for all nouns and verbs for which we find at least 100 pairs of triplets in the validation set. We did not find enough examples for any adjectives, and thus did not include them in our evaluation.

We implement the architecture in PyTorch (Paszke et al., 2019). We use the Adam optimizer (Kingma and Ba, 2015) with the scheduling described in Devlin et al. (2019). We train every configuration on a single GPU and stop training after 48 hours, with batch-size 8 and accumulating gradients over 8 batches, in 16 bit precision mode. For each model configuration we save model weights after each epoch and report results for the checkpoint which gets the best triplet accuracy on the narration validation data.

Our code is publicly available at https://github.com/gchrupala/peppa, and can be consulted for further details of the experimental setup.

We account for two sources of variance in the results. Firstly, for each model configuration we ran four separate training runs in order to account for the effect of random initialization. Secondly, we estimate the variance due to validation/test sample by resampling validation and test items 500 times. In the case of the minimal pairs evaluation, we utilize bootstrapping with 100 re-samples. In most cases in Section 5, we pool variance from both sources and report overall spread, except when specifically focusing on the contribution of each source.

Figure 3 shows recall@N for values of N between 1 and 10 for the complete model on test narration data in both fixed and jitter conditions. Both plots show that the value of recall@N increases monotnically. For the rest of this paper, we only report recall@10.

Table 2 presents the recall@10 and triplet accuracy scores on test narration data obtained with the complete model. In Section 5.1 we investigate the impact of various components of our training setup on performance as measured by recall@10 and triplet accuracy. In Section 5.2 we focus on the targeted evaluation via minimal pairs.

For completeness, we report results on both dialog and narration data. However, the scores on narration are our main focus as they are not confounded by speaker-based clues, and thus indicate to what extent the model learns aspects of utterance meaning.

For experiments in Section 5.1.1 we include each run as a separate boxplot to show the consistency of the results between runs in different training conditions. For the other experiments we collapse the results of the four runs to avoid clutter.

Results on different pretraining configurations are shown in Figure 4. The best overall performance on both the dialog and the narration data is achieved with a model where both the video and audio encoder are pre-trained before being fine-tuned on our data. On narration data, for both metrics, we see a clear ranking of configurations from best to worst: audio and video pretraining (AV), audio pretraining (A), video pretraining (V), and no training (None). Meanwhile, for dialog data, the performance between A and V conditions is comparable. In the absence of any pretraining (None), some runs fail to converge, thus performing at chance level.

To further understand and disentangle the effects of audio pretraining and fine-tuning, we train a model with frozen parameters of the wav2vec2 module. The effect of this condition is shown in Figure 5. We find that, without fine-tuning of the wav2vec2 module, performance decreases substantially on both metrics. In other words, best performance is only achieved with pre-trained and fine-tuned models.

Next, we evaluate a model that has been trained with varying video and audio lengths (jitter). For fair comparison, here we report recall@10 for both fixed and jitter validation configurations. As seen in Figure 6, the effect of jitter is only minor and the performance is comparable. However, we do observe some performance improvements when using jitter in the minimal pairs evaluation (cf. Section 5.2).

Finally, we explore the role of the temporal nature of the visual modality. Figure 7 compares the model with the regular video encoder with one using the static baseline encoder. For this comparison we did not pretrain the video encoder in either condition, in order to remove the confound of the pretraining data.¹⁰ Across all metrics, we observe substantial performance drops for the static model, which has access to the same video frames, but does not have access to their temporal ordering.

Additionally, we investigate the effect of clip duration on this same comparison, using the triplet evaluation data. Figure 8 shows that the effect is nonlinear, and for the shortest clips temporal information does not help and may even have a detrimental effect.

Figure 9 shows the effect of scrambling the video frames along the temporal dimension at test time (note that here the video encoders are pretrained). As expected, we observe substantial performance drops when the model does not see the video frames in the correct order.

For this ablation the differential impact of clip duration on the two conditions is very similar as in the static ablation (figure not included).

Table 3 presents results for the minimal pair evaluation along with several ablations. Models that are pretrained and fine-tuned with jitter (row 0) perform best. In the first two configurations (rows 0 and 1), there is not much difference in the scores for verbs and nouns. However, we observe a substantial performance drop for both nouns and verbs if the wav2vec2 module is not fine-tuned.

If the model is trained without jitter (row 2), performance drops substantially for nouns, but not for verbs. One possible explanation for this could be that the evaluation samples for nouns are on average shorter than those for verbs (nouns: 0.43s vs. verbs: 0.49s), and a model trained with jitter performs better on short clips because it has been exposed to clips of varying duration during training. Supporting this hypothesis, we find a positive correlation between log duration of clips and accuracy, which is lower for models trained with jitter (Pearson r = 0.52, p < 0.001) than for models without jitter (Pearson r = 0.69, p < 0.001).

In line with the general results, we find that the benefit of audio pretraining (row 5) is greater than that of video pretraining (row 4). A model without any pretraining (row 3) only performs marginally above chance.

For a model trained with a static video encoder (row 6), we compare performance to a model that was also trained without video pretraining (row 5) as done for the general results. We observe a slight performance improvement for nouns, and no significant difference for verbs. We suspect that temporal information is not crucial for the minimal pairs evaluation, because most evaluation samples are clips of short duration (on average: 0.44s, i.e., 4–5 frames), thus limiting the benefit of the time dimension. As we saw in the analysis of clip duration (Figure 8), temporal information for such short clips does not improve performance, and could even have detrimental effects. In the alternative temporal ablation with scrambled video frames (row 7), we observe no significant performance drop compared to the base condition (row 0).

Figures 10 and 11 show per-word accuracy for nouns and verbs for the best performing model configuration. We observe substantial variance in the accuracy scores, suggesting that the difficulty to learn certain words varies. For example, the scores for house, car, and cake are the lowest. This could be because these concepts are not easy to ground, either because they are used in displaced speech or because they do not often refer to a similar visual entity. When looking at our evaluation samples, we find that indeed the word house is used in varying visual contexts (house entrance, whole house, inside the house, rabbit’s house) and in displaced speech (talking about going to somebody’s house). Cars are only sometimes completely visible, often we see only cartoon characters in a car. Regarding cake, it refers to either a whole cake, a slice, batter, or crumbs.

On the other end, performance for concrete words such as ice, box, and sand is the best, and indeed we find that in the evaluation examples these concepts are always present in the corresponding video and visually highly similar. Additionally, the words Pedro and Rebecca are learned very well: They refer to Pedro Pony and Rebecca Rabbit, easily visually distinguishable from characters belonging to other species.

Further investigations with larger datasets are necessary to reveal the underlying reasons for difficulty, and relating them to predictors of age of acquisition in the child language acquisition literature (Roy et al., 2015; Frank et al., 2021).

We simulate grounded language learning in a naturalistic setting, where the connection between the linguistic and visual modalities is not always strong and is potentially confounded by correlations with non-semantic aspects of the speech signal. Our experimental results suggest that despite the challenges inherent to the naturalistic aspects of our training dataset, a simple bimodal architecture can capture aspects of visual meaning of individual words as well as full utterances, and generalize well to narrative utterances featuring a single unseen speaker and a descriptive rather than conversational style. Our analyses show that generalization is substantially boosted by fine-tuning audio representations pretrained on unlabeled single-modality speech data. Fine-tuning a pretrained video encoder also makes a contribution, but is less crucial to generalization from dialog to narration.

We also investigate the role of temporal information in learning form-meaning mappings and show that having access to time information facilitates learning, except for very short video segments.

To better understand what aspects of language are learning in our setting, we need to carry out in-depth analyses of learned representations on sub-word, lexical, and phrasal levels. It would also be worthwhile to figure out the details of how specifically temporal information in video contributes to acquiring linguistic knowledge. Some analyses in this direction are currently constrained by the size of the evaluation dataset, and more large-scale datasets are needed in the future.

We model the acquisition of spoken language from language-internal correlations as well as from grounding in vision by fine-tuning an audio encoder pretrained on read speech. This approach is rather simplistic and does not match the real experience of language learners. It would be interesting to make the setting more realistic by using pretraining data which reflect a young learner’s experience more closely, and to realistically interleave learning via self-supervision from speech and via grounding in vision. Ideally we would want to dispense with supervised pretraining of the video encoder as well and rather use a model pretrained in a self-supervised way also for this modality.

We would like to thank Nikos Papasarantopoulos and Shay B. Cohen for creating the Peppa Pig annotations and for sharing them with us. Part of the NWO/E-Science Center grant number 027.018.G03 was used to purchase the video data on DVD. Thanks to Bertrand Higy for taking care of this purchase, as well as for sharing his ideas with us in the initial stages of this work. We also thank Abdellah Fourtassi, and three anonymous reviewers for their useful feedback.

This work was supported by grants ANR-16-CONV-0002 (ILCB), AMX-19-IET-009 (Archimedes Institute), and the Excellence Initiative of Aix-Marseille University (A*MIDEX).